{"totalScanned":5212,"collections":{"others":336,"neurips-2026":4876},"defaultVisible":2068,"severities":{"no_issues":2323,"minor_cosmetic":123,"repairable":1384,"major":520,"critical":41},"papers":[{"slug":"review-e-arc-01","title":"How Powerful are Graph Neural Networks?","authors":["Keyulu Xu","Weihua Hu","Jure Leskovec","Stefanie Jegelka"],"year":2018,"category":"Graphs & equivariance","status":"no-error","issueTitle":"How Powerful are Graph Neural Networks?","plainExplanation":"No error found in this selected proof. Countability and bounded multiplicity matter; the argument does not make the factor function continuous or guarantee learned MLP weights realize arbitrary infinite-alphabet mappings.","primaryUrl":"https://arxiv.org/abs/1810.00826","noteUrl":"/review?paper=review-e-arc-01","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-arc-02","title":"Weisfeiler and Leman Go Neural: Higher-order Graph Neural Networks","authors":["Christopher Morris","Martin Ritzert","Matthias Fey","William L. Hamilton","Jan Eric Lenssen","Gaurav Rattan","Martin Grohe"],"year":2018,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Weisfeiler and Leman Go Neural: Higher-order Graph Neural Networks","plainExplanation":"No error found in the 1-WL upper-bound proof. Does not certify the separate constructive lower bound or higher-order claims.","primaryUrl":"https://arxiv.org/abs/1810.02244","noteUrl":"/review?paper=review-e-arc-02","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-e-arc-03","title":"Invariant and Equivariant Graph Networks","authors":["Haggai Maron","Heli Ben-Hamu","Nadav Shamir","Yaron Lipman"],"year":2018,"category":"Graphs & equivariance","status":"addressed","issueTitle":"The equivariant basis count depends on graph size","plainExplanation":"Partitions with more blocks than nodes describe empty index patterns. Counting only realizable partitions gives the exact dimension; the unrestricted Bell number applies once \\(n\\ge k+\\ell\\). Pearce-Crump derives this correction in a separate paper.","primaryUrl":"https://arxiv.org/abs/1812.09902","citationCount":583,"citationUrl":"https://www.semanticscholar.org/paper/6541afa4b4061a7d5c8387514bedea9dc249fd80","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-e-arc-05","title":"Provably Powerful Graph Networks","authors":["Haggai Maron","Heli Ben-Hamu","Hadar Serviansky","Yaron Lipman"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Provably Powerful Graph Networks","plainExplanation":"No error found in selected separation proof. Generator theorem is an external algebraic input, not independently reproved here; neural approximation stages were not comprehensively audited.","primaryUrl":"https://arxiv.org/abs/1905.11136","noteUrl":"/review?paper=review-e-arc-05","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-arc-06","title":"Universal Invariant and Equivariant Graph Neural Networks","authors":["Nicolas Keriven","Gabriel Peyré"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Universal Invariant and Equivariant Graph Neural Networks","plainExplanation":"No error found in this substantive algebra-closure lemma. Does not certify the full equivariant Stone-Weierstrass proof or every graph-size topology assertion.","primaryUrl":"https://arxiv.org/abs/1905.04943","noteUrl":"/review?paper=review-e-arc-06","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-arc-07","title":"Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting","authors":["Giorgos Bouritsas","Fabrizio Frasca","Stefanos Zafeiriou","Michael M. Bronstein"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting","plainExplanation":"No issue was found in the checked claims. For a non-singleton connected H and vertex orbit i, each occurrence containing v contributes deg_H(i) incident directed edge counts. Summing all edge-orbit coordinates whose v-end lies in i and dividing by deg_H(i) reconstructs the vertex count. One extra layer can compute these counts then simulate each GSN-v layer. Singleton H has constant vertex count 1 and is handled by a bias.","primaryUrl":"https://arxiv.org/abs/2006.09252","noteUrl":"/review?paper=review-e-arc-07","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-arc-08","title":"Can Graph Neural Networks Count Substructures?","authors":["Zhengdao Chen","Lei Chen","Soledad Villar","Joan Bruna"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Can Graph Neural Networks Count Substructures?","plainExplanation":"No error found in selected initial-colour counting result. This paper uses its own k-WL convention; it is essential not to conflate its 2-WL with the folklore joint-neighbour version.","primaryUrl":"https://arxiv.org/abs/2002.04025","noteUrl":"/review?paper=review-e-arc-08","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-arc-11","title":"Are Transformers universal approximators of sequence-to-sequence functions?","authors":["Chulhee Yun","Srinadh Bhojanapalli","Ankit Singh Rawat","Sashank J. Reddi","Sanjiv Kumar"],"year":2019,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Are Transformers universal approximators of sequence-to-sequence functions?","plainExplanation":"No error found in this selected step of the main universality proof. The deeper contextual mapping and arbitrary modified-layer composition lemma were not fully independently verified; this is not certification of the entire theorem.","primaryUrl":"https://arxiv.org/abs/1912.10077","noteUrl":"/review?paper=review-e-arc-11","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-e-arc-13","title":"Theoretical Limitations of Self-Attention in Neural Sequence Models","authors":["Michael Hahn"],"year":2019,"category":"Transformers & sequence models","status":"addressed","issueTitle":"The proof must include changes in attention weights","plainExplanation":"Changing a token changes both values and attention weights. The original proof controls only the values. Hahn’s dated supplement includes the missing weight term and bounds it by \\(O(1/n)\\) under the original assumptions, preserving the influence theorem.","primaryUrl":"https://arxiv.org/abs/1906.06755","citationCount":420,"citationUrl":"https://www.semanticscholar.org/paper/b3564be8b79f25585acb035f3deaf4ae93c26d8f","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"big-bird-infinity-norm","title":"Big Bird: Transformers for Longer Sequences","authors":["Manzil Zaheer","Guru Guruganesh","Avinava Dubey","Joshua Ainslie","Chris Alberti","Santiago Ontanon","Philip Pham","Anirudh Ravula","Qifan Wang","Li Yang","Amr Ahmed"],"year":2020,"category":"Transformers & in-context learning","status":"finding","issueTitle":"Big Bird’s uniform-approximation lemma fails at a jump","plainExplanation":"Lemma 5 claims that an ordinary softmax/ReLU transformer can approximate every network in an intermediate class, even in L∞. That class permits discontinuous activations, whereas every finite ordinary transformer is continuous. Across a unit jump, a continuous approximation must have error at least 1/2. The finite-p argument can tolerate this error on a shrinking region; the L∞ claim cannot.","primaryUrl":"https://arxiv.org/pdf/2007.14062v2#page=21","citationCount":3071,"citationUrl":"https://www.semanticscholar.org/paper/044e13d7dd4e0655eb76f0bd00b2c1bdb44e2be3","severity":"repairable","issueTypes":["citation_hypothesis_mismatch","lemma_error"],"classifications":["repairable"]},{"slug":"review-e-arc-21","title":"Group Equivariant Convolutional Networks","authors":["Taco S. Cohen","Max Welling"],"year":2016,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Group Equivariant Convolutional Networks","plainExplanation":"No error found in the central mathematical claim (paper uses equation derivation rather than named theorem). Finite crops, boundaries and stride may restrict practical equivariance.","primaryUrl":"https://arxiv.org/abs/1602.07576","noteUrl":"/review?paper=review-e-arc-21","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-arc-22","title":"Steerable CNNs","authors":["Taco S. Cohen","Max Welling"],"year":2016,"category":"Machine learning","status":"no-error","issueTitle":"Steerable CNNs","plainExplanation":"No error found in full selected induction derivation. This establishes the feature transformation law; admissibility of nonlinearities for arbitrary feature types is a separate question.","primaryUrl":"https://arxiv.org/abs/1612.08498","noteUrl":"/review?paper=review-e-arc-22","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-arc-23","title":"A General Theory of Equivariant CNNs on Homogeneous Spaces","authors":["Taco Cohen","Mario Geiger","Maurice Weiler"],"year":2018,"category":"Graphs & equivariance","status":"unresolved","issueTitle":"Ordinary kernels need an additional operator hypothesis","plainExplanation":"The initial integral representation is too broad on continuous groups. Boundedness and equivariance allow point evaluation through a Dirac kernel, which is not an ordinary integrable function. The subsequent equivariance calculation is valid for operators that do admit the specified kernels.","primaryUrl":"https://arxiv.org/abs/1811.02017","noteUrl":"/review?paper=review-e-arc-23","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-e-arc-26","title":"E(n) Equivariant Graph Neural Networks","authors":["Victor Garcia Satorras","Emiel Hoogeboom","Max Welling"],"year":2021,"category":"Graphs & equivariance","status":"no-error","issueTitle":"E(n) Equivariant Graph Neural Networks","plainExplanation":"No error found in full selected equivariance proof; executable rotation/reflection/translation check included. E(n) includes reflections, so this result is not a guarantee to distinguish mirror configurations.","primaryUrl":"https://arxiv.org/abs/2102.09844","noteUrl":"/review?paper=review-e-arc-26","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-01","title":"Denoising Diffusion Implicit Models","authors":["Jiaming Song","Chenlin Meng","Stefano Ermon"],"year":2020,"category":"Generative models","status":"unresolved","issueTitle":"The Gaussian KL weights omit the conditional-mean coefficient","plainExplanation":"The equal-covariance Gaussian KL contains the square of that clean-input coefficient. Equation (30) drops it, and the conversion to noise-prediction error replaces the required \\(1-\\alpha_t\\) factor by \\(1/d\\). The existence of positive weights remains valid.","primaryUrl":"https://arxiv.org/abs/2010.02502","noteUrl":"/review?paper=review-e-gen-01","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-e-gen-02","title":"Variational Diffusion Models","authors":["Diederik P. Kingma","Tim Salimans","Ben Poole","Jonathan Ho"],"year":2021,"category":"Generative models","status":"no-error","issueTitle":"Variational Diffusion Models","plainExplanation":"No error in selected continuous-time equivalence argument. Does not establish finite-step invariance or equality for fixed untransformed network parameters.","primaryUrl":"https://arxiv.org/abs/2107.00630","noteUrl":"/review?paper=review-e-gen-02","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-03","title":"Maximum Likelihood Training of Score-Based Diffusion Models","authors":["Yang Song","Conor Durkan","Iain Murray","Stefano Ermon"],"year":2021,"category":"Generative models","status":"no-error","issueTitle":"Maximum Likelihood Training of Score-Based Diffusion Models","plainExplanation":"Selected Girsanov/data-processing argument checks out after the terminal average. No guarantee for singular data outside stated regularity or for the probability-flow ODE likelihood with an imperfect score.","primaryUrl":"https://arxiv.org/abs/2101.09258","noteUrl":"/review?paper=review-e-gen-03","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-04","title":"Neural Ordinary Differential Equations","authors":["Ricky T. Q. Chen","Yulia Rubanova","Jesse Bettencourt","David Duvenaud"],"year":2018,"category":"Machine learning","status":"unresolved","issueTitle":"The pointwise density ODE needs a differentiability qualification","plainExplanation":"The v6 statement and proof retain the same regularity issue as v5: a Lipschitz vector field need not possess a classical spatial derivative at every point. The new revision's documented correction to Fig. 3c does not change Theorem 1 or the differentiability step in Appendix A.","primaryUrl":"https://arxiv.org/abs/1806.07366","noteUrl":"/review?paper=review-e-gen-04","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-e-gen-05","title":"FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models","authors":["Will Grathwohl","Ricky T. Q. Chen","Jesse Bettencourt","Ilya Sutskever","David Duvenaud"],"year":2018,"category":"Generative models","status":"no-error","issueTitle":"FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models","plainExplanation":"No error in this central estimator derivation. Adaptive numerical solver error is outside the exact-integral unbiasedness result.","primaryUrl":"https://arxiv.org/abs/1810.01367","noteUrl":"/review?paper=review-e-gen-05","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-06","title":"Neural Autoregressive Flows","authors":["Chin-Wei Huang","David Krueger","Alexandre Lacoste","Aaron Courville"],"year":2018,"category":"Machine learning","status":"no-error","issueTitle":"Neural Autoregressive Flows","plainExplanation":"No counterexample to the selected weak-approximation construction. Compact approximation and conditional-CDF regularity are inherited from its lemmas, not independently strengthened to total variation/KL or finite network likelihood guarantees.","primaryUrl":"https://arxiv.org/abs/1804.00779","noteUrl":"/review?paper=review-e-gen-06","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-10","title":"Stochastic Interpolants: A Unifying Framework for Flows and Diffusions","authors":["Michael S. Albergo","Nicholas M. Boffi","Eric Vanden-Eijnden"],"year":2023,"category":"Generative models","status":"no-error","issueTitle":"Stochastic Interpolants: A Unifying Framework for Flows and Diffusions","plainExplanation":"No error in selected velocity-regression characterization. Existence/regularity and positivity supplied by preceding theorem were not independently audited in full.","primaryUrl":"https://arxiv.org/abs/2303.08797","noteUrl":"/review?paper=review-e-gen-10","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-11","title":"Building Normalizing Flows with Stochastic Interpolants","authors":["Michael S. Albergo","Eric Vanden-Eijnden"],"year":2022,"category":"Machine learning","status":"no-error","issueTitle":"Building Normalizing Flows with Stochastic Interpolants","plainExplanation":"No error in selected projection/minimization argument. Optimal-transport max-min proposition was not reviewed.","primaryUrl":"https://arxiv.org/abs/2209.15571","noteUrl":"/review?paper=review-e-gen-11","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-12","title":"Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions","authors":["Sitan Chen","Sinho Chewi","Jerry Li","Yuanzhi Li","Adil Salim","Anru R. Zhang"],"year":2022,"category":"Generative models","status":"no-error","issueTitle":"Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions","plainExplanation":"Selected theorem assembly and moment dependencies checked; full stopped-Girsanov proof of Theorem 9 not certified. No claim of assumptions-free sampling.","primaryUrl":"https://arxiv.org/abs/2209.11215","noteUrl":"/review?paper=review-e-gen-12","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-13","title":"A Theoretical Analysis of Contrastive Unsupervised Representation Learning","authors":["Sanjeev Arora","Hrishikesh Khandeparkar","Mikhail Khodak","Orestis Plevrakis","Nikunj Saunshi"],"year":2019,"category":"Representation learning","status":"no-error","issueTitle":"A Theoretical Analysis of Contrastive Unsupervised Representation Learning","plainExplanation":"No error in selected population-to-mean-classifier step and theorem assembly. Generalization lemma imported; guarantees average latent-class tasks, not arbitrary augmentations.","primaryUrl":"https://arxiv.org/abs/1902.09229","noteUrl":"/review?paper=review-e-gen-13","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-14","title":"Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere","authors":["Tongzhou Wang","Phillip Isola"],"year":2020,"category":"Representation learning","status":"no-error","issueTitle":"Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere","plainExplanation":"Latest v10 explicitly records the corrected convergence rate. No unresolved error in the selected asymptotic subresult; uniform-minimizer characterization not independently re-proved here.","primaryUrl":"https://arxiv.org/abs/2005.10242","noteUrl":"/review?paper=review-e-gen-14","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-15","title":"Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss","authors":["Jeff Z. HaoChen","Colin Wei","Adrien Gaidon","Tengyu Ma"],"year":2021,"category":"Representation learning","status":"no-error","issueTitle":"Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss","plainExplanation":"Selected central loss identity verified, including degree square roots. No certification of the subsequent multiway-Cheeger generalization theorem.","primaryUrl":"https://arxiv.org/abs/2106.04156","noteUrl":"/review?paper=review-e-gen-15","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-16","title":"Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style","authors":["Julius von Kügelgen","Yash Sharma","Luigi Gresele","Wieland Brendel","Bernhard Schölkopf","Michel Besserve","Francesco Locatello"],"year":2021,"category":"Representation learning","status":"no-error","issueTitle":"Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style","plainExplanation":"No counterexample to selected identification proof under its product-space and smoothness hypotheses. Result is block identification, not coordinatewise disentanglement.","primaryUrl":"https://arxiv.org/abs/2106.04619","noteUrl":"/review?paper=review-e-gen-16","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-19","title":"On Variational Bounds of Mutual Information","authors":["Ben Poole","Sherjil Ozair","Aaron van den Oord","Alexander A. Alemi","George Tucker"],"year":2019,"category":"Generative models","status":"no-error","issueTitle":"On Variational Bounds of Mutual Information","plainExplanation":"No issue was found in the checked claims. Selected central bound verified. This paper supplies an addressing proof for the CPC finite-negative InfoNCE derivation; no new error retained.","primaryUrl":"https://arxiv.org/abs/1905.06922","noteUrl":"/review?paper=review-e-gen-19","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-e-gen-20","title":"Representation Learning with Contrastive Predictive Coding","authors":["Aaron van den Oord","Yazhe Li","Oriol Vinyals"],"year":2018,"category":"Representation learning","status":"addressed","issueTitle":"The finite-batch InfoNCE bound needs an exact argument","plainExplanation":"Replacing a random negative-sample denominator by its mean is not an exact finite-sample argument. A candidate-index KL proof establishes the stated InfoNCE lower bound, as the later mutual-information analysis explains.","primaryUrl":"https://arxiv.org/abs/1807.03748","citationCount":14412,"citationUrl":"https://www.semanticscholar.org/paper/b227f3e4c0dc96e5ac5426b85485a70f2175a205","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-e-gen-21","title":"On Mutual Information Maximization for Representation Learning","authors":["Michael Tschannen","Josip Djolonga","Paul K. Rubenstein","Sylvain Gelly","Mario Lucic"],"year":2019,"category":"Representation learning","status":"no-error","issueTitle":"On Mutual Information Maximization for Representation Learning","plainExplanation":"No error in selected named result. This is a bound, not an assertion that maximizing MI guarantees useful representations.","primaryUrl":"https://arxiv.org/abs/1907.13625","noteUrl":"/review?paper=review-e-gen-21","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-22","title":"Understanding Dimensional Collapse in Contrastive Self-supervised Learning","authors":["Li Jing","Pascal Vincent","Yann LeCun","Yuandong Tian"],"year":2021,"category":"Representation learning","status":"unresolved","issueTitle":"Positive InfoNCE growth need not produce divergent weight norms","plainExplanation":"A strictly positive derivative can have finite integral. The actual finite-data InfoNCE dynamics can converge to finite, distinct singular values while retaining a nonzero Gram difference. Thus neither the divergence step nor the subsequent limiting Gram equality follows from the stated premises.","primaryUrl":"https://arxiv.org/abs/2110.09348","noteUrl":"/review?paper=review-e-gen-22","severity":"repairable","issueTypes":["invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-e-gen-23","title":"Understanding self-supervised Learning Dynamics without Contrastive Pairs","authors":["Yuandong Tian","Xinlei Chen","Surya Ganguli"],"year":2021,"category":"Representation learning","status":"no-error","issueTitle":"Understanding self-supervised Learning Dynamics without Contrastive Pairs","plainExplanation":"No issue was found in the checked claims. Balance invariant verified. Appendix Eq.(43) omits alpha_p^-1, but proof Eq.(46) and final line explicitly provide the corrected expression; excluded as an already-addressed local formula slip.","primaryUrl":"https://arxiv.org/abs/2102.06810","noteUrl":"/review?paper=review-e-gen-23","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-24","title":"Understanding the Behaviour of Contrastive Loss","authors":["Feng Wang","Huaping Liu"],"year":2020,"category":"Representation learning","status":"no-error","issueTitle":"Understanding the Behaviour of Contrastive Loss","plainExplanation":"No error in selected central entropy claim under fixed similarities. Does not prove entropy increases along training when similarities also change.","primaryUrl":"https://arxiv.org/abs/2012.09740","noteUrl":"/review?paper=review-e-gen-24","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-gen-25","title":"Debiased Contrastive Learning","authors":["Ching-Yao Chuang","Joshua Robinson","Lin Yen-Chen","Antonio Torralba","Stefanie Jegelka"],"year":2020,"category":"Representation learning","status":"unresolved","issueTitle":"A direct expectation bound repairs the debiasing proof","plainExplanation":"The two displayed tail integrals give twice the coefficient in the theorem. The proof also treats a signed loss difference as a nonnegative variable. These steps do not prove the stated constant, but a direct bounded-variance argument proves a stronger absolute-error bound.","primaryUrl":"https://arxiv.org/abs/2007.00224","noteUrl":"/review?paper=review-e-gen-25","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-e-gen-26","title":"Contrastive Learning with Hard Negative Samples","authors":["Joshua Robinson","Ching-Yao Chuang","Suvrit Sra","Stefanie Jegelka"],"year":2020,"category":"Representation learning","status":"no-error","issueTitle":"Contrastive Learning with Hard Negative Samples","plainExplanation":"No issue was found in the checked claims. Selected Gibbs concentration claim verified by reconstruction. Printed essential-supremum definition and intermediate signs contain slips, but the essential-supremum route states the needed object; no separate unresolved issue retained.","primaryUrl":"https://arxiv.org/abs/2010.04592","noteUrl":"/review?paper=review-e-gen-26","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-02","title":"Benefits of depth in neural networks","authors":["Matus Telgarsky"],"year":2016,"category":"Machine learning","status":"no-error","issueTitle":"Benefits of depth in neural networks","plainExplanation":"No error found in selected lemma.","primaryUrl":"https://arxiv.org/pdf/1602.04485v2","noteUrl":"/review?paper=review-e-opt-02","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-03","title":"The Power of Depth for Feedforward Neural Networks","authors":["Ronen Eldan","Ohad Shamir"],"year":2015,"category":"Machine learning","status":"no-error","issueTitle":"The Power of Depth for Feedforward Neural Networks","plainExplanation":"No error found in selected Fourier-geometric lemma.","primaryUrl":"https://arxiv.org/pdf/1512.03965v4","noteUrl":"/review?paper=review-e-opt-03","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-04","title":"Error bounds for approximations with deep ReLU networks","authors":["Dmitry Yarotsky"],"year":2016,"category":"Machine learning","status":"no-error","issueTitle":"Error bounds for approximations with deep ReLU networks","plainExplanation":"No error found in selected proposition.","primaryUrl":"https://arxiv.org/pdf/1610.01145v3","noteUrl":"/review?paper=review-e-opt-04","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-05","title":"The Expressive Power of Neural Networks: A View from the Width","authors":["Zhou Lu","Hongming Pu","Feicheng Wang","Zhiqiang Hu","Liwei Wang"],"year":2017,"category":"Machine learning","status":"no-error","issueTitle":"The Expressive Power of Neural Networks: A View from the Width","plainExplanation":"No error found in selected theorem after reconstructing its compactness step.","primaryUrl":"https://arxiv.org/pdf/1709.02540v3","noteUrl":"/review?paper=review-e-opt-05","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-08","title":"Universal Approximation with Deep Narrow Networks","authors":["Patrick Kidger","Terry Lyons"],"year":2019,"category":"Machine learning","status":"no-error","issueTitle":"Universal Approximation with Deep Narrow Networks","plainExplanation":"No error found in selected construction lemmas.","primaryUrl":"https://arxiv.org/pdf/1905.08539v2","noteUrl":"/review?paper=review-e-opt-08","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-09","title":"On the Expressive Power of Deep Neural Networks","authors":["Maithra Raghu","Ben Poole","Jon Kleinberg","Surya Ganguli","Jascha Sohl-Dickstein"],"year":2016,"category":"Machine learning","status":"no-error","issueTitle":"On the Expressive Power of Deep Neural Networks","plainExplanation":"No error found in selected region-count proof.","primaryUrl":"https://arxiv.org/pdf/1606.05336v6","noteUrl":"/review?paper=review-e-opt-09","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-10","title":"On the Expressive Power of Deep Learning: A Tensor Analysis","authors":["Nadav Cohen","Or Sharir","Amnon Shashua"],"year":2015,"category":"Machine learning","status":"no-error","issueTitle":"On the Expressive Power of Deep Learning: A Tensor Analysis","plainExplanation":"No error found in selected rank proof.","primaryUrl":"https://arxiv.org/pdf/1509.05009v3","noteUrl":"/review?paper=review-e-opt-10","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-e-opt-13","title":"Deep Learning without Poor Local Minima","authors":["Kenji Kawaguchi"],"year":2016,"category":"Optimization & generalization","status":"addressed","issueTitle":"The proposed saddle family need not have zero Hessian","plainExplanation":"Zero hidden factors with an arbitrary output factor can have nonzero Hessian and a strict saddle. Setting every factor to zero supplies the non-strict saddle needed by the corollary. The later classification explains which other critical points are strict.","primaryUrl":"https://arxiv.org/pdf/1605.07110v3","citationCount":1015,"citationUrl":"https://www.semanticscholar.org/paper/07925910d45761d96269fc3bdfdc21b1d20d84ad","severity":"minor_cosmetic","issueTypes":["incorrect_bound"],"classifications":["minor_cosmetic"]},{"slug":"review-e-opt-15","title":"Sharp Minima Can Generalize For Deep Nets","authors":["Laurent Dinh","Razvan Pascanu","Samy Bengio","Yoshua Bengio"],"year":2017,"category":"Machine learning","status":"no-error","issueTitle":"Sharp Minima Can Generalize For Deep Nets","plainExplanation":"No error found in selected symmetry proofs.","primaryUrl":"https://arxiv.org/pdf/1703.04933v2","noteUrl":"/review?paper=review-e-opt-15","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-16","title":"Train faster, generalize better: Stability of stochastic gradient descent","authors":["Moritz Hardt","Benjamin Recht","Yoram Singer"],"year":2015,"category":"Optimization & generalization","status":"no-error","issueTitle":"Train faster, generalize better: Stability of stochastic gradient descent","plainExplanation":"No error found in selected convex-stability proof.","primaryUrl":"https://arxiv.org/pdf/1509.01240v2","noteUrl":"/review?paper=review-e-opt-16","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-18","title":"Spectrally-normalized margin bounds for neural networks","authors":["Peter Bartlett","Dylan J. Foster","Matus Telgarsky"],"year":2017,"category":"Machine learning","status":"no-error","issueTitle":"Spectrally-normalized margin bounds for neural networks","plainExplanation":"No error found in selected covering lemma.","primaryUrl":"https://arxiv.org/pdf/1706.08498v2","noteUrl":"/review?paper=review-e-opt-18","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-19","title":"Norm-Based Capacity Control in Neural Networks","authors":["Behnam Neyshabur","Ryota Tomioka","Nathan Srebro"],"year":2015,"category":"Machine learning","status":"no-error","issueTitle":"Norm-Based Capacity Control in Neural Networks","plainExplanation":"No error found in selected norm-equivalence proof.","primaryUrl":"https://arxiv.org/pdf/1503.00036v2","noteUrl":"/review?paper=review-e-opt-19","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-20","title":"Neural Tangent Kernel: Convergence and Generalization in Neural Networks","authors":["Arthur Jacot","Franck Gabriel","Clément Hongler"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Neural Tangent Kernel: Convergence and Generalization in Neural Networks","plainExplanation":"No error found in selected initialization recursion.","primaryUrl":"https://arxiv.org/pdf/1806.07572v4","noteUrl":"/review?paper=review-e-opt-20","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-21","title":"Wide Neural Networks of Any Depth Evolve as Linear Models Under Gradient Descent","authors":["Jaehoon Lee","Lechao Xiao","Samuel S. Schoenholz","Yasaman Bahri","Roman Novak","Jascha Sohl-Dickstein","Jeffrey Pennington"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Wide Neural Networks of Any Depth Evolve as Linear Models Under Gradient Descent","plainExplanation":"No error found in selected predictive-law derivation.","primaryUrl":"https://arxiv.org/pdf/1902.06720v4","noteUrl":"/review?paper=review-e-opt-21","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-22","title":"On Lazy Training in Differentiable Programming","authors":["Lenaic Chizat","Edouard Oyallon","Francis Bach"],"year":2018,"category":"Machine learning","status":"no-error","issueTitle":"On Lazy Training in Differentiable Programming","plainExplanation":"No error found in selected finite-horizon proof.","primaryUrl":"https://arxiv.org/pdf/1812.07956v5","noteUrl":"/review?paper=review-e-opt-22","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-23","title":"A Mean Field View of the Landscape of Two-Layers Neural Networks","authors":["Song Mei","Andrea Montanari","Phan-Minh Nguyen"],"year":2018,"category":"Machine learning","status":"no-error","issueTitle":"A Mean Field View of the Landscape of Two-Layers Neural Networks","plainExplanation":"No error found in selected dissipation result.","primaryUrl":"https://arxiv.org/pdf/1804.06561v2","noteUrl":"/review?paper=review-e-opt-23","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-25","title":"On the Global Convergence of Gradient Descent for Over-parameterized Models using Optimal Transport","authors":["Lenaic Chizat","Francis Bach"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"On the Global Convergence of Gradient Descent for Over-parameterized Models using Optimal Transport","plainExplanation":"No error found in selected proposition; one displayed one-sided derivative was read with its required inequality.","primaryUrl":"https://arxiv.org/pdf/1805.09545v2","noteUrl":"/review?paper=review-e-opt-25","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-26","title":"Gradient Descent Provably Optimizes Over-parameterized Neural Networks","authors":["Simon S. Du","Xiyu Zhai","Barnabas Poczos","Aarti Singh"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Gradient Descent Provably Optimizes Over-parameterized Neural Networks","plainExplanation":"No error found in selected positive-definiteness proof.","primaryUrl":"https://arxiv.org/pdf/1810.02054v2","noteUrl":"/review?paper=review-e-opt-26","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-27","title":"A Convergence Theory for Deep Learning via Over-Parameterization","authors":["Zeyuan Allen-Zhu","Yuanzhi Li","Zhao Song"],"year":2018,"category":"Optimization & generalization","status":"unresolved","issueTitle":"The Gaussian gate fact needs a nondegenerate gate direction","plainExplanation":"The two definitions of \\(v\\) need not agree for arbitrary \\(q\\). Under the gated definition, the allowed choice \\(q=-h\\) activates every row because the inequality is non-strict, invalidating the half-active distribution.","primaryUrl":"https://arxiv.org/pdf/1811.03962v5","noteUrl":"/review?paper=review-e-opt-27","severity":"minor_cosmetic","issueTypes":["definition_error"],"classifications":["minor_cosmetic"]},{"slug":"review-e-opt-29","title":"Implicit Regularization in Matrix Factorization","authors":["Suriya Gunasekar","Blake Woodworth","Srinadh Bhojanapalli","Behnam Neyshabur","Nathan Srebro"],"year":2017,"category":"Machine learning","status":"no-error","issueTitle":"Implicit Regularization in Matrix Factorization","plainExplanation":"No error found after reconstructing the limiting dual argument; a normalization typo is repairable.","primaryUrl":"https://arxiv.org/pdf/1705.09280v1","noteUrl":"/review?paper=review-e-opt-29","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-opt-30","title":"Gradient Descent Maximizes the Margin of Homogeneous Neural Networks","authors":["Kaifeng Lyu","Jian Li"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Gradient Descent Maximizes the Margin of Homogeneous Neural Networks","plainExplanation":"No error found in selected smoothed-margin proof.","primaryUrl":"https://arxiv.org/pdf/1906.05890v4","noteUrl":"/review?paper=review-e-opt-30","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-rl-01","title":"GFlowNet Foundations","authors":["Yoshua Bengio","Salem Lahlou","Tristan Deleu","Edward J. Hu","Mo Tiwari","Emmanuel Bengio"],"year":2023,"category":"Machine learning","status":"no-error","issueTitle":"GFlowNet Foundations","plainExplanation":"No error found in the positive-flow conservation/uniqueness argument.","primaryUrl":"https://www.jmlr.org/papers/volume24/22-0364/22-0364.pdf","noteUrl":"/review?paper=review-e-rl-01","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-e-rl-02","title":"Trajectory balance: Improved credit assignment in GFlowNets","authors":["Esmeralda S. Whitammer","Moksh Jain","Emmanuel Bengio","Chen Sun","Yoshua Bengio"],"year":2022,"category":"Reinforcement learning","status":"addressed","issueTitle":"The gradient-variance formula reverses the likelihood ratio","plainExplanation":"Expanding the two covariance matrices gives \\(-\\mathbb E[(1+2a)SS^T]\\), where \\(a=\\log(P_F/T)\\). The printed inverse ratio changes that sign. Correcting it preserves a precise local variance comparison near the target.","primaryUrl":"https://arxiv.org/pdf/2201.13259","citationCount":302,"citationUrl":"https://www.semanticscholar.org/paper/0116eb4208e9abff3064934d178e655e716d92ad","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"cql-linear-projection","title":"Conservative Q-Learning for Offline Reinforcement Learning","authors":["Aviral Kumar","Aurick Zhou","George Tucker","Sergey Levine"],"year":2020,"category":"Reinforcement learning","status":"finding","issueTitle":"CQL's projected conservative penalty can be negative","plainExplanation":"Projection can reverse the sign of the averaged conservative term. In a one-state, two-action example with full data support and an exactly representable Bellman target, the projected term is negative. A positive CQL weight then raises the evaluated policy value above its true value.","primaryUrl":"https://arxiv.org/pdf/2006.04779v3","citationCount":2786,"citationUrl":"https://www.semanticscholar.org/paper/28db20a81eec74a50204686c3cf796c42a020d2e","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-e-rl-06","title":"Offline Reinforcement Learning with Implicit Q-Learning","authors":["Ilya Kostrikov","Ashvin Nair","Sergey Levine"],"year":2022,"category":"Reinforcement learning","status":"no-error","issueTitle":"Offline Reinforcement Learning with Implicit Q-Learning","plainExplanation":"No false result established; supplied the contraction step omitted by the short proof.","primaryUrl":"https://arxiv.org/pdf/2110.06169","noteUrl":"/review?paper=review-e-rl-06","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-e-rl-07","title":"A General Theoretical Paradigm to Understand Learning from Human Preferences","authors":["Mohammad Gheshlaghi Azar","Mark Rowland","Bilal Piot","Daniel Guo","Daniele Calandriello","Michal Valko","Rémi Munos"],"year":2024,"category":"Machine learning","status":"no-error","issueTitle":"A General Theoretical Paradigm to Understand Learning from Human Preferences","plainExplanation":"No issue was found in the checked claims. Selected uniqueness and population/sample-loss equivalence verified.","primaryUrl":"https://arxiv.org/pdf/2310.12036","noteUrl":"/review?paper=review-e-rl-07","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-e-rl-08","title":"ORPO: Monolithic Preference Optimization without Reference Model","authors":["Jiwoo Hong","Noah Lee","James Thorne"],"year":2024,"category":"Reinforcement learning","status":"addressed","issueTitle":"The printed odds-ratio gradient is missing a minus sign","plainExplanation":"The defined negative log-sigmoid loss differentiates with a minus sign. Correcting the displayed manual derivative agrees with the later tempered-gradient formula at unit temperature; the scalar loss itself is unchanged.","primaryUrl":"https://arxiv.org/pdf/2403.07691","citationCount":689,"citationUrl":"https://www.semanticscholar.org/paper/973814cd535facbf4f27c3de477b05bf19366030","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-o-2505-13447","title":"Mean Flows for One-step Generative Modeling","authors":["Geng, Zhengyang","Deng, Mingyang","Bai, Xingjian","Kolter, Zico","He, Kaiming"],"year":2025,"category":"Generative models","status":"no-error","issueTitle":"Mean Flows for One-step Generative Modeling","plainExplanation":"No issue was found in the checked claims. Identity and selected unweighted update proof hold under stated regularity. Adaptive-weight observation is a limit of that justification, not a disproof of the identity or empirical algorithm.","primaryUrl":"https://papers.nips.cc/paper_files/paper/2025/file/6d13e085b79d454da5910e4ca82a3d9d-Paper-Conference.pdf","noteUrl":"/review?paper=review-o-2505-13447","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-o-1710-10345","title":"The Implicit Bias of Gradient Descent on Separable Data","authors":["Daniel Soudry","Elad Hoffer","Mor Shpigel Nacson","Suriya Gunasekar","Nathan Srebro"],"year":2017,"category":"Optimization & generalization","status":"addressed","issueTitle":"An unproved multiclass assumption was added to the theorem","plainExplanation":"Version 7 explicitly assumes a solution to the support-vector equations and acknowledges that the earlier binary-case existence proof did not establish generic multiclass solvability. The existing addressed note already records this correction. The earlier almost-all claim is not treated as disproved, and the binary result is unaffected.","primaryUrl":"https://arxiv.org/pdf/1710.10345v6","citationCount":1160,"citationUrl":"https://www.semanticscholar.org/paper/11adc8bd35bd897502f9b5452ab7ac668ec9b0fb","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-o-2103-03404","title":"Attention is Not All You Need: Pure Attention Loses Rank Doubly Exponentially with Depth","authors":["Yihe Dong","Jean-Baptiste Cordonnier","Andreas Loukas"],"year":2021,"category":"Transformers & sequence models","status":"unresolved","issueTitle":"A signed residual invalidates an intermediate attention inequality","plainExplanation":"An entrywise inequality for nonnegative attention weights does not preserve its direction after multiplication by signed features. Cancellation in the displayed right side can make it zero while the true residual is nonzero, even under the added oscillation condition.","primaryUrl":"https://arxiv.org/pdf/2103.03404v2","noteUrl":"/review?paper=review-o-2103-03404","severity":"repairable","issueTypes":["invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"addressed-o-2303-01469","title":"Consistency Models","authors":["Yang Song","Prafulla Dhariwal","Mark Chen","Ilya Sutskever"],"year":2023,"category":"Generative models","status":"addressed","issueTitle":"The loss-limit argument does not justify a moving-average teacher","plainExplanation":"The original loss approximation does not establish that the limiting objective remains useful when teacher and student parameters differ. The authors' follow-up gives an explicit model where the consistency losses agree but their limit loses dependence on the data, and removes the teacher's moving average. This corrects the training justification; it does not refute the original loss-approximation identity.","primaryUrl":"https://proceedings.mlr.press/v202/song23a/song23a.pdf","citationCount":2155,"citationUrl":"https://www.semanticscholar.org/paper/ac974291d7e3a152067382675524f3e3c2ded11b","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-o-2209-03003","title":"Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow","authors":["Xingchao Liu","Chengyue Gong","Qiang Liu"],"year":2022,"category":"Machine learning","status":"no-error","issueTitle":"Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow","plainExplanation":"No error found in the selected convex-cost proof under ODE and integrability hypotheses; not a guarantee for every numerical solver or trained velocity approximation.","primaryUrl":"https://arxiv.org/pdf/2209.03003","noteUrl":"/review?paper=review-o-2209-03003","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-o-1811-00995","title":"Invertible Residual Networks","authors":["Jens Behrmann","Will Grathwohl","Ricky T. Q. Chen","David Duvenaud","Joern-Henrik Jacobsen"],"year":2018,"category":"Machine learning","status":"no-error","issueTitle":"Invertible Residual Networks","plainExplanation":"No issue was found in the checked claims. For target y, x->y-g(x) is a contraction, so Banach gives a unique inverse. Triangle inequalities give (1-L)||x-xprime|| <= ||F(x)-F(xprime)||. Eigenvalues of I+Jg need only have positive real parts, not be real positive: a half-rotation gives1±i/2. Real eigenvalues are positive and complex eigenvalues pair conjugately, so the determinant is positive.","primaryUrl":"https://proceedings.mlr.press/v97/behrmann19a/behrmann19a.pdf","noteUrl":"/review?paper=review-o-1811-00995","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-o-1802-05957","title":"Spectral Normalization for Generative Adversarial Networks","authors":["Takeru Miyato","Toshiki Kataoka","Masanori Koyama","Yuichi Yoshida"],"year":2018,"category":"Generative models","status":"no-error","issueTitle":"Spectral Normalization for Generative Adversarial Networks","plainExplanation":"No issue was found in the checked claims. Selected exact-operator argument is valid. Implementation approximation is not an unconditional certificate and is acknowledged; no retained error.","primaryUrl":"https://arxiv.org/pdf/1802.05957","noteUrl":"/review?paper=review-o-1802-05957","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir001","title":"Trust Region Policy Optimization","authors":["John Schulman","Sergey Levine","Philipp Moritz","Michael I. Jordan","Pieter Abbeel"],"year":2015,"category":"Optimization & generalization","status":"no-error","issueTitle":"Trust Region Policy Optimization","plainExplanation":"No issue was found in the checked claims. Validates the uniform-divergence performance bound; the implemented average-KL constrained approximation is not given this same guarantee. Does not certify optimisation of the neural surrogate.","primaryUrl":"https://arxiv.org/abs/1502.05477","noteUrl":"/review?paper=review-ir001","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir002","title":"High-Dimensional Continuous Control Using Generalized Advantage Estimation","authors":["John Schulman","Philipp Moritz","Sergey Levine","Michael Jordan","Pieter Abbeel"],"year":2015,"category":"Reinforcement learning","status":"no-error","issueTitle":"High-Dimensional Continuous Control Using Generalized Advantage Estimation","plainExplanation":"No issue was found in the checked claims. Checks unbiasedness for the paper's discounted gradient g_gamma. The paper explicitly acknowledges that discounting already changes the undiscounted objective; this is not an unacknowledged error.","primaryUrl":"https://arxiv.org/abs/1506.02438","noteUrl":"/review?paper=review-ir002","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir003","title":"Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor","authors":["Tuomas Haarnoja","Aurick Zhou","Pieter Abbeel","Sergey Levine"],"year":2018,"category":"Reinforcement learning","status":"no-error","issueTitle":"Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor","plainExplanation":"No issue was found in the checked claims. Checks Lemma 2 in the exact tabular setting. Does not certify Theorem 1's policy-sequence convergence for arbitrary restricted nonconvex classes or the practical approximate algorithm.","primaryUrl":"https://arxiv.org/abs/1801.01290","noteUrl":"/review?paper=review-ir003","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir004","title":"Reinforcement Learning with Deep Energy-Based Policies","authors":["Tuomas Haarnoja","Haoran Tang","Pieter Abbeel","Sergey Levine"],"year":2017,"category":"Reinforcement learning","status":"no-error","issueTitle":"Reinforcement Learning with Deep Energy-Based Policies","plainExplanation":"No issue was found in the checked claims. Checks the exact soft Bellman operator only; Monte Carlo partition estimates and amortised Stein sampling introduce approximation outside this contraction proof.","primaryUrl":"https://arxiv.org/abs/1702.08165","noteUrl":"/review?paper=review-ir004","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-ir005","title":"A Distributional Perspective on Reinforcement Learning","authors":["Marc G. Bellemare","Will Dabney","Rémi Munos"],"year":2017,"category":"Reinforcement learning","status":"addressed","issueTitle":"Bellman variance includes variation across successor means","plainExplanation":"Equal within-state variances do not imply equal variances after mixing states with different means. The authors’ later book gives the complete total-variance identity. A variance contraction follows on a common-mean class, not for arbitrary input distributions.","primaryUrl":"https://arxiv.org/abs/1707.06887","citationCount":1952,"citationUrl":"https://www.semanticscholar.org/paper/c1f4ef741242d629d1f56e442a09a7ba29595a0e","severity":"repairable","issueTypes":["formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-ir006","title":"Distributional Reinforcement Learning with Quantile Regression","authors":["Will Dabney","Mark Rowland","Marc G. Bellemare","Rémi Munos"],"year":2017,"category":"Reinforcement learning","status":"no-error","issueTitle":"Distributional Reinforcement Learning with Quantile Regression","plainExplanation":"No error in the central contraction result. The Lemma 2 literal equality F(theta)=midpoint does not describe every atomic-CDF minimiser, but its supplied generalised-inverse choice remains valid; this local set-characterisation issue is not promoted.","primaryUrl":"https://arxiv.org/abs/1710.10044","noteUrl":"/review?paper=review-ir006","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir007","title":"Soft Actor-Critic Algorithms and Applications","authors":["Tuomas Haarnoja","Aurick Zhou","Kristian Hartikainen","George Tucker","Sehoon Ha","Jie Tan","Vikash Kumar","Henry Zhu","Abhishek Gupta","Pieter Abbeel","Sergey Levine"],"year":2018,"category":"Reinforcement learning","status":"unresolved","issueTitle":"The exact temperature objective drops a temperature-dependent reward term","plainExplanation":"Dropping a temperature-dependent optimized reward changes the exact objective. A fixed-policy partial derivative gives the correct dual gradient, but it is not the derivative of the composed function printed inside the argmin.","primaryUrl":"https://arxiv.org/abs/1812.05905","noteUrl":"/review?paper=review-ir007","severity":"repairable","issueTypes":["formula_error","invalid_inference"],"classifications":["repairable"]},{"slug":"review-ir008","title":"Off-Policy Deep Reinforcement Learning without Exploration","authors":["Scott Fujimoto","David Meger","Doina Precup"],"year":2018,"category":"Reinforcement learning","status":"no-error","issueTitle":"Off-Policy Deep Reinforcement Learning without Exploration","plainExplanation":"No substantive error in these selected tabular claims. The neural generative-model approximation, stochastic finite-batch transition errors, and performance against policies leaving the batch are outside this guarantee.","primaryUrl":"https://arxiv.org/abs/1812.02900","noteUrl":"/review?paper=review-ir008","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"bear-suboptimality-constant","title":"Stabilizing Off-Policy Q-Learning via Bootstrapping Error Reduction","authors":["Aviral Kumar","Justin Fu","George Tucker","Sergey Levine"],"year":2019,"category":"Reinforcement learning","status":"finding","issueTitle":"The BEAR bound misses the cost of the current action","plainExplanation":"A Q-backup evaluates future rewards after the current action has already been chosen. Theorem 4.1 uses the difference between two such backups to bound the cost of restricting the policy. That difference can miss a reward lost by excluding the current action. Here both actions lead immediately to a zero-reward absorbing state, so the Q-backups agree. The constrained policy still loses the reward of one at the initial state.","primaryUrl":"https://arxiv.org/pdf/1906.00949v2","citationCount":1344,"citationUrl":"https://www.semanticscholar.org/paper/82b4b03a4659d6e04bd7cbf51d6e08fde1348dbd","severity":"repairable","issueTypes":["definition_error","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-ir010","title":"MOPO: Model-based Offline Policy Optimization","authors":["Tianhe Yu","Garrett Thomas","Lantao Yu","Stefano Ermon","James Zou","Sergey Levine","Chelsea Finn","Tengyu Ma"],"year":2020,"category":"Optimization & generalization","status":"no-error","issueTitle":"MOPO: Model-based Offline Policy Optimization","plainExplanation":"No issue was found in the checked claims. The exact identity and immediate absolute-error lower bound are valid. Whether ensemble uncertainty upper-bounds the required value-integral discrepancy is a separate practical assumption, not proved by this lemma.","primaryUrl":"https://arxiv.org/abs/2005.13239","noteUrl":"/review?paper=review-ir010","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir011","title":"When to Trust Your Model: Model-Based Policy Optimization","authors":["Michael Janner","Justin Fu","Marvin Zhang","Sergey Levine"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"When to Trust Your Model: Model-Based Policy Optimization","plainExplanation":"No issue was found in the checked claims. Main common-reward, TV version reconstructed. The KL-labelled auxiliary statement would need Pinsker's square root; the same paper supplies the intended TV condition, so this local mismatch is excluded. Reward-model error and branched-rollout theorems not certified by this selected review.","primaryUrl":"https://arxiv.org/abs/1906.08253","noteUrl":"/review?paper=review-ir011","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"combo-zero-penalty","title":"COMBO: Conservative Offline Model-Based Policy Optimization","authors":["Tianhe Yu","Aviral Kumar","Rafael Rafailov","Aravind Rajeswaran","Sergey Levine","Chelsea Finn"],"year":2021,"category":"Reinforcement learning","status":"finding","issueTitle":"A zero penalty cannot offset model bias","plainExplanation":"COMBO’s conservative term depends on the difference between model and data occupancies. When the occupancies coincide, that term vanishes, regardless of β. A positive learned-reward bias can nevertheless remain. The one-state example below therefore has an estimated value above the true value for every β. The preceding bound retains an explicit model-error term and is consistent with the example; the unconditional lower-bound corollary is not.","primaryUrl":"https://arxiv.org/pdf/2102.08363v2","citationCount":544,"citationUrl":"https://www.semanticscholar.org/paper/245682e8b3fa76f4a3e2991b5497577af95cbb3f","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-ir013","title":"MOReL : Model-Based Offline Reinforcement Learning","authors":["Rahul Kidambi","Aravind Rajeswaran","Praneeth Netrapalli","Thorsten Joachims"],"year":2020,"category":"Reinforcement learning","status":"no-error","issueTitle":"MOReL : Model-Based Offline Reinforcement Learning","plainExplanation":"No substantive error in Theorem1. Its guarantee depends on a valid unknown-state detector; ensemble disagreement in the practical implementation does not automatically establish that premise.","primaryUrl":"https://arxiv.org/abs/2005.05951","noteUrl":"/review?paper=review-ir013","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir014","title":"Constrained Policy Optimization","authors":["Joshua Achiam","David Held","Aviv Tamar","Pieter Abbeel"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Constrained Policy Optimization","plainExplanation":"No substantive error in this selected lemma. It does not by itself prove that practical constrained updates always satisfy safety constraints or solve the nonconvex neural problem.","primaryUrl":"https://arxiv.org/abs/1705.10528","noteUrl":"/review?paper=review-ir014","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir015","title":"Online Meta-Learning","authors":["Chelsea Finn","Aravind Rajeswaran","Sham Kakade","Sergey Levine"],"year":2019,"category":"Reinforcement learning","status":"no-error","issueTitle":"Online Meta-Learning","plainExplanation":"A proof gap is repairable with the actual sign information. Global bounded gradient and positive strong convexity on all R^d are incompatible; the paper itself mentions bounded-domain examples. No unaddressed substantive error retained, and no claim that arbitrary neural objectives meet these hypotheses.","primaryUrl":"https://arxiv.org/abs/1902.08438","noteUrl":"/review?paper=review-ir015","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir016","title":"Meta-Learning with Implicit Gradients","authors":["Aravind Rajeswaran","Chelsea Finn","Sham Kakade","Sergey Levine"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Meta-Learning with Implicit Gradients","plainExplanation":"No issue was found in the checked claims. Checks the implicit Jacobian, not the entire approximate-solver complexity theorem. The main guarantee's stronger assumptions address branch-selection ambiguity; no unaddressed finding retained.","primaryUrl":"https://arxiv.org/abs/1909.04630","noteUrl":"/review?paper=review-ir016","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir017","title":"On First-Order Meta-Learning Algorithms","authors":["Alex Nichol","Joshua Achiam","John Schulman"],"year":2018,"category":"Reinforcement learning","status":"no-error","issueTitle":"On First-Order Meta-Learning Algorithms","plainExplanation":"The leading-order algebra is correct under its Taylor/expectation regularity. A prose sentence gives the alignment direction the wrong sign, but the adjacent equations and algorithm supply the correct sign; this is not retained. No guarantee for nonsmooth crossing points or large steps follows.","primaryUrl":"https://arxiv.org/abs/1803.02999","noteUrl":"/review?paper=review-ir017","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir018","title":"Randomized Prior Functions for Deep Reinforcement Learning","authors":["Ian Osband","John Aslanides","Albin Cassirer"],"year":2018,"category":"Reinforcement learning","status":"no-error","issueTitle":"Randomized Prior Functions for Deep Reinforcement Learning","plainExplanation":"No issue was found in the checked claims. Exact in this linear conjugate case. The paper explicitly says the nonlinear neural construction is no longer an exact posterior, so that extension is not being certified or criticised as an unstated equivalence.","primaryUrl":"https://arxiv.org/abs/1806.03335","noteUrl":"/review?paper=review-ir018","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir019","title":"Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation","authors":["Emmanuel Bengio","Moksh Jain","Maksym Korablyov","Doina Precup","Yoshua Bengio"],"year":2021,"category":"Generative models","status":"no-error","issueTitle":"Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation","plainExplanation":"No substantive error in the finite-DAG sampling theorem. It requires exact flow conservation; finite training loss and neural function approximation do not automatically yield exact samples.","primaryUrl":"https://arxiv.org/abs/2106.04399","noteUrl":"/review?paper=review-ir019","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir020","title":"Direct Preference Optimization: Your Language Model is Secretly a Reward Model","authors":["Rafael Rafailov","Archit Sharma","Eric Mitchell","Stefano Ermon","Christopher D. Manning","Chelsea Finn"],"year":2023,"category":"Reinforcement learning","status":"no-error","issueTitle":"Direct Preference Optimization: Your Language Model is Secretly a Reward Model","plainExplanation":"No issue was found in the checked claims. Uniqueness and cancellation are valid. This is not a certificate that every unbounded reward on an unrestricted infinite response space is normalisable, or that finite preference data identifies a reward/policy globally.","primaryUrl":"https://arxiv.org/abs/2305.18290","noteUrl":"/review?paper=review-ir020","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ir021","title":"Generative Adversarial Imitation Learning","authors":["Jonathan Ho","Stefano Ermon"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Generative Adversarial Imitation Learning","plainExplanation":"No substantive error in this selected lemma. Policies at unreachable states are only defined up to arbitrary choices; strict uniqueness concerns occupancies. The entire IRL minimax theorem and attainment of unbounded cost optimisation are not certified.","primaryUrl":"https://arxiv.org/abs/1606.03476","noteUrl":"/review?paper=review-ir021","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-ir022","title":"Learning Robust Rewards with Adversarial Inverse Reinforcement Learning","authors":["Justin Fu","Katie Luo","Sergey Levine"],"year":2017,"category":"Reinforcement learning","status":"addressed","issueTitle":"The stochastic reward-transfer proof needs a stronger transition condition","plainExplanation":"Theorem 5.1 allows stochastic dynamics, but equal expected successor potentials do not imply equal potential values. Its decomposability assumption is insufficient for that step. A later analysis identifies the missing transition-kernel completeness condition in Section 3.3, Eqs. (45)-(46). The separate deterministic recovery argument in the original Appendix C is unaffected.","primaryUrl":"https://arxiv.org/abs/1710.11248","citationCount":882,"citationUrl":"https://www.semanticscholar.org/paper/5e2c4e7b3302549b3718601c44d9af6c7554efef","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-iarc-001","title":"Semi-Supervised Classification with Graph Convolutional Networks","authors":["Thomas N. Kipf","Max Welling"],"year":2016,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Semi-Supervised Classification with Graph Convolutional Networks","plainExplanation":"No error found in selected polynomial-localization argument.","primaryUrl":"https://arxiv.org/pdf/1609.02907","noteUrl":"/review?paper=review-iarc-001","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-002","title":"Inductive Representation Learning on Large Graphs","authors":["William L. Hamilton","Rex Ying","Jure Leskovec"],"year":2017,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Inductive Representation Learning on Large Graphs","plainExplanation":"No issue was found in the checked claims. A literal scalar reading of Algorithm 1’s final normalization cannot represent arbitrary clustering coefficients. Appendix E explicitly describes three GraphSAGE iterations followed by an MLP readout, supplying the intended output architecture. This is a same-paper clarification of the theorem–algorithm interpretation, not a later erratum or a claim that practical GraphSAGE lacks this expressive ability.","primaryUrl":"https://arxiv.org/pdf/1706.02216","noteUrl":"/review?paper=review-iarc-002","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-003","title":"How Attentive are Graph Attention Networks?","authors":["Shaked Brody","Uri Alon","Eran Yahav"],"year":2021,"category":"Graphs & equivariance","status":"no-error","issueTitle":"How Attentive are Graph Attention Networks?","plainExplanation":"No error found in selected theorem.","primaryUrl":"https://arxiv.org/pdf/2105.14491","noteUrl":"/review?paper=review-iarc-003","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-004","title":"Simplifying Graph Convolutional Networks","authors":["Felix Wu","Tianyi Zhang","Amauri Holanda de Souza","Christopher Fifty","Tao Yu","Kilian Q. Weinberger"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Simplifying Graph Convolutional Networks","plainExplanation":"No issue was found in the checked claims. Main selected theorem is supported. An unused sublemma upper-bound display is false but repairable; not retained as an eligible finding.","primaryUrl":"https://arxiv.org/pdf/1902.07153","noteUrl":"/review?paper=review-iarc-004","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-005","title":"Predict then Propagate: Graph Neural Networks meet Personalized PageRank","authors":["Johannes Gasteiger","Aleksandar Bojchevski","Stephan Günnemann"],"year":2018,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Predict then Propagate: Graph Neural Networks meet Personalized PageRank","plainExplanation":"No issue was found in the checked claims. Selected convergence result is valid; determinant wording is an insufficient criterion already repaired by the surrounding spectral argument.","primaryUrl":"https://arxiv.org/pdf/1810.05997","noteUrl":"/review?paper=review-iarc-005","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-006","title":"Principal Neighbourhood Aggregation for Graph Nets","authors":["Gabriele Corso","Luca Cavalleri","Dominique Beaini","Pietro Liò","Petar Veličković"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Principal Neighbourhood Aggregation for Graph Nets","plainExplanation":"No issue was found in the checked claims. Theorem 1 and the moment reconstruction are sound. The arbitrary-scaler statement needs the positive shift explicitly supplied in its own proof; not retained as an unaddressed failure.","primaryUrl":"https://arxiv.org/pdf/2004.05718","noteUrl":"/review?paper=review-iarc-006","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-007","title":"PairNorm: Tackling Oversmoothing in GNNs","authors":["Lingxiao Zhao","Leman Akoglu"],"year":2019,"category":"Transformers & sequence models","status":"no-error","issueTitle":"PairNorm: Tackling Oversmoothing in GNNs","plainExplanation":"No error found in the selected normalization identity on its natural nonzero-denominator domain.","primaryUrl":"https://arxiv.org/pdf/1909.12223","noteUrl":"/review?paper=review-iarc-007","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-008","title":"DropEdge: Towards Deep Graph Convolutional Networks on Node Classification","authors":["Yu Rong","Wenbing Huang","Tingyang Xu","Junzhou Huang"],"year":2019,"category":"Graphs & equivariance","status":"unresolved","issueTitle":"An increasing eigenvalue lower bound does not prove slower smoothing","plainExplanation":"An increase in a lower bound does not establish an increase in the quantity bounded. Reaching the disconnected case does not prove monotonicity before disconnection. The proof also needs to compare initial distances to the two different stationary subspaces.","primaryUrl":"https://arxiv.org/pdf/1907.10903","noteUrl":"/review?paper=review-iarc-008","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-iarc-009","title":"Deeper Insights into Graph Convolutional Networks for Semi-Supervised Learning","authors":["Qimai Li","Zhichao Han","Xiao-Ming Wu"],"year":2018,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Deeper Insights into Graph Convolutional Networks for Semi-Supervised Learning","plainExplanation":"No issue was found in the checked claims. The displayed symmetric-smoothing limit uses an inverse square-root degree factor, whereas its stationary vectors require the positive square-root factor. The prose immediately preceding Theorem 1 explicitly gives the square-root dependence, clarifying the intended limit. This same-paper clarification preserves the central linear oversmoothing conclusion; it is not a documented later erratum.","primaryUrl":"https://arxiv.org/pdf/1801.07606","noteUrl":"/review?paper=review-iarc-009","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-010","title":"PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation","authors":["Charles R. Qi","Hao Su","Kaichun Mo","Leonidas J. Guibas"],"year":2016,"category":"Transformers & sequence models","status":"no-error","issueTitle":"PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation","plainExplanation":"No error found in selected critical-set theorem after reading set-type notation as subsets of the ambient input domain.","primaryUrl":"https://arxiv.org/pdf/1612.00593","noteUrl":"/review?paper=review-iarc-010","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-011","title":"Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering","authors":["Michaël Defferrard","Xavier Bresson","Pierre Vandergheynst"],"year":2016,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering","plainExplanation":"No error found in the selected localization, recurrence, or gradient derivations.","primaryUrl":"https://arxiv.org/pdf/1606.09375","noteUrl":"/review?paper=review-iarc-011","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-012","title":"Understanding over-squashing and bottlenecks on graphs via curvature","authors":["Jake Topping","Francesco Di Giovanni","Benjamin Paul Chamberlain","Xiaowen Dong","Michael M. Bronstein"],"year":2021,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Understanding over-squashing and bottlenecks on graphs via curvature","plainExplanation":"No issue was found in the checked claims. Unroll the chain rule across r+1 layers. Reaching s at distance r+1 requires moving across an edge at every layer; any self-state derivative term would allow fewer than r+1 moves and cannot reach s. Each remaining shortest-walk term has r+1 factors from each update/message derivative, hence magnitude≤(αβ)^(r+1) times the product of edge coefficients. Summing nonnegative path weights is exactly (Ahat^(r+1))is.","primaryUrl":"https://arxiv.org/pdf/2111.14522","noteUrl":"/review?paper=review-iarc-012","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-013","title":"Tensor Field Networks: Rotation- and Translation-Equivariant Neural Networks for 3D Point Clouds","authors":["Nathaniel Thomas","Tess Smidt","Steven Kearnes","Lusann Yang","Li Li","Kai Kohlhoff","Patrick Riley"],"year":2018,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Tensor Field Networks: Rotation- and Translation-Equivariant Neural Networks for 3D Point Clouds","plainExplanation":"No error found in the selected point-convolution equivariance proof.","primaryUrl":"https://arxiv.org/pdf/1802.08219","noteUrl":"/review?paper=review-iarc-013","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-014","title":"SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks","authors":["Fabian B. Fuchs","Daniel E. Worrall","Volker Fischer","Max Welling"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks","plainExplanation":"No error found in the selected attention-equivariance argument.","primaryUrl":"https://arxiv.org/pdf/2006.10503","noteUrl":"/review?paper=review-iarc-014","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-015","title":"3D Steerable CNNs: Learning Rotationally Equivariant Features in Volumetric Data","authors":["Maurice Weiler","Mario Geiger","Max Welling","Wouter Boomsma","Taco Cohen"],"year":2018,"category":"Graphs & equivariance","status":"no-error","issueTitle":"3D Steerable CNNs: Learning Rotationally Equivariant Features in Volumetric Data","plainExplanation":"No error found in the selected integral-kernel lemma and its conditional cross-correlation reduction.","primaryUrl":"https://arxiv.org/pdf/1807.02547","noteUrl":"/review?paper=review-iarc-015","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-016","title":"Gauge Equivariant Convolutional Networks and the Icosahedral CNN","authors":["Taco S. Cohen","Maurice Weiler","Berkay Kicanaoglu","Max Welling"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Gauge Equivariant Convolutional Networks and the Icosahedral CNN","plainExplanation":"No error found in the selected gauge-equivariance derivation.","primaryUrl":"https://arxiv.org/pdf/1902.04615","noteUrl":"/review?paper=review-iarc-016","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-017","title":"Generalizing Convolutional Neural Networks for Equivariance to Lie Groups on Arbitrary Continuous Data","authors":["Marc Finzi","Samuel Stanton","Pavel Izmailov","Andrew Gordon Wilson"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Generalizing Convolutional Neural Networks for Equivariance to Lie Groups on Arbitrary Continuous Data","plainExplanation":"No issue was found in the checked claims. Selected equivariance-in-distribution result follows with the correct transformed local sampling law. The displayed vtilde definition has the inverse direction reversed; not retained as a substantive theorem failure.","primaryUrl":"https://arxiv.org/pdf/2002.12880","noteUrl":"/review?paper=review-iarc-017","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-018","title":"Spherical CNNs","authors":["Taco S. Cohen","Mario Geiger","Jonas Koehler","Max Welling"],"year":2018,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Spherical CNNs","plainExplanation":"No error found in the selected continuous correlation equivariance proof.","primaryUrl":"https://arxiv.org/pdf/1801.10130","noteUrl":"/review?paper=review-iarc-018","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-019","title":"Harmonic Networks: Deep Translation and Rotation Equivariance","authors":["Daniel E. Worrall","Stephan J. Garbin","Daniyar Turmukhambetov","Gabriel J. Brostow"],"year":2016,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Harmonic Networks: Deep Translation and Rotation Equivariance","plainExplanation":"No error found in the selected transformation and chaining laws.","primaryUrl":"https://arxiv.org/pdf/1612.04642","noteUrl":"/review?paper=review-iarc-019","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-020","title":"Vector Neurons: A General Framework for SO(3)-Equivariant Networks","authors":["Congyue Deng","Or Litany","Yueqi Duan","Adrien Poulenard","Andrea Tagliasacchi","Leonidas Guibas"],"year":2021,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Vector Neurons: A General Framework for SO(3)-Equivariant Networks","plainExplanation":"No error found in the selected vector-neuron equivariance argument.","primaryUrl":"https://arxiv.org/pdf/2104.12229","noteUrl":"/review?paper=review-iarc-020","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-021","title":"On the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact Groups","authors":["Risi Kondor","Shubhendu Trivedi"],"year":2018,"category":"Graphs & equivariance","status":"unresolved","issueTitle":"Equivariance of reachable activations does not force every weight map to be a convolution","plainExplanation":"Equivariance of activations reached from network inputs constrains later weights only on reachable hidden states. A convolutional linear map must intertwine the group actions on the whole layer space.","primaryUrl":"https://arxiv.org/pdf/1802.03690","noteUrl":"/review?paper=review-iarc-021","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-iarc-022","title":"e3nn: Euclidean Neural Networks","authors":["Mario Geiger","Tess Smidt"],"year":2022,"category":"Transformers & sequence models","status":"no-error","issueTitle":"e3nn: Euclidean Neural Networks","plainExplanation":"No error found in selected lemma.","primaryUrl":"https://arxiv.org/pdf/2207.09453","noteUrl":"/review?paper=review-iarc-022","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-023","title":"General E(2)-Equivariant Steerable CNNs","authors":["Maurice Weiler","Gabriele Cesa"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"General E(2)-Equivariant Steerable CNNs","plainExplanation":"No error found in the selected convolution proof.","primaryUrl":"https://arxiv.org/pdf/1911.08251","noteUrl":"/review?paper=review-iarc-023","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-024","title":"Equivariant message passing for the prediction of tensorial properties and molecular spectra","authors":["Kristof T. Schütt","Oliver T. Unke","Michael Gastegger"],"year":2021,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Equivariant message passing for the prediction of tensorial properties and molecular spectra","plainExplanation":"No error found in the selected architecture derivation.","primaryUrl":"https://arxiv.org/pdf/2102.03150","noteUrl":"/review?paper=review-iarc-024","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-025","title":"Learning Steerable Filters for Rotation Equivariant CNNs","authors":["Maurice Weiler","Fred A. Hamprecht","Martin Storath"],"year":2017,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Learning Steerable Filters for Rotation Equivariant CNNs","plainExplanation":"No error found in the selected continuous-domain proof.","primaryUrl":"https://arxiv.org/pdf/1711.07289","noteUrl":"/review?paper=review-iarc-025","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-026","title":"Directional Message Passing for Molecular Graphs","authors":["Johannes Gasteiger","Janek Groß","Stephan Günnemann"],"year":2020,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Directional Message Passing for Molecular Graphs","plainExplanation":"No error found in these selected derivations under the spherical measure and normalized-coordinate convention.","primaryUrl":"https://arxiv.org/pdf/2003.03123","noteUrl":"/review?paper=review-iarc-026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-iarc-027","title":"GemNet: Universal Directional Graph Neural Networks for Molecules","authors":["Johannes Gasteiger","Florian Becker","Stephan Günnemann"],"year":2021,"category":"Molecular models & equivariance","status":"addressed","issueTitle":"A zero Clebsch-Gordan coefficient invalidates the spherical-product equivalence","plainExplanation":"Appendix A treats multiplication by a Clebsch-Gordan coefficient as preserving expressivity, but that coefficient is zero when the three angular degrees sum to an odd number. This removes interactions needed by some SO(3)-invariant scalar functions. The Price of Freedom independently derives the same selection rule for spherical products; it does not present a GemNet erratum.","primaryUrl":"https://arxiv.org/abs/2106.08903","citationCount":651,"citationUrl":"https://www.semanticscholar.org/paper/5542d0ff99767f75f8c8a329fc3d88d73ff470c3","severity":"critical","issueTypes":["false_statement","invalid_inference","proof_gap"],"classifications":["critical"]},{"slug":"review-iarc-028","title":"E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials","authors":["Simon Batzner","Albert Musaelian","Lixin Sun","Mario Geiger","Jonathan P. Mailoa","Mordechai Kornbluth","Nicola Molinari","Tess E. Smidt","Boris Kozinsky"],"year":2021,"category":"Graphs & equivariance","status":"no-error","issueTitle":"E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials","plainExplanation":"No error found in the selected tensor-convolution argument.","primaryUrl":"https://arxiv.org/pdf/2101.03164","noteUrl":"/review?paper=review-iarc-028","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-029","title":"Learning local equivariant representations for large-scale atomistic dynamics","authors":["Albert Musaelian","Simon Batzner","Anders Johansson","Lixin Sun","Cameron J. Owen","Mordechai Kornbluth","Boris Kozinsky"],"year":2022,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Learning local equivariant representations for large-scale atomistic dynamics","plainExplanation":"No error found in the selected recursive expansion.","primaryUrl":"https://arxiv.org/pdf/2204.05249","noteUrl":"/review?paper=review-iarc-029","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-030","title":"Equiformer: Equivariant Graph Attention Transformer for 3D Atomistic Graphs","authors":["Yi-Lun Liao","Tess Smidt"],"year":2022,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Equiformer: Equivariant Graph Attention Transformer for 3D Atomistic Graphs","plainExplanation":"No error found in the selected equivariance argument.","primaryUrl":"https://arxiv.org/pdf/2206.11990","noteUrl":"/review?paper=review-iarc-030","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-031","title":"MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing","authors":["Sami Abu-El-Haija","Bryan Perozzi","Amol Kapoor","Nazanin Alipourfard","Kristina Lerman","Hrayr Harutyunyan","Greg Ver Steeg","Aram Galstyan"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing","plainExplanation":"No issue was found in the checked claims. The selected expressivity statements withstand checking; one final proof sentence needs the explicit equal-output argument reconstructed here.","primaryUrl":"https://arxiv.org/pdf/1905.00067","noteUrl":"/review?paper=review-iarc-031","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-032","title":"Adaptive Universal Generalized PageRank Graph Neural Network","authors":["Eli Chien","Jianhao Peng","Pan Li","Olgica Milenkovic"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Adaptive Universal Generalized PageRank Graph Neural Network","plainExplanation":"No error found in the selected formal theorem.","primaryUrl":"https://arxiv.org/pdf/2006.07988","noteUrl":"/review?paper=review-iarc-032","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-033","title":"Simple and Deep Graph Convolutional Networks","authors":["Ming Chen","Zhewei Wei","Zengfeng Huang","Bolin Ding","Yaliang Li"],"year":2020,"category":"Graphs & equivariance","status":"unresolved","issueTitle":"Final ReLU features cannot represent arbitrary signed polynomial filters","plainExplanation":"ReLU makes every final hidden coordinate nonnegative, whereas an arbitrary polynomial filter can produce a negative output from positive input. The proof drops ReLU after checking the input sign, but its learned scalar multipliers can still make a preactivation negative.","primaryUrl":"https://arxiv.org/pdf/2007.02133","noteUrl":"/review?paper=review-iarc-033","severity":"repairable","issueTypes":["false_statement","invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-iarc-034","title":"Beyond Low-frequency Information in Graph Convolutional Networks","authors":["Deyu Bo","Xiao Wang","Chuan Shi","Huawei Shen"],"year":2021,"category":"Graphs & equivariance","status":"unresolved","issueTitle":"GCN averaging need not shorten every adjacent-pair distance","plainExplanation":"Equal degrees do not remove the other neighbors’ contributions. Those contributions can increase the distance between an adjacent pair, even with unit weight and nonnegative features.","primaryUrl":"https://arxiv.org/pdf/2101.00797","noteUrl":"/review?paper=review-iarc-034","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-iarc-035","title":"Diffusion Improves Graph Learning","authors":["Johannes Gasteiger","Stefan Weißenberger","Stephan Günnemann"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Diffusion Improves Graph Learning","plainExplanation":"No error found in the selected derivation under the symmetric analysis convention.","primaryUrl":"https://arxiv.org/pdf/1911.05485","noteUrl":"/review?paper=review-iarc-035","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-036","title":"Adaptive Sampling Towards Fast Graph Representation Learning","authors":["Wenbing Huang","Tong Zhang","Yu Rong","Junzhou Huang"],"year":2018,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Adaptive Sampling Towards Fast Graph Representation Learning","plainExplanation":"No error found in the selected fixed-parent result for nonnegative scalar features.","primaryUrl":"https://arxiv.org/pdf/1809.05343","noteUrl":"/review?paper=review-iarc-036","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-037","title":"BernNet: Learning Arbitrary Graph Spectral Filters via Bernstein Approximation","authors":["Mingguo He","Zhewei Wei","Zengfeng Huang","Hongteng Xu"],"year":2021,"category":"Graphs & equivariance","status":"unresolved","issueTitle":"BernNet: Learning Arbitrary Graph Spectral Filters via Bernstein Approximation","plainExplanation":"An interior zero prevents exact finite representation by nonnegative Bernstein coefficients. Uniform approximation remains valid, and gives the appropriate correction to the filter-class statement.","primaryUrl":"https://arxiv.org/pdf/2106.10994","noteUrl":"/review?paper=review-iarc-037","severity":"repairable","issueTypes":["citation_hypothesis_mismatch","false_statement","lemma_error"],"classifications":["repairable"]},{"slug":"review-iarc-038","title":"FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling","authors":["Jie Chen","Tengfei Ma","Cao Xiao"],"year":2018,"category":"Graphs & equivariance","status":"finding","issueTitle":"FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling","plainExplanation":"The proof sums conditional variances as though the outer estimates had independent inner samples. Their shared inner sample induces covariance even when the outer vertices are independent.","primaryUrl":"https://arxiv.org/pdf/1801.10247","noteUrl":"/review?paper=review-iarc-038","severity":"major","issueTypes":[],"classifications":["major"]},{"slug":"review-iarc-039","title":"Layer-Dependent Importance Sampling for Training Deep and Large Graph Convolutional Networks","authors":["Difan Zou","Ziniu Hu","Yewen Wang","Song Jiang","Yizhou Sun","Quanquan Gu"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Layer-Dependent Importance Sampling for Training Deep and Large Graph Convolutional Networks","plainExplanation":"No error found in the selected sketch-variance lemma on its active support.","primaryUrl":"https://arxiv.org/pdf/1911.07323","noteUrl":"/review?paper=review-iarc-039","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-040","title":"GraphSAINT: Graph Sampling Based Inductive Learning Method","authors":["Hanqing Zeng","Hongkuan Zhou","Ajitesh Srivastava","Rajgopal Kannan","Viktor Prasanna"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Variance-optimal edge probabilities require a cap at one","plainExplanation":"The proportional allocation solves the problem without the Bernoulli constraints \\(p_e\\le1\\). It can therefore assign an impossible probability. Saturating a sampler’s input is not in itself a proof of the budget-constrained optimum.","primaryUrl":"https://arxiv.org/pdf/1907.04931","noteUrl":"/review?paper=review-iarc-040","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-iarc-041","title":"Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks","authors":["Wei-Lin Chiang","Xuanqing Liu","Si Si","Yang Li","Samy Bengio","Cho-Jui Hsieh"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks","plainExplanation":"No error found in the selected block-decomposition derivation.","primaryUrl":"https://arxiv.org/pdf/1905.07953","noteUrl":"/review?paper=review-iarc-041","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-042","title":"What graph neural networks cannot learn: depth vs width","authors":["Andreas Loukas"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"What graph neural networks cannot learn: depth vs width","plainExplanation":"No error found in the selected idealized equivalence theorem with the stated computational assumptions.","primaryUrl":"https://arxiv.org/pdf/1907.03199","noteUrl":"/review?paper=review-iarc-042","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-043","title":"Identity-aware Graph Neural Networks","authors":["Jiaxuan You","Jonathan Gomes-Selman","Rex Ying","Jure Leskovec"],"year":2021,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Identity-aware Graph Neural Networks","plainExplanation":"No issue was found in the checked claims. The constructed recursion correctly counts walks/closed walks. Its terminology should not be interpreted as arbitrary simple-cycle counting.","primaryUrl":"https://arxiv.org/pdf/2101.10320","noteUrl":"/review?paper=review-iarc-043","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-044","title":"Graph Random Neural Networks for Semi-Supervised Learning on Graphs","authors":["Wenzheng Feng","Jie Zhang","Yuxiao Dong","Yu Han","Huanbo Luan","Qian Xu","Qiang Yang","Evgeny Kharlamov","Jie Tang"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Graph Random Neural Networks for Semi-Supervised Learning on Graphs","plainExplanation":"No error found in the selected exact decomposition and expressly approximate Taylor calculations.","primaryUrl":"https://arxiv.org/pdf/2005.11079","noteUrl":"/review?paper=review-iarc-044","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-045","title":"Do Transformers Really Perform Bad for Graph Representation?","authors":["Chengxuan Ying","Tianle Cai","Shengjie Luo","Shuxin Zheng","Guolin Ke","Di He","Yanming Shen","Tie-Yan Liu"],"year":2021,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Do Transformers Really Perform Bad for Graph Representation?","plainExplanation":"No issue was found in the checked claims. Main aggregation construction is valid in the stated approximation/ideal-mask sense. The Fact2 proof contains a local false step; its conclusion has a direct parameter-choice repair.","primaryUrl":"https://arxiv.org/pdf/2106.05234","noteUrl":"/review?paper=review-iarc-045","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-046","title":"Recipe for a General, Powerful, Scalable Graph Transformer","authors":["Ladislav Rampášek","Mikhail Galkin","Vijay Prakash Dwivedi","Anh Tuan Luu","Guy Wolf","Dominique Beaini"],"year":2022,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Recipe for a General, Powerful, Scalable Graph Transformer","plainExplanation":"No issue was found in the checked claims. The conditional existence claim withstands checking, with an added presence coordinate needed for its all-zero-feature example.","primaryUrl":"https://arxiv.org/pdf/2205.12454","noteUrl":"/review?paper=review-iarc-046","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-047","title":"Pure Transformers are Powerful Graph Learners","authors":["Jinwoo Kim","Tien Dat Nguyen","Seonwoo Min","Sungjun Cho","Moontae Lee","Honglak Lee","Seunghoon Hong"],"year":2022,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Pure Transformers are Powerful Graph Learners","plainExplanation":"No error found in the selected normalized nonzero-row approximation argument.","primaryUrl":"https://arxiv.org/pdf/2207.02505","noteUrl":"/review?paper=review-iarc-047","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-048","title":"Hierarchical Graph Representation Learning with Differentiable Pooling","authors":["Rex Ying","Jiaxuan You","Christopher Morris","Xiang Ren","William L. Hamilton","Jure Leskovec"],"year":2018,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Hierarchical Graph Representation Learning with Differentiable Pooling","plainExplanation":"No issue was found in the checked claims. The intended pooling invariance result is correct. The displayed hypothesis uses invariant where equivariant is required and omits a feature permutation; the orthogonality calculation identifies the intended formulation.","primaryUrl":"https://arxiv.org/pdf/1806.08804","noteUrl":"/review?paper=review-iarc-048","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-049","title":"Rethinking Graph Transformers with Spectral Attention","authors":["Devin Kreuzer","Dominique Beaini","William L. Hamilton","Vincent Létourneau","Prudencio Tossou"],"year":2021,"category":"Graphs & equivariance","status":"unresolved","issueTitle":"Ambient Lp approximation does not control graph classification error","plainExplanation":"Small ambient integral error need not give small error on a finite set of graph encodings, or on a graph distribution supported there. Acknowledging possible mislabelling does not supply a relation between the approximation norm and the graph error. The displayed composition is also ill-typed for the maps defined immediately above it.","primaryUrl":"https://arxiv.org/pdf/2106.03893","noteUrl":"/review?paper=review-iarc-049","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-iarc-050","title":"Reformer: The Efficient Transformer","authors":["Nikita Kitaev","Łukasz Kaiser","Anselm Levskaya"],"year":2020,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Reformer: The Efficient Transformer","plainExplanation":"No issue was found in the checked claims. Selected reversibility and chunking argument checks out.","primaryUrl":"https://arxiv.org/pdf/2001.04451","noteUrl":"/review?paper=review-iarc-050","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-051","title":"Efficient Attention: Attention with Linear Complexities","authors":["Zhuoran Shen","Mingyuan Zhang","Haiyu Zhao","Shuai Yi","Hongsheng Li"],"year":2018,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Efficient Attention: Attention with Linear Complexities","plainExplanation":"No issue was found in the checked claims. Exact scaled-attention equivalence is sound.","primaryUrl":"https://arxiv.org/pdf/1812.01243","noteUrl":"/review?paper=review-iarc-051","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-052","title":"RoFormer: Enhanced Transformer with Rotary Position Embedding","authors":["Jianlin Su","Yu Lu","Shengfeng Pan","Ahmed Murtadha","Bo Wen","Yunfeng Liu"],"year":2021,"category":"Transformers & sequence models","status":"unresolved","issueTitle":"Finite rotary frequencies do not imply decay at all long distances","plainExplanation":"A finite sum of fixed-frequency complex exponentials is recurrent. Its phases need not cancel at every large distance, so the valid Abel inequality does not establish an unrestricted long-distance decay property.","primaryUrl":"https://arxiv.org/pdf/2104.09864","noteUrl":"/review?paper=review-iarc-052","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-iarc-053","title":"Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention","authors":["Angelos Katharopoulos","Apoorv Vyas","Nikolaos Pappas","François Fleuret"],"year":2020,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention","plainExplanation":"No issue was found in the checked claims. Selected gradient proof is sound.","primaryUrl":"https://arxiv.org/pdf/2006.16236","noteUrl":"/review?paper=review-iarc-053","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-054","title":"Linear Transformers Are Secretly Fast Weight Programmers","authors":["Imanol Schlag","Kazuki Irie","Jürgen Schmidhuber"],"year":2021,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Linear Transformers Are Secretly Fast Weight Programmers","plainExplanation":"No issue was found in the checked claims. The selected conditional overwrite and normalization derivations check out.","primaryUrl":"https://arxiv.org/pdf/2102.11174","noteUrl":"/review?paper=review-iarc-054","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-055","title":"FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness","authors":["Tri Dao","Daniel Y. Fu","Stefano Ermon","Atri Rudra","Christopher Ré"],"year":2022,"category":"Transformers & sequence models","status":"no-error","issueTitle":"FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness","plainExplanation":"No issue was found in the checked claims. Selected exactness and FLOP/auxiliary-memory proof is sound.","primaryUrl":"https://arxiv.org/pdf/2205.14135","noteUrl":"/review?paper=review-iarc-055","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-056","title":"Nyströmformer: A Nystöm-based Algorithm for Approximating Self-Attention","authors":["Yunyang Xiong","Zhanpeng Zeng","Rudrasis Chakraborty","Mingxing Tan","Glenn Fung","Yin Li","Vikas Singh"],"year":2021,"category":"Transformers & sequence models","status":"unresolved","issueTitle":"The pseudoinverse iteration needs a range-compatible initialization","plainExplanation":"The residual condition does not detect components of the initial iterate in the nullspace of \\(A\\). The recurrence can amplify those components while the tested residual stays zero.","primaryUrl":"https://arxiv.org/pdf/2102.03902","noteUrl":"/review?paper=review-iarc-056","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-iarc-057","title":"Graph Neural Networks Exponentially Lose Expressive Power for Node Classification","authors":["Kenta Oono","Taiji Suzuki"],"year":2019,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Graph Neural Networks Exponentially Lose Expressive Power for Node Classification","plainExplanation":"No issue was found in the checked claims. The stated selected contraction is valid; there is a local algebraic sign typo in the longer proof which does not refute it.","primaryUrl":"https://arxiv.org/pdf/1905.10947","noteUrl":"/review?paper=review-iarc-057","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-058","title":"Learning Important Features Through Propagating Activation Differences","authors":["Avanti Shrikumar","Peyton Greenside","Anshul Kundaje"],"year":2017,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Learning Important Features Through Propagating Activation Differences","plainExplanation":"No issue was found in the checked claims. The selected conservation proof is correct on its defined domain.","primaryUrl":"https://arxiv.org/pdf/1704.02685","noteUrl":"/review?paper=review-iarc-058","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-059","title":"On the (In)fidelity and Sensitivity for Explanations.","authors":["Chih-Kuan Yeh","Cheng-Yu Hsieh","Arun Sai Suggala","David I. Inouye","Pradeep Ravikumar"],"year":2019,"category":"Transformers & sequence models","status":"no-error","issueTitle":"On the (In)fidelity and Sensitivity for Explanations.","plainExplanation":"No issue was found in the checked claims. The selected least-squares proof is sound with the stated invertibility and integrability conditions.","primaryUrl":"https://arxiv.org/pdf/1901.09392","noteUrl":"/review?paper=review-iarc-059","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-iarc-060","title":"Axiomatic Attribution for Deep Networks","authors":["Mukund Sundararajan","Ankur Taly","Qiqi Yan"],"year":2017,"category":"Explainability","status":"addressed","issueTitle":"The stated completeness assumptions and path-uniqueness claim need correction","plainExplanation":"Almost-everywhere differentiability in the ambient space is insufficient for the stated completeness result along a chosen path. The symmetry definition also leaves paths unconstrained when the symmetric coordinates differ at an endpoint, so it does not establish a unique path method. Lerma and Lucas identify both problems and give precise replacements. Integrated gradients retains its completeness guarantee when the required path regularity holds.","primaryUrl":"https://arxiv.org/abs/1703.01365","citationCount":8651,"citationUrl":"https://www.semanticscholar.org/paper/f302e136c41db5de1d624412f68c9174cf7ae8be","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-iarc-061","title":"A Unified Approach to Interpreting Model Predictions","authors":["Scott Lundberg","Su-In Lee"],"year":2017,"category":"Transformers & sequence models","status":"no-error","issueTitle":"A Unified Approach to Interpreting Model Predictions","plainExplanation":"No issue was found in the checked claims. The selected Shapley-kernel characterization is valid with endpoint constraints.","primaryUrl":"https://arxiv.org/pdf/1705.07874","noteUrl":"/review?paper=review-iarc-061","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-062","title":"GNNExplainer: Generating Explanations for Graph Neural Networks","authors":["Rex Ying","Dylan Bourgeois","Jiaxuan You","Marinka Zitnik","Jure Leskovec"],"year":2019,"category":"Graphs & equivariance","status":"unresolved","issueTitle":"The Jensen substitution is a lower bound under convexity","plainExplanation":"The bound label reverses Jensen’s inequality. Under the stated convexity assumption, evaluating at the mean gives a lower bound on the expected value.","primaryUrl":"https://arxiv.org/pdf/1903.03894","noteUrl":"/review?paper=review-iarc-062","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-iarc-063","title":"Parameterized Explainer for Graph Neural Network","authors":["Dongsheng Luo","Wei Cheng","Dongkuan Xu","Wenchao Yu","Bo Zong","Haifeng Chen","Xiang Zhang"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Parameterized Explainer for Graph Neural Network","plainExplanation":"No issue was found in the checked claims. The intended relaxed sampling limit and Monte Carlo objective are sound; the displayed point-mass limit notation is imprecise.","primaryUrl":"https://arxiv.org/pdf/2011.04573","noteUrl":"/review?paper=review-iarc-063","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-064","title":"On Explainability of Graph Neural Networks via Subgraph Explorations","authors":["Hao Yuan","Haiyang Yu","Jie Wang","Kang Li","Shuiwang Ji"],"year":2021,"category":"Graphs & equivariance","status":"unresolved","issueTitle":"On Explainability of Graph Neural Networks via Subgraph Explorations","plainExplanation":"No issue was found in the checked claims. The factorial weights in Eq. (6) are correct, and the released mc_shapley and mc_l_shapley functions sample predecessor coalitions of uniform random player permutations. The proposed uniform-subset objection therefore does not describe these implementations. Section 3.5 explicitly introduces a local approximation, without claiming an exact global-error bound.","primaryUrl":"https://arxiv.org/pdf/2102.05152","noteUrl":"/review?paper=review-iarc-064","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-065","title":"Hopfield Networks is All You Need","authors":["Hubert Ramsauer","Bernhard Schäfl","Johannes Lehner","Philipp Seidl","Michael Widrich","Thomas Adler","Lukas Gruber","Markus Holzleitner","Milena Pavlović","Geir Kjetil Sandve","Victor Greiff","David Kreil","Michael Kopp","Günter Klambauer","Johannes Brandstetter","Sepp Hochreiter"],"year":2020,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Hopfield Networks is All You Need","plainExplanation":"No issue was found in the checked claims. The selected energy-convergence proof is sound.","primaryUrl":"https://arxiv.org/pdf/2008.02217","noteUrl":"/review?paper=review-iarc-065","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-066","title":"Rethinking Attention with Performers","authors":["Krzysztof Choromanski","Valerii Likhosherstov","David Dohan","Xingyou Song","Andreea Gane","Tamas Sarlos","Peter Hawkins","Jared Davis","Afroz Mohiuddin","Lukasz Kaiser","David Belanger","Lucy Colwell","Adrian Weller"],"year":2020,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Rethinking Attention with Performers","plainExplanation":"No issue was found in the checked claims. The selected unbiasedness and MSE calculations are sound.","primaryUrl":"https://arxiv.org/pdf/2009.14794","noteUrl":"/review?paper=review-iarc-066","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-067","title":"Linformer: Self-Attention with Linear Complexity","authors":["Sinong Wang","Belinda Z. Li","Madian Khabsa","Han Fang","Hao Ma"],"year":2020,"category":"Transformers & sequence models","status":"unresolved","issueTitle":"Linformer: Self-Attention with Linear Complexity","plainExplanation":"The relative inner-product estimate requires an exact zero after projection whenever the original inner product is zero. Gaussian projection does not preserve such zeros. The exponential-sketch argument has a separate failure: the proposed sketch produces a centered random output even on zero logits and constant values, whereas exact normalized attention returns one.","primaryUrl":"https://arxiv.org/pdf/2006.04768","noteUrl":"/review?paper=review-iarc-067","severity":"major","issueTypes":["incorrect_bound","invalid_inference","proof_gap"],"classifications":["major"]},{"slug":"review-iarc-068","title":"On Layer Normalization in the Transformer Architecture","authors":["Ruibin Xiong","Yunchang Yang","Di He","Kai Zheng","Shuxin Zheng","Chen Xing","Huishuai Zhang","Yanyan Lan","Liwei Wang","Tie-Yan Liu"],"year":2020,"category":"Transformers & sequence models","status":"unresolved","issueTitle":"The layer-normalization Jacobian depends on the centered norm","plainExplanation":"Layer normalization removes the mean. Adding an arbitrarily large constant to every coordinate changes the raw norm but leaves the normalized vector and its Jacobian unchanged. The final uncentered-norm bound cannot hold uniformly for all inputs.","primaryUrl":"https://arxiv.org/pdf/2002.04745","noteUrl":"/review?paper=review-iarc-068","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-iarc-069","title":"Understanding the Difficulty of Training Transformers","authors":["Liyuan Liu","Xiaodong Liu","Jianfeng Gao","Weizhu Chen","Jiawei Han"],"year":2020,"category":"Transformers & sequence models","status":"unresolved","issueTitle":"The variance derivation omits possible cross-layer covariances","plainExplanation":"Successive modules share upstream random inputs. Independent weights alone do not imply independent outputs or perturbations, so they do not justify omitting the covariance terms in the derivation.","primaryUrl":"https://arxiv.org/pdf/2004.08249","noteUrl":"/review?paper=review-iarc-069","severity":"repairable","issueTypes":["invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-iarc-070","title":"Transformers learn in-context by gradient descent","authors":["Johannes von Oswald","Eyvind Niklasson","Ettore Randazzo","João Sacramento","Alexander Mordvintsev","Andrey Zhmoginov","Max Vladymyrov"],"year":2022,"category":"Optimization & generalization","status":"no-error","issueTitle":"Transformers learn in-context by gradient descent","plainExplanation":"No issue was found in the checked claims. The selected constructed update is algebraically correct.","primaryUrl":"https://arxiv.org/pdf/2212.07677","noteUrl":"/review?paper=review-iarc-070","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-071","title":"What learning algorithm is in-context learning? Investigations with linear models","authors":["Ekin Akyürek","Dale Schuurmans","Jacob Andreas","Tengyu Ma","Denny Zhou"],"year":2022,"category":"Transformers & sequence models","status":"no-error","issueTitle":"What learning algorithm is in-context learning? Investigations with linear models","plainExplanation":"No issue was found in the checked claims. The multiplication construction checks out as an approximation; the division primitive needs the positive-denominator restriction.","primaryUrl":"https://arxiv.org/pdf/2211.15661","noteUrl":"/review?paper=review-iarc-071","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"implicit-bayesian-icl-prior-shift","title":"An Explanation of In-context Learning as Implicit Bayesian Inference","authors":["Sang Michael Xie","Aditi Raghunathan","Percy Liang","Tengyu Ma"],"year":2021,"category":"Transformers & in-context learning","status":"finding","issueTitle":"Close priors can give different conditional predictions","plainExplanation":"Two hidden-state priors can be close before observing an input and far apart after conditioning on it. Assumption 3 controls the first distance. The proof then drops the two evidence normalisers and uses that bound for conditional predictions. The five-state HMM below has strictly positive transitions and satisfies the prior-TV condition, yet the two predictors choose different labels. Label agreement needs a posterior-TV condition or a bound that accounts for the evidence probability.","primaryUrl":"https://arxiv.org/pdf/2111.02080v6#page=17","citationCount":1152,"citationUrl":"https://www.semanticscholar.org/paper/10bd4160b44803ada6a3d2e366c44b7e2a4ffe90","severity":"repairable","issueTypes":["invalid_inference","lemma_error","missing_assumption"],"classifications":["repairable"]},{"slug":"transformers-statisticians-local-smoothness","title":"Transformers as Statisticians: Provable In-Context Learning with In-Context Algorithm Selection","authors":["Yu Bai","Fan Chen","Huan Wang","Caiming Xiong","Song Mei"],"year":2023,"category":"Transformers & in-context learning","status":"finding","issueTitle":"A perturbed gradient step can leave the smoothness region","plainExplanation":"The stability proof bounds distance from the minimizer but uses it as distance from the origin. Starting from zero gives an origin bound of $2\\|w^*\\|_2$ before perturbations, which can use the entire smoothness radius. A permitted perturbation then leaves that region. Explicit headroom and a corrected radius bound restore the local stability induction.","primaryUrl":"https://papers.neurips.cc/paper_files/paper/2023/file/b2e63e36c57e153b9015fece2352a9f9-Paper-Conference.pdf#page=28","citationCount":348,"citationUrl":"https://www.semanticscholar.org/paper/70c3d5ab03a54281be91709b19e3f50a2e4be0e3","severity":"repairable","issueTypes":["incorrect_bound","lemma_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-iarc-074","title":"Universal Transformers","authors":["Mostafa Dehghani","Stephan Gouws","Oriol Vinyals","Jakob Uszkoreit","Łukasz Kaiser"],"year":2018,"category":"Transformers & sequence models","status":"unresolved","issueTitle":"Universal Transformers","plainExplanation":"A universal simulation needs an invariant state code through the full normalized encoder and explicit computation resources. The supplied finite-budget construction establishes that restricted result rather than the original unpadded reduction.","primaryUrl":"https://arxiv.org/pdf/1807.03819","noteUrl":"/review?paper=review-iarc-074","severity":"repairable","issueTypes":["invalid_reduction","proof_gap"],"classifications":["repairable"]},{"slug":"review-iarc-075","title":"On the Bottleneck of Graph Neural Networks and its Practical Implications","authors":["Uri Alon","Eran Yahav"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"On the Bottleneck of Graph Neural Networks and its Practical Implications","plainExplanation":"No issue was found in the checked claims. The binary finite-precision counting lower bound checks out on the full task family.","primaryUrl":"https://arxiv.org/pdf/2006.05205","noteUrl":"/review?paper=review-iarc-075","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-076","title":"Neural Message Passing for Quantum Chemistry","authors":["Justin Gilmer","Samuel S. Schoenholz","Patrick F. Riley","Oriol Vinyals","George E. Dahl"],"year":2017,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Neural Message Passing for Quantum Chemistry","plainExplanation":"No issue was found in the checked claims. The selected algebraic reduction is valid with dense messages allowed for unrestricted spectral filters; the localGCN specialization is exact.","primaryUrl":"https://arxiv.org/pdf/1704.01212","noteUrl":"/review?paper=review-iarc-076","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-077","title":"Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules","authors":["Johannes Gasteiger","Shankari Giri","Johannes T. Margraf","Stephan Günnemann"],"year":2020,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules","plainExplanation":"No issue was found in the checked claims. The selected uncertainty argument is sound and its practical limitation has an exact witness.","primaryUrl":"https://arxiv.org/pdf/2011.14115","noteUrl":"/review?paper=review-iarc-077","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-078","title":"SchNet: A continuous-filter convolutional neural network for modeling quantum interactions","authors":["Kristof T. Schütt","Pieter-Jan Kindermans","Huziel E. Sauceda","Stefan Chmiela","Alexandre Tkatchenko","Klaus-Robert Müller"],"year":2017,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"SchNet: A continuous-filter convolutional neural network for modeling quantum interactions","plainExplanation":"No issue was found in the checked claims. The selected symmetry/conservative-force construction is sound on its smooth configuration domain.","primaryUrl":"https://arxiv.org/pdf/1706.08566","noteUrl":"/review?paper=review-iarc-078","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-079","title":"EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations","authors":["Yi-Lun Liao","Brandon Wood","Abhishek Das","Tess Smidt"],"year":2023,"category":"Graphs & equivariance","status":"no-error","issueTitle":"EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations","plainExplanation":"No issue was found in the checked claims. The selected equivariant convolution reduction checks out.","primaryUrl":"https://arxiv.org/pdf/2306.12059","noteUrl":"/review?paper=review-iarc-079","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-iarc-080","title":"On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation","authors":["Sebastian Bach","Alexander Binder","Grégoire Montavon","Frederick Klauschen","Klaus-Robert Müller","Wojciech Samek"],"year":2015,"category":"Transformers & sequence models","status":"no-error","issueTitle":"On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation","plainExplanation":"No issue was found in the checked claims. The selected conservation proof and leakage identities are algebraically sound.","primaryUrl":"https://journals.plos.org/plosone/article/file?id=10.1371%2Fjournal.pone.0130140&type=printable","noteUrl":"/review?paper=review-iarc-080","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig001","title":"Variational Inference with Normalizing Flows","authors":["Danilo Jimenez Rezende","Shakir Mohamed"],"year":2015,"category":"Generative models","status":"no-error","issueTitle":"Variational Inference with Normalizing Flows","plainExplanation":"No error established in these two central constructions. The boundary parameters may yield a zero Jacobian at one point; the paper's strict softplus parameterizations avoid that. Universal approximation and numerical conditioning not audited.","primaryUrl":"https://arxiv.org/pdf/1505.05770","noteUrl":"/review?paper=review-ig001","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-ig002","title":"Importance Weighted Autoencoders","authors":["Yuri Burda","Roger Grosse","Ruslan Salakhutdinov"],"year":2015,"category":"Generative models","status":"addressed","issueTitle":"Convergence of the expected log needs control near zero","plainExplanation":"Positive bounded weights can approach zero so rapidly that every finite-sample expected log is negative infinity, even though their mean converges almost surely. Uniform integrability of the log averages restores the limit; later variational-bound analyses state this condition.","primaryUrl":"https://arxiv.org/pdf/1509.00519","citationCount":1369,"citationUrl":"https://www.semanticscholar.org/paper/3e47c4c2dd98c49b7771c7228812d5fd9eee56a3","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-ig003","title":"f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization","authors":["Sebastian Nowozin","Botond Cseke","Ryota Tomioka"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization","plainExplanation":"Selected contraction route verified within an invariant local neighborhood. It does not establish entry into that neighborhood, stochastic/minibatch convergence, or nonconvex neural-network convergence. No theorem counterexample retained from the printed factor slip.","primaryUrl":"https://arxiv.org/pdf/1606.00709","noteUrl":"/review?paper=review-ig003","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig004","title":"Adversarial Feature Learning","authors":["Jeff Donahue","Philipp Krähenbühl","Trevor Darrell"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Adversarial Feature Learning","plainExplanation":"No issue was found in the checked claims. Almost-everywhere inversion follows at exact joint matching. It is not pointwise inversion on all inputs, an optimization convergence guarantee, or an assertion that a finite neural class attains the ideal global value.","primaryUrl":"https://arxiv.org/pdf/1605.09782","noteUrl":"/review?paper=review-ig004","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig005","title":"Adversarial Variational Bayes: Unifying Variational Autoencoders and Generative Adversarial Networks","authors":["Lars Mescheder","Sebastian Nowozin","Andreas Geiger"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"Adversarial Variational Bayes: Unifying Variational Autoencoders and Generative Adversarial Networks","plainExplanation":"No issue was found in the checked claims. Selected global-equilibrium characterization checks out. The score-identity step in Proposition 2 needs ordinary differentiation-under-integral/fixed-support regularity and is not certified for arbitrary moving-support families. Existence or discovery of a global Nash equilibrium is not established.","primaryUrl":"https://arxiv.org/pdf/1701.04722","noteUrl":"/review?paper=review-ig005","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig006","title":"Wasserstein Auto-Encoders","authors":["Ilya Tolstikhin","Olivier Bousquet","Sylvain Gelly","Bernhard Schoelkopf"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"Wasserstein Auto-Encoders","plainExplanation":"No error in the selected constrained transport identity under regular conditional probabilities. It does not say the finite-penalty or restricted deterministic-encoder implementation equals optimal transport exactly.","primaryUrl":"https://arxiv.org/pdf/1711.01558","noteUrl":"/review?paper=review-ig006","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"infovae-divergence-sign","title":"InfoVAE: Balancing Learning and Inference in Variational Autoencoders","authors":["Shengjia Zhao","Jiaming Song","Stefano Ermon"],"year":2019,"category":"Generative models","status":"finding","issueTitle":"InfoVAE’s divergence substitution needs a coefficient restriction","plainExplanation":"InfoVAE penalizes disagreement between the encoder’s latent distribution and the prior. Proposition 2 permits replacing the marginal KL term by any strict divergence, but its parameter range allows the coefficient of that penalty to become negative. The objective then rewards disagreement. For α=0, λ=1/2 and D=3 KL, shifting an independent Gaussian encoder to mean m leaves its mutual information at zero while increasing the objective by m²/4. The matched model at m=0 cannot be a global maximum.","primaryUrl":"https://arxiv.org/pdf/1706.02262v3","citationCount":495,"citationUrl":"https://www.semanticscholar.org/paper/6f88397a7df1861d18303b19a77d6453ed32621e","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-ig008","title":"MMD GAN: Towards Deeper Understanding of Moment Matching Network","authors":["Chun-Liang Li","Wei-Cheng Chang","Yu Cheng","Yiming Yang","Barnabás Póczos"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"MMD GAN: Towards Deeper Understanding of Moment Matching Network","plainExplanation":"No issue was found in the checked claims. Selected conclusion verified with a repaired inequality and a different harmless constant. Uniformity in phi is essential and is stronger than simply saying the generator is a feed-forward network. The separate weak-topology Theorem 4 is not certified by this check.","primaryUrl":"https://arxiv.org/pdf/1705.08584","noteUrl":"/review?paper=review-ig008","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig009","title":"Density estimation using Real NVP","authors":["Laurent Dinh","Jascha Sohl-Dickstein","Samy Bengio"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Density estimation using Real NVP","plainExplanation":"No error in this central affine-coupling construction. The check concerns a fixed layer with finite outputs; floating-point overflow and batch-dependent normalization elsewhere in the architecture are separate issues.","primaryUrl":"https://arxiv.org/pdf/1605.08803","noteUrl":"/review?paper=review-ig009","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig010","title":"Score-Based Generative Modeling through Stochastic Differential Equations","authors":["Yang Song","Jascha Sohl-Dickstein","Diederik P. Kingma","Abhishek Kumar","Stefano Ermon","Ben Poole"],"year":2020,"category":"Generative models","status":"no-error","issueTitle":"Score-Based Generative Modeling through Stochastic Differential Equations","plainExplanation":"No error in the selected smooth exact-score derivation. Equality of marginals does not imply equal sample paths or equality after replacing the score by an imperfect network. Boundary/singular-data regularity is not established by the algebra alone.","primaryUrl":"https://arxiv.org/pdf/2011.13456","noteUrl":"/review?paper=review-ig010","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig011","title":"Wasserstein GAN","authors":["Martin Arjovsky","Soumith Chintala","Léon Bottou"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"Wasserstein GAN","plainExplanation":"No issue was found in the checked claims. The continuity part is verified. The local Lipschitz portion invokes z-dependent neighborhoods before taking expectations, so that stronger portion is not certified by this review; an integrable Lipschitz envelope on one common parameter neighborhood repairs the argument.","primaryUrl":"https://arxiv.org/pdf/1701.07875","noteUrl":"/review?paper=review-ig011","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig012","title":"Improved Training of Wasserstein GANs","authors":["Ishaan Gulrajani","Faruk Ahmed","Martin Arjovsky","Vincent Dumoulin","Aaron Courville"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"Improved Training of Wasserstein GANs","plainExplanation":"No issue was found in the checked claims. Endpoint normalization is undefined at y and is read by its interior limit. The theorem concerns optimally coupled pairs; independent real/fake pair penalties used in training are a heuristic and do not follow from this proof.","primaryUrl":"https://arxiv.org/pdf/1704.00028","noteUrl":"/review?paper=review-ig012","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig013","title":"GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium","authors":["Martin Heusel","Hubert Ramsauer","Thomas Unterthiner","Bernhard Nessler","Sepp Hochreiter"],"year":2017,"category":"Generative models","status":"finding","issueTitle":"GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium","plainExplanation":"A player can improve by changing its own parameters while holding its opponent fixed, even at a stable reduced-dynamics limit. Convergence to that limit therefore does not establish local Nash.","primaryUrl":"https://arxiv.org/pdf/1706.08500","noteUrl":"/review?paper=review-ig013","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-ig014","title":"InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets","authors":["Xi Chen","Yan Duan","Rein Houthooft","John Schulman","Ilya Sutskever","Pieter Abbeel"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets","plainExplanation":"No issue was found in the checked claims. The discrete entropy ceiling is not transferred to continuous codes; interpretability/disentanglement does not follow merely from high mutual information.","primaryUrl":"https://arxiv.org/pdf/1606.03657","noteUrl":"/review?paper=review-ig014","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig015","title":"Least Squares Generative Adversarial Networks","authors":["Xudong Mao","Qing Li","Haoran Xie","Raymond Y. K. Lau","Zhen Wang","Stephen Paul Smolley"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Least Squares Generative Adversarial Networks","plainExplanation":"No issue was found in the checked claims. The Pearson-divergence argument-order convention varies. The equality is for the augmented functional Eq.(4); after optimizing D, the added real-data term depends on G, so this review does not equate the original Eq.(2) value function or its total derivative to that divergence.","primaryUrl":"https://arxiv.org/pdf/1611.04076","noteUrl":"/review?paper=review-ig015","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig016","title":"Energy-based Generative Adversarial Network","authors":["Junbo Zhao","Michael Mathieu","Yann LeCun"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Energy-based Generative Adversarial Network","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/1609.03126","noteUrl":"/review?paper=review-ig016","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig017","title":"BEGAN: Boundary Equilibrium Generative Adversarial Networks","authors":["David Berthelot","Thomas Schumm","Luke Metz"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"BEGAN: Boundary Equilibrium Generative Adversarial Networks","plainExplanation":"No issue was found in the checked claims. The lower bound is valid: the Wasserstein distance between two loss distributions is at least the absolute difference of their mean losses. Equal means do not imply equal loss distributions or equal image distributions.","primaryUrl":"https://arxiv.org/pdf/1703.10717","noteUrl":"/review?paper=review-ig017","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig018","title":"Adversarially Learned Inference","authors":["Vincent Dumoulin","Ishmael Belghazi","Ben Poole","Olivier Mastropietro","Alex Lamb","Martin Arjovsky","Aaron Courville"],"year":2016,"category":"Statistical learning","status":"no-error","issueTitle":"Adversarially Learned Inference","plainExplanation":"No issue was found in the checked claims. The stochastic architecture is compatible with this joint-law statement. The separately stated invertibility proposition contains encoder/decoder naming ambiguities and is not certified here.","primaryUrl":"https://arxiv.org/pdf/1606.00704","noteUrl":"/review?paper=review-ig018","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig019","title":"The Cramer Distance as a Solution to Biased Wasserstein Gradients","authors":["Marc G. Bellemare","Ivo Danihelka","Will Dabney","Shakir Mohamed","Balaji Lakshminarayanan","Stephan Hoyer","Rémi Munos"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"The Cramer Distance as a Solution to Biased Wasserstein Gradients","plainExplanation":"No issue was found in the checked claims. The sample-loss claim is verified; it does not prove bias for every GAN critic-training protocol. Some informal median/SGD sentences and Theorem 3 dominated-convergence assumptions are not certified.","primaryUrl":"https://arxiv.org/pdf/1705.10743","noteUrl":"/review?paper=review-ig019","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig020","title":"Demystifying MMD GANs","authors":["Mikołaj Bińkowski","Danica J. Sutherland","Michael Arbel","Arthur Gretton"],"year":2018,"category":"Generative models","status":"no-error","issueTitle":"Demystifying MMD GANs","plainExplanation":"No issue was found in the checked claims. Theorem 2 is the primary verified result. Theorem 3 needs the selected pair to be distinguishable by D (implicit in calling D a metric); it does not apply to squared MMD, whose unbiased U-statistic the paper explicitly discusses.","primaryUrl":"https://arxiv.org/pdf/1801.01401","noteUrl":"/review?paper=review-ig020","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig021","title":"Generative Moment Matching Networks","authors":["Yujia Li","Kevin Swersky","Richard Zemel"],"year":2015,"category":"Generative models","status":"no-error","issueTitle":"Generative Moment Matching Networks","plainExplanation":"No issue was found in the checked claims. This is a central explicit derivation, not a numbered theorem. The paper imports the characteristic-kernel theorem from Gretton; this review certifies the finite-sample derivation and limits the informal all-moments wording.","primaryUrl":"https://arxiv.org/pdf/1502.02761","noteUrl":"/review?paper=review-ig021","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig022","title":"Training generative neural networks via Maximum Mean Discrepancy optimization","authors":["Gintare Karolina Dziugaite","Daniel M. Roy","Zoubin Ghahramani"],"year":2015,"category":"Optimization & generalization","status":"no-error","issueTitle":"Training generative neural networks via Maximum Mean Discrepancy optimization","plainExplanation":"No issue was found in the checked claims. The finite-class theorem is the reviewed result. The following corollary appears to put delta rather than 1/delta inside a logarithm, and the infinite-class symmetrization is not certified here.","primaryUrl":"https://arxiv.org/pdf/1505.03906","noteUrl":"/review?paper=review-ig022","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"numerics-gan-skew-scale","title":"The Numerics of GANs","authors":["Lars Mescheder","Sebastian Nowozin","Andreas Geiger"],"year":2017,"category":"Generative models","status":"finding","issueTitle":"The consensus damping bound omits the scale of the skew part","plainExplanation":"Consensus regularization changes the local training matrix from A to A−γAᵀA. Lemma 9 bounds how much oscillation remains relative to decay, but its derivation drops the scale of the skew part of A. For A=[(−1,−1),(1,−1)] and γ=1, the eigenvalues after regularization are −3±i. Their imaginary-to-real ratio is 1/3, above the claimed bound of 1/5. Keeping ‖A−Aᵀ‖₂ in the estimate repairs this bound and preserves the conclusion that sufficiently strong damping reduces the ratio.","primaryUrl":"https://arxiv.org/pdf/1705.10461v3","citationCount":490,"citationUrl":"https://www.semanticscholar.org/paper/459fbc416eb9a55920645c741b1e4cce95f39786","severity":"repairable","issueTypes":["incorrect_bound","lemma_error"],"classifications":["repairable"]},{"slug":"review-ig024","title":"Which Training Methods for GANs do actually Converge?","authors":["Lars Mescheder","Andreas Geiger","Sebastian Nowozin"],"year":2018,"category":"Generative models","status":"no-error","issueTitle":"Which Training Methods for GANs do actually Converge?","plainExplanation":"No issue was found in the checked claims. The stationary trajectory initialized exactly at the origin is the trivial exception to the informal all-curves sentence. The substantive nonconvergence claim is verified; the general manifold convergence theorem is outside this selection.","primaryUrl":"https://arxiv.org/pdf/1801.04406","noteUrl":"/review?paper=review-ig024","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig025","title":"Generalization and Equilibrium in Generative Adversarial Nets (GANs)","authors":["Sanjeev Arora","Rong Ge","Yingyu Liang","Tengyu Ma","Yi Zhang"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Generalization and Equilibrium in Generative Adversarial Nets (GANs)","plainExplanation":"The net bound uses the bounded-parameter convention. Adaptive training corollary additionally uses fresh generated samples, explicitly stated in its appendix proof. No assertion that neural-net distance identifies the full distribution.","primaryUrl":"https://arxiv.org/pdf/1703.00573","noteUrl":"/review?paper=review-ig025","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig026","title":"Improved Variational Inference with Inverse Autoregressive Flow","authors":["Diederik P. Kingma","Tim Salimans","Rafal Jozefowicz","Xi Chen","Ilya Sutskever","Max Welling"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Improved Variational Inference with Inverse Autoregressive Flow","plainExplanation":"No issue was found in the checked claims. The sentence immediately before Eq8 prints sigma_i instead of1/sigma_i; Eq8 and the implemented update are consistent with the correct derivative. This local reciprocal typo does not invalidate the flow construction.","primaryUrl":"https://arxiv.org/pdf/1606.04934","noteUrl":"/review?paper=review-ig026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig027","title":"Masked Autoregressive Flow for Density Estimation","authors":["George Papamakarios","Theo Pavlakou","Iain Murray"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"Masked Autoregressive Flow for Density Estimation","plainExplanation":"No issue was found in the checked claims. This is equality of objectives under inverse transport, not equality of parallel sampling costs or an efficiently evaluable density for arbitrary implicit base data law.","primaryUrl":"https://arxiv.org/pdf/1705.07057","noteUrl":"/review?paper=review-ig027","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig028","title":"Glow: Generative Flow with Invertible 1x1 Convolutions","authors":["Diederik P. Kingma","Prafulla Dhariwal"],"year":2018,"category":"Generative models","status":"no-error","issueTitle":"Glow: Generative Flow with Invertible 1x1 Convolutions","plainExplanation":"No issue was found in the checked claims. These identities require nonzero s throughout training. Orthogonal initialization alone does not guarantee every unconstrained update stays nonsingular; the selected claim is conditional on invertibility.","primaryUrl":"https://arxiv.org/pdf/1807.03039","noteUrl":"/review?paper=review-ig028","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig029","title":"Neural Spline Flows","authors":["Conor Durkan","Artur Bekasov","Iain Murray","George Papamakarios"],"year":2019,"category":"Generative models","status":"no-error","issueTitle":"Neural Spline Flows","plainExplanation":"No issue was found in the checked claims. Strict positive derivatives used in the actual flow avoid zero-Jacobian endpoint cases permitted by a broader nonnegative interpolation statement. Finite-precision root stability is separate.","primaryUrl":"https://arxiv.org/pdf/1906.04032","noteUrl":"/review?paper=review-ig029","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig030","title":"Residual Flows for Invertible Generative Modeling","authors":["Ricky T. Q. Chen","Jens Behrmann","David Duvenaud","Jörn-Henrik Jacobsen"],"year":2019,"category":"Generative models","status":"no-error","issueTitle":"Residual Flows for Invertible Generative Modeling","plainExplanation":"No issue was found in the checked claims. The appendix contains an incorrect E|quadratic form|=trace expression, but the direct norm envelope above proves the claimed theorem. Full support alone does not guarantee finite variance or finite expected compute; only finite runtime almost surely is claimed in the selection.","primaryUrl":"https://arxiv.org/pdf/1906.02735","noteUrl":"/review?paper=review-ig030","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig031","title":"Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design","authors":["Jonathan Ho","Xi Chen","Aravind Srinivas","Yan Duan","Pieter Abbeel"],"year":2019,"category":"Generative models","status":"no-error","issueTitle":"Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design","plainExplanation":"No issue was found in the checked claims. If q has smaller support the Jensen lower bound still holds but the initial importance-integral equality needs adjustment. Actual sigmoid flow provides full interior support.","primaryUrl":"https://arxiv.org/pdf/1902.00275","noteUrl":"/review?paper=review-ig031","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig032","title":"Deep Unsupervised Learning using Nonequilibrium Thermodynamics","authors":["Jascha Sohl-Dickstein","Eric A. Weiss","Niru Maheswaranathan","Surya Ganguli"],"year":2015,"category":"Generative models","status":"no-error","issueTitle":"Deep Unsupervised Learning using Nonequilibrium Thermodynamics","plainExplanation":"No issue was found in the checked claims. Maximum entropy of the terminal law alone would not imply monotonicity, but the particular diffusion kernels supply it. The conditional sampling perturbation derivation is not certified.","primaryUrl":"https://arxiv.org/pdf/1503.03585","noteUrl":"/review?paper=review-ig032","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig033","title":"Denoising Diffusion Probabilistic Models","authors":["Jonathan Ho","Ajay Jain","Pieter Abbeel"],"year":2020,"category":"Generative models","status":"no-error","issueTitle":"Denoising Diffusion Probabilistic Models","plainExplanation":"No issue was found in the checked claims. The unweighted simple noise loss later used for sample quality is not identical to the likelihood bound; the paper labels it a reweighting. This review checks the exact decomposition.","primaryUrl":"https://arxiv.org/pdf/2006.11239","noteUrl":"/review?paper=review-ig033","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig034","title":"Improved Denoising Diffusion Probabilistic Models","authors":["Alex Nichol","Prafulla Dhariwal"],"year":2021,"category":"Generative models","status":"no-error","issueTitle":"Improved Denoising Diffusion Probabilistic Models","plainExplanation":"No issue was found in the checked claims. The formula minimizes variance of the scalar loss estimator, not automatically gradient variance. This is a central derivation rather than a numbered theorem; empirical gradient-noise improvement is separately measured.","primaryUrl":"https://arxiv.org/pdf/2102.09672","noteUrl":"/review?paper=review-ig034","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig035","title":"Diffusion Models Beat GANs on Image Synthesis","authors":["Prafulla Dhariwal","Alex Nichol"],"year":2021,"category":"Generative models","status":"no-error","issueTitle":"Diffusion Models Beat GANs on Image Synthesis","plainExplanation":"No issue was found in the checked claims. The Gaussian tilt is an approximation when classifier log density has curvature; finite-step sampling from it is not proved to yield the exact class conditional. Arbitrary guidance scaling additionally changes the target.","primaryUrl":"https://arxiv.org/pdf/2105.05233","noteUrl":"/review?paper=review-ig035","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig036","title":"Generative Modeling by Estimating Gradients of the Data Distribution","authors":["Yang Song","Stefano Ermon"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Generative Modeling by Estimating Gradients of the Data Distribution","plainExplanation":"No issue was found in the checked claims. The equivalence is credited in the paper to Vincent2011. It identifies the score of smoothed data; exact Langevin convergence additionally requires its own regularity and limiting conditions.","primaryUrl":"https://arxiv.org/pdf/1907.05600","noteUrl":"/review?paper=review-ig036","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig037","title":"Improved Techniques for Training Score-Based Generative Models","authors":["Yang Song","Stefano Ermon"],"year":2020,"category":"Generative models","status":"no-error","issueTitle":"Improved Techniques for Training Score-Based Generative Models","plainExplanation":"No issue was found in the checked claims. The overlap inequality is for distinct mixture components; applying the denominator bound withi=j double-counts the same component. It explains component interaction, not a general guarantee that a chosen noise level trains well.","primaryUrl":"https://arxiv.org/pdf/2006.09011","noteUrl":"/review?paper=review-ig037","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig038","title":"DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps","authors":["Cheng Lu","Yuhao Zhou","Fan Bao","Jianfei Chen","Chongxuan Li","Jun Zhu"],"year":2022,"category":"Generative models","status":"no-error","issueTitle":"DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps","plainExplanation":"No issue was found in the checked claims. The exact representation is verified. Infinite-logSNR endpoint singularities and the order3 global-error proof are outside this selection; a finite cutoff is needed for ordinary bounded-coefficient estimates.","primaryUrl":"https://arxiv.org/pdf/2206.00927","noteUrl":"/review?paper=review-ig038","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig039","title":"DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models","authors":["Cheng Lu","Yuhao Zhou","Fan Bao","Jianfei Chen","Chongxuan Li","Jun Zhu"],"year":2022,"category":"Generative models","status":"no-error","issueTitle":"DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models","plainExplanation":"No issue was found in the checked claims. Equivalent exact solutions can lead to different finite-order solvers because different terms are integrated analytically. Stochastic extension and claimed global convergence order are not certified here.","primaryUrl":"https://arxiv.org/pdf/2211.01095","noteUrl":"/review?paper=review-ig039","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig040","title":"UniPC: A Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models","authors":["Wenliang Zhao","Lujia Bai","Yongming Rao","Jie Zhou","Jiwen Lu"],"year":2023,"category":"Generative models","status":"unresolved","issueTitle":"The UniC order proof needs a Lipschitz highest derivative","plainExplanation":"Continuity of the highest stated derivative supplies a smaller-order remainder, but not the extra full power of \\(h\\) used in the proof. The moment cancellation is valid once an appropriate remainder bound is supplied.","primaryUrl":"https://arxiv.org/pdf/2302.04867","noteUrl":"/review?paper=review-ig040","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-ig041","title":"Elucidating the Design Space of Diffusion-Based Generative Models","authors":["Tero Karras","Miika Aittala","Timo Aila","Samuli Laine"],"year":2022,"category":"Generative models","status":"no-error","issueTitle":"Elucidating the Design Space of Diffusion-Based Generative Models","plainExplanation":"No issue was found in the checked claims. Checks the central explicit optimization derivation; does not establish that variance scaling is universally optimal for training or certify empirical design choices.","primaryUrl":"https://arxiv.org/pdf/2206.00364","noteUrl":"/review?paper=review-ig041","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig042","title":"Pseudo Numerical Methods for Diffusion Models on Manifolds","authors":["Luping Liu","Yi Ren","Zhijie Lin","Zhou Zhao"],"year":2022,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Pseudo Numerical Methods for Diffusion Models on Manifolds","plainExplanation":"No issue was found in the checked claims. The exact-noise hypothesis is much stronger than an MSE-optimal predictor of conditional mean noise; property does not itself prove numerical order or learned-sampler accuracy. Appendix A.6 order proof only partially inspected and not certified.","primaryUrl":"https://arxiv.org/pdf/2202.09778","noteUrl":"/review?paper=review-ig042","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig043","title":"Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models","authors":["Fan Bao","Chongxuan Li","Jun Zhu","Bo Zhang"],"year":2022,"category":"Generative models","status":"no-error","issueTitle":"Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models","plainExplanation":"No issue was found in the checked claims. The optimality holds within the stated isotropic reverse family and for the true score; estimated scores can make the plug-in variance inaccurate and the paper separately clips/bounds it.","primaryUrl":"https://arxiv.org/pdf/2201.06503","noteUrl":"/review?paper=review-ig043","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig044","title":"Denoising Diffusion Restoration Models","authors":["Bahjat Kawar","Michael Elad","Stefano Ermon","Jiaming Song"],"year":2022,"category":"Generative models","status":"no-error","issueTitle":"Denoising Diffusion Restoration Models","plainExplanation":"No issue was found in the checked claims. The proposition explicitly integrates over the measurement distribution given clean data. It is a forward marginal statement, not a guarantee that the trained algorithm samples the posterior conditional on an arbitrary fixed observation.","primaryUrl":"https://arxiv.org/pdf/2201.11793","noteUrl":"/review?paper=review-ig044","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"dps-gaussian-noise-scaling","title":"Diffusion Posterior Sampling for General Noisy Inverse Problems","authors":["Hyungjin Chung","Jeongsol Kim","Michael T. Mccann","Marc L. Klasky","Jong Chul Ye"],"year":2023,"category":"Generative models","status":"finding","issueTitle":"The DPS likelihood bound has the wrong dependence on noise","plainExplanation":"DPS replaces an average likelihood by the likelihood evaluated at a denoised mean. Theorem 1 bounds the resulting error, but the Gaussian constant inherited from Lemma 2 is incorrect. With identity observation, σ=0.1, y=0.2 and conditional state distribution N(0,0.01), the signed gap is approximately 0.497859; the stated upper bound is only 6.14×10⁻²³. Differentiating the Gaussian gives the constant exp(−1/2)/(√(2π)σ²), which restores the likelihood-bound argument.","primaryUrl":"https://arxiv.org/pdf/2209.14687v4","citationCount":1867,"citationUrl":"https://www.semanticscholar.org/paper/61e46884567be7cad12e999365b16a8d3414b678","severity":"repairable","issueTypes":["incorrect_bound","lemma_error"],"classifications":["repairable"]},{"slug":"review-ig046","title":"Classifier-Free Diffusion Guidance","authors":["Jonathan Ho","Tim Salimans"],"year":2022,"category":"Generative models","status":"no-error","issueTitle":"Classifier-Free Diffusion Guidance","plainExplanation":"No issue was found in the checked claims. A fixed-time tilted score identity does not establish that integrating the guided diffusion samples the same power tilt at time zero. The paper notes generic neural score fields need not be conservative; no exact endpoint-distribution guarantee is certified.","primaryUrl":"https://arxiv.org/pdf/2207.12598","noteUrl":"/review?paper=review-ig046","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig047","title":"DreamFusion: Text-to-3D using 2D Diffusion","authors":["Ben Poole","Ajay Jain","Jonathan T. Barron","Ben Mildenhall"],"year":2022,"category":"Generative models","status":"no-error","issueTitle":"DreamFusion: Text-to-3D using 2D Diffusion","plainExplanation":"No issue was found in the checked claims. Approximate or nonconservative CFG scores weaken the interpretation as an exact KL gradient. Lower variance from the control variate is empirical unless covariance assumptions are supplied.","primaryUrl":"https://arxiv.org/pdf/2209.14988","noteUrl":"/review?paper=review-ig047","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig048","title":"Deep Variational Information Bottleneck","authors":["Alexander A. Alemi","Ian Fischer","Joshua V. Dillon","Kevin Murphy"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"Deep Variational Information Bottleneck","plainExplanation":"No issue was found in the checked claims. Only the variational bound is certified; finite Gaussian parameterization and stochastic optimization need not attain the information-theoretic optimum. Dropped label entropy is constant, with the classification setting ensuring its usual nonnegative discrete interpretation.","primaryUrl":"https://arxiv.org/pdf/1612.00410","noteUrl":"/review?paper=review-ig048","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig049","title":"Mutual Information Neural Estimation","authors":["Mohamed Ishmael Belghazi","Aristide Baratin","Sai Rajeswar","Sherjil Ozair","Yoshua Bengio","Aaron Courville","R Devon Hjelm"],"year":2018,"category":"Generative models","status":"no-error","issueTitle":"Mutual Information Neural Estimation","plainExplanation":"No issue was found in the checked claims. This checks the foundational variational identity, not the subsequent finite-network approximation or sample estimator strong-consistency claims, which require additional assumptions and optimization control.","primaryUrl":"https://arxiv.org/pdf/1801.04062","noteUrl":"/review?paper=review-ig049","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig050","title":"Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning","authors":["Jean-Bastien Grill","Florian Strub","Florent Altché","Corentin Tallec","Pierre H. Richemond","Elena Buchatskaya","Carl Doersch","Bernardo Avila Pires","Zhaohan Daniel Guo","Mohammad Gheshlaghi Azar","Bilal Piot","Koray Kavukcuoglu","Rémi Munos","Michal Valko"],"year":2020,"category":"Representation learning","status":"no-error","issueTitle":"Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning","plainExplanation":"No issue was found in the checked claims. The actual finite predictor is not guaranteed to be globally optimal and target parameters evolve. This identity does not prove avoidance of collapse; the paper explicitly frames stability of collapsed equilibria as a hypothesis.","primaryUrl":"https://arxiv.org/pdf/2006.07733","noteUrl":"/review?paper=review-ig050","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig051","title":"Barlow Twins: Self-Supervised Learning via Redundancy Reduction","authors":["Jure Zbontar","Li Jing","Ishan Misra","Yann LeCun","Stéphane Deny"],"year":2021,"category":"Representation learning","status":"no-error","issueTitle":"Barlow Twins: Self-Supervised Learning via Redundancy Reduction","plainExplanation":"No issue was found in the checked claims. The paper explicitly calls its use of cross-correlation instead of autocorrelation a further surrogate. Gaussian differential entropy does not literally have H(Z|Y)=0 for deterministic continuous variables, so this is a discretized/approximate motivation, not an exact IB equivalence or a certified anti-collapse theorem.","primaryUrl":"https://arxiv.org/pdf/2103.03230","noteUrl":"/review?paper=review-ig051","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig052","title":"Contrastive Multiview Coding","authors":["Yonglong Tian","Dilip Krishnan","Phillip Isola"],"year":2019,"category":"Representation learning","status":"no-error","issueTitle":"Contrastive Multiview Coding","plainExplanation":"Density-ratio result verified; the displayed MI argument has the same known shortcut as CPC, and the valid alternative proof was already addressed by Poole et al. 2019 in the original frozen audit. No unresolved error retained.","primaryUrl":"https://arxiv.org/pdf/1906.05849","noteUrl":"/review?paper=review-ig052","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig053","title":"What Makes for Good Views for Contrastive Learning?","authors":["Yonglong Tian","Chen Sun","Ben Poole","Dilip Krishnan","Cordelia Schmid","Phillip Isola"],"year":2020,"category":"Representation learning","status":"unresolved","issueTitle":"The appendix’s information lower bound need not be attainable","plainExplanation":"The chain-rule lower bound does not prove that a pair of admissible views attains it. Noisy labels provide a finite counterexample to the appendix’s equality, while the main minimal-sufficiency conclusion can still hold.","primaryUrl":"https://arxiv.org/pdf/2005.10243","noteUrl":"/review?paper=review-ig053","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-ig054","title":"Supervised Contrastive Learning","authors":["Prannay Khosla","Piotr Teterwak","Chen Wang","Aaron Sarna","Yonglong Tian","Phillip Isola","Aaron Maschinot","Ce Liu","Dilip Krishnan"],"year":2020,"category":"Representation learning","status":"no-error","issueTitle":"Supervised Contrastive Learning","plainExplanation":"No issue was found in the checked claims. The full batch gradient also includes terms in which this embedding is another anchor’s positive/negative. The paper explicitly derives the individual anchor loss. Qualitative hard-mining claims do not cover anti-aligned positives, whose tangent factor also vanishes.","primaryUrl":"https://arxiv.org/pdf/2004.11362","noteUrl":"/review?paper=review-ig054","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig055","title":"Domain-Adversarial Training of Neural Networks","authors":["Yaroslav Ganin","Evgeniya Ustinova","Hana Ajakan","Pascal Germain","Hugo Larochelle","François Laviolette","Mario Marchand","Victor Lempitsky"],"year":2015,"category":"Generative models","status":"finding","issueTitle":"Domain-Adversarial Training of Neural Networks","plainExplanation":"Risk transfer depends on disagreements between pairs of label predictors. The stated $\\mathcal H$-divergence only compares the positive prediction rate of each individual predictor. Equal rates need not imply equal disagreement rates, even when $\\mathcal H$ is finite and closed under label complementation.","primaryUrl":"https://arxiv.org/pdf/1505.07818","noteUrl":"/review?paper=review-ig055","severity":"repairable","issueTypes":["citation_hypothesis_mismatch","incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-ig056","title":"Learning Transferable Features with Deep Adaptation Networks","authors":["Mingsheng Long","Yue Cao","Jianmin Wang","Michael I. Jordan"],"year":2015,"category":"Generative models","status":"unresolved","issueTitle":"Learning Transferable Features with Deep Adaptation Networks","plainExplanation":"No issue was found in the checked claims. The final risk inequality is not contradicted: the proof permits C=C0+C1+2, which makes the bound valid but uninformative for risks at most one. Selecting one Parzen classifier does not bound an arbitrary hypothesis-class supremum, but that reduction is unnecessary to obtain this particular loose inequality.","primaryUrl":"https://arxiv.org/pdf/1502.02791","noteUrl":"/review?paper=review-ig056","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig057","title":"Deep CORAL: Correlation Alignment for Deep Domain Adaptation","authors":["Baochen Sun","Kate Saenko"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Deep CORAL: Correlation Alignment for Deep Domain Adaptation","plainExplanation":"No issue was found in the checked claims. Checks the central analytic gradient, not a domain adaptation generalization theorem. Equal covariance does not imply equal distributions; the paper itself notes degenerate constant features can minimize this penalty.","primaryUrl":"https://arxiv.org/pdf/1607.01719","noteUrl":"/review?paper=review-ig057","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig058","title":"Deep Transfer Learning with Joint Adaptation Networks","authors":["Mingsheng Long","Han Zhu","Jianmin Wang","Michael I. Jordan"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Deep Transfer Learning with Joint Adaptation Networks","plainExplanation":"No issue was found in the checked claims. The zero-discrepancy implication requires a characteristic product kernel; Gaussian kernels meet it on Euclidean layer spaces. Matching predicted-output tuples is not by itself matching unknown target labels.","primaryUrl":"https://arxiv.org/pdf/1605.06636","noteUrl":"/review?paper=review-ig058","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig059","title":"Maximum Classifier Discrepancy for Unsupervised Domain Adaptation","authors":["Kuniaki Saito","Kohei Watanabe","Yoshitaka Ushiku","Tatsuya Harada"],"year":2017,"category":"Generative models","status":"finding","issueTitle":"Maximum Classifier Discrepancy for Unsupervised Domain Adaptation","plainExplanation":"Individual classifiers can have identical prediction rates in two domains while their disagreements have different rates. The second inequality in Equation (7) therefore does not follow from the definitions. The first inequality remains a valid domain-adaptation bound.","primaryUrl":"https://arxiv.org/pdf/1712.02560","noteUrl":"/review?paper=review-ig059","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-ig060","title":"Bridging Theory and Algorithm for Domain Adaptation","authors":["Yuchen Zhang","Tianle Liu","Mingsheng Long","Michael I. Jordan"],"year":2019,"category":"Generative models","status":"no-error","issueTitle":"Bridging Theory and Algorithm for Domain Adaptation","plainExplanation":"No issue was found in the checked claims. The separate uniform-complexity proof is not certified. Official supplement saved at corrections/mdd_supp.pdf; its A.3 also provides the addressing source for the earlier divergence-class mismatch.","primaryUrl":"https://arxiv.org/pdf/1904.05801","noteUrl":"/review?paper=review-ig060","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig061","title":"Invariant Risk Minimization","authors":["Martin Arjovsky","Léon Bottou","Ishaan Gulrajani","David Lopez-Paz"],"year":2019,"category":"Generative models","status":"no-error","issueTitle":"Invariant Risk Minimization","plainExplanation":"No issue was found in the checked claims. This characterizes linear invariant predictors, not nonlinear IRMv1 generalization or the later transversality theorem. The surrounding claim that all zero sets/intersections are manifolds is not needed and not certified.","primaryUrl":"https://arxiv.org/pdf/1907.02893","noteUrl":"/review?paper=review-ig061","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig062","title":"Out-of-Distribution Generalization via Risk Extrapolation (REx)","authors":["David Krueger","Ethan Caballero","Joern-Henrik Jacobsen","Amy Zhang","Jonathan Binas","Dinghuai Zhang","Remi Le Priol","Aaron Courville"],"year":2020,"category":"Optimization & generalization","status":"no-error","issueTitle":"Out-of-Distribution Generalization via Risk Extrapolation (REx)","plainExplanation":"No issue was found in the checked claims. If only proper rather than strictly proper, prediction need not be unique. Finite domain sets or finite risk-equality penalties do not satisfy the selected theorem’s all-intervention hypothesis.","primaryUrl":"https://arxiv.org/pdf/2003.00688","noteUrl":"/review?paper=review-ig062","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig063","title":"Learning Representations for Counterfactual Inference","authors":["Fredrik D. Johansson","Uri Shalit","David Sontag"],"year":2016,"category":"Representation learning","status":"no-error","issueTitle":"Learning Representations for Counterfactual Inference","plainExplanation":"The first ridge stability inequality is imported as Theorem A1 and is not independently certified; review covers the paper’s explicit new Lipschitz/nearest-neighbor step. No claim that learned representations preserve causal identification.","primaryUrl":"https://arxiv.org/pdf/1605.03661","noteUrl":"/review?paper=review-ig063","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig064","title":"Causal Effect Inference with Deep Latent-Variable Models","authors":["Christos Louizos","Uri Shalit","Joris Mooij","David Sontag","Richard Zemel","Max Welling"],"year":2017,"category":"Statistical learning","status":"no-error","issueTitle":"Causal Effect Inference with Deep Latent-Variable Models","plainExplanation":"No issue was found in the checked claims. The key full-latent-joint recovery assumption is explicitly assumed, not established by the VAE. Conditional positivity is needed to evaluate both treatments; no claim of identification from arbitrary observed proxies is certified.","primaryUrl":"https://arxiv.org/pdf/1705.08821","noteUrl":"/review?paper=review-ig064","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig065","title":"Weight Uncertainty in Neural Networks","authors":["Charles Blundell","Julien Cornebise","Koray Kavukcuoglu","Daan Wierstra"],"year":2015,"category":"Statistical learning","status":"no-error","issueTitle":"Weight Uncertainty in Neural Networks","plainExplanation":"No issue was found in the checked claims. The proposition’s differentiability wording alone does not ensure dominated differentiation; the paper introduces it as holding under certain conditions. This is a regularity qualification, not a demonstrated failure in its Gaussian neural setup.","primaryUrl":"https://arxiv.org/pdf/1505.05424","noteUrl":"/review?paper=review-ig065","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig066","title":"Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning","authors":["Yarin Gal","Zoubin Ghahramani"],"year":2015,"category":"Statistical learning","status":"no-error","issueTitle":"Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning","plainExplanation":"No issue was found in the checked claims. Finite-T plug-in covariance with denominator T is downward biased, not an unbiased sample-covariance claim. This review certifies the moment formulas, not calibration or the singular Bernoulli-to-Gaussian KL approximation. Separate technical appendix saved in corrections/dropout_appendix.pdf.","primaryUrl":"https://arxiv.org/pdf/1506.02142","noteUrl":"/review?paper=review-ig066","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig067","title":"Concrete Dropout","authors":["Yarin Gal","Jiri Hron","Alex Kendall"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"Concrete Dropout","plainExplanation":"No issue was found in the checked claims. The prose claim that minimizing KL is equivalent solely to maximizing Bernoulli entropy overlooks its weight term: for fixed M, c(1-p)||M||²-KH(p) is minimized at sigmoid(c||M||²/K), generally above.5. This local interpretive error is not retained as a substantial unresolved theorem defect.","primaryUrl":"https://arxiv.org/pdf/1705.07832","noteUrl":"/review?paper=review-ig067","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig068","title":"Variational Dropout and the Local Reparameterization Trick","authors":["Diederik P. Kingma","Tim Salimans","Max Welling"],"year":2015,"category":"Generative models","status":"no-error","issueTitle":"Variational Dropout and the Local Reparameterization Trick","plainExplanation":"No issue was found in the checked claims. The appendix’s intermediate displayed q factor omits scale factors, but the conditional-expectation argument repairs the claimed variance comparison. This covers fully connected layers with independent output weights; convolutional spatial correlations require care.","primaryUrl":"https://arxiv.org/pdf/1506.02557","noteUrl":"/review?paper=review-ig068","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig069","title":"Variational Dropout Sparsifies Deep Neural Networks","authors":["Dmitry Molchanov","Arsenii Ashukha","Dmitry Vetrov"],"year":2017,"category":"Generative models","status":"no-error","issueTitle":"Variational Dropout Sparsifies Deep Neural Networks","plainExplanation":"No issue was found in the checked claims. Printed Theta(alpha_i^-1) needs a nonzero residual numerator b; if y=0 then theta_i=0 exactly. The claimed vanishing limit still holds. The improper-prior Bayesian interpretation and diagonal convolution covariance are not certified.","primaryUrl":"https://arxiv.org/pdf/1701.05369","noteUrl":"/review?paper=review-ig069","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig070","title":"Disentangling by Factorising","authors":["Hyunjik Kim","Andriy Mnih"],"year":2018,"category":"Generative models","status":"no-error","issueTitle":"Disentangling by Factorising","plainExplanation":"No issue was found in the checked claims. This decomposition is exact; the adversarial total-correlation estimator and identifiability of disentangled factors are separate and not established by it. The paper explicitly documents failure of a naive minibatch density estimate.","primaryUrl":"https://arxiv.org/pdf/1802.05983","noteUrl":"/review?paper=review-ig070","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig071","title":"Isolating Sources of Disentanglement in Variational Autoencoders","authors":["Ricky T. Q. Chen","Xuechen Li","Roger Grosse","David Duvenaud"],"year":2018,"category":"Generative models","status":"no-error","issueTitle":"Isolating Sources of Disentanglement in Variational Autoencoders","plainExplanation":"No issue was found in the checked claims. Paper’s compressed EqS8 omits conditioning in an intermediate line, and M=N cannot sample N other distinct indices from N-1; full-data endpoint needs an explicit convention. Neither affects the interior estimator/decomposition checked here; no substantive unresolved error retained.","primaryUrl":"https://arxiv.org/pdf/1802.04942","noteUrl":"/review?paper=review-ig071","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig072","title":"Variational Autoencoders and Nonlinear ICA: A Unifying Framework","authors":["Ilyes Khemakhem","Diederik P. Kingma","Ricardo Pio Monti","Aapo Hyvärinen"],"year":2019,"category":"Generative models","status":"unresolved","issueTitle":"Minimal sufficient statistics need not have nonzero derivatives almost everywhere","plainExplanation":"A statistic can be strictly increasing yet have derivative zero on a set of positive measure. Strong exponentiality therefore does not justify the derivative-rank assertion. The change of variables for lower-dimensional images also needs a measure formulation rather than an ambient Lebesgue density.","primaryUrl":"https://arxiv.org/pdf/1907.04809","noteUrl":"/review?paper=review-ig072","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-ig073","title":"Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning","authors":["Aapo Hyvarinen","Hiroaki Sasaki","Richard E. Turner"],"year":2018,"category":"Representation learning","status":"no-error","issueTitle":"Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning","plainExplanation":"No issue was found in the checked claims. Requires the stated smooth inverse and connected R^n domain; mere bijectivity is insufficient. Population-optimal logistic fitting is idealized, not a finite-sample training guarantee.","primaryUrl":"https://arxiv.org/pdf/1805.08651","noteUrl":"/review?paper=review-ig073","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig074","title":"Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA","authors":["Aapo Hyvarinen","Hiroshi Morioka"],"year":2016,"category":"Representation learning","status":"no-error","issueTitle":"Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA","plainExplanation":"No issue was found in the checked claims. Checks the affine-statistic recovery theorem, not the stronger subsequent linear-ICA/sign corollaries. Conditional independence within segments need not imply independence after mixing segments; iVAE’s later discussion already addresses that distinction.","primaryUrl":"https://arxiv.org/pdf/1605.06336","noteUrl":"/review?paper=review-ig074","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig075","title":"The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables","authors":["Chris J. Maddison","Andriy Mnih","Yee Whye Teh"],"year":2016,"category":"Statistical learning","status":"no-error","issueTitle":"The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables","plainExplanation":"The relaxed objective has unbiased pathwise gradients for itself; it is generally biased as an estimator of the original discrete-objective gradient. No claim of uniform gradient convergence at zero temperature is certified.","primaryUrl":"https://arxiv.org/pdf/1611.00712","noteUrl":"/review?paper=review-ig075","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig076","title":"Categorical Reparameterization with Gumbel-Softmax","authors":["Eric Jang","Shixiang Gu","Ben Poole"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Categorical Reparameterization with Gumbel-Softmax","plainExplanation":"No issue was found in the checked claims. This establishes the continuous distribution and its reparameterization, not unbiasedness of the straight-through discrete estimator, which is a heuristic variant.","primaryUrl":"https://arxiv.org/pdf/1611.01144","noteUrl":"/review?paper=review-ig076","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig077","title":"REBAR: Low-variance, unbiased gradient estimates for discrete latent variable models","authors":["George Tucker","Andriy Mnih","Chris J. Maddison","Dieterich Lawson","Jascha Sohl-Dickstein"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"REBAR: Low-variance, unbiased gradient estimates for discrete latent variable models","plainExplanation":"No issue was found in the checked claims. Fresh-sample unbiasedness remains conditional on learned control parameters/history. Variance improvement is not guaranteed for arbitrary control scales or arbitrary coupling; unbiasedness is the selected exact conclusion.","primaryUrl":"https://arxiv.org/pdf/1703.07370","noteUrl":"/review?paper=review-ig077","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig078","title":"Backpropagation through the Void: Optimizing control variates for black-box gradient estimation","authors":["Will Grathwohl","Dami Choi","Yuhuai Wu","Geoffrey Roeder","David Duvenaud"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Backpropagation through the Void: Optimizing control variates for black-box gradient estimation","plainExplanation":"No issue was found in the checked claims. The proof covers the single stochastic-node estimator; the separate RL trajectory theorems were not reviewed. Minimizing estimated second moment equals variance minimization only because the mean is independent of control parameters.","primaryUrl":"https://arxiv.org/pdf/1711.00123","noteUrl":"/review?paper=review-ig078","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig079","title":"Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm","authors":["Qiang Liu","Dilin Wang"],"year":2016,"category":"Optimization & generalization","status":"no-error","issueTitle":"Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm","plainExplanation":"Appendix Lemma A.1 display omits a minus sign, restored by its own proof and the main theorem. Inverse function theorem alone gives local invertibility; bounded derivative/small-step conditions supply global diffeomorphism. No global particle-convergence claim certified.","primaryUrl":"https://arxiv.org/pdf/1608.04471","noteUrl":"/review?paper=review-ig079","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-ig080","title":"Automatic Differentiation Variational Inference","authors":["Alp Kucukelbir","Dustin Tran","Rajesh Ranganath","Andrew Gelman","David M. Blei"],"year":2016,"category":"Generative models","status":"no-error","issueTitle":"Automatic Differentiation Variational Inference","plainExplanation":"No issue was found in the checked claims. Appendix A prose incorrectly says every affine map has Jacobian1; its determinant is generally constant, not1. Appendix C correctly retains covariance log-determinant and its derivative, so this does not invalidate the selected algorithm formulas. Broader optimization/calibration claims not certified.","primaryUrl":"https://arxiv.org/pdf/1603.00788","noteUrl":"/review?paper=review-ig080","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1611-03530","title":"Understanding deep learning requires rethinking generalization","authors":["Chiyuan Zhang","Samy Bengio","Moritz Hardt","Benjamin Recht","Oriol Vinyals"],"year":2016,"category":"Optimization & generalization","status":"no-error","issueTitle":"Understanding deep learning requires rethinking generalization","plainExplanation":"No issue was found in the checked claims. The selected interpolation construction works; no retained mathematical error.","primaryUrl":"https://arxiv.org/pdf/1611.03530v2","noteUrl":"/review?paper=review-io-1611-03530","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"pac-bayes-finite-symmetry-mixture","title":"Computing Nonvacuous Generalization Bounds for Deep (Stochastic) Neural Networks with Many More Parameters than Training Data","authors":["Gintare Karolina Dziugaite","Daniel M. Roy"],"year":2017,"category":"Optimization & generalization","status":"finding","issueTitle":"A finite set of symmetries need not give the stated KL reduction","plainExplanation":"Averaging prior-preserving symmetries cannot increase KL for the invertible maps considered here, but the exact reduction depends on every component of the mixture. Lemma B.1 replaces their average divergence from the mixture with the divergence of Q alone. Full finite-group averaging justifies that replacement. The lemma allows any finite set, and a partial cyclic average violates the identity. Depending on the posterior, the formula can either overstate or understate the mixture complexity.","primaryUrl":"https://arxiv.org/pdf/1703.11008v2","citationCount":975,"citationUrl":"https://www.semanticscholar.org/paper/540c226fdf7047ac602c7cb05a18db19ee595df0","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-io-1707-09564","title":"A PAC-Bayesian Approach to Spectrally-Normalized Margin Bounds for Neural Networks","authors":["Behnam Neyshabur","Srinadh Bhojanapalli","Nathan Srebro"],"year":2017,"category":"Statistical learning","status":"no-error","issueTitle":"A PAC-Bayesian Approach to Spectrally-Normalized Margin Bounds for Neural Networks","plainExplanation":"No issue was found in the checked claims. Selected proofs reconstruct correctly; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1707.09564v2","noteUrl":"/review?paper=review-io-1707-09564","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"batchnorm-weight-gradient-bound","title":"How Does Batch Normalization Help Optimization?","authors":["Shibani Santurkar","Dimitris Tsipras","Andrew Ilyas","Aleksander Madry"],"year":2018,"category":"Optimization & generalization","status":"finding","issueTitle":"The BatchNorm bound can be negative for a squared gradient","plainExplanation":"A squared gradient cannot be negative. Yet Theorem 4.4 gives it a negative upper bound for the two-example batch below. With two examples, centering and normalization remove both directions in the batch space, so the idealized BatchNorm gradient is zero. The theorem subtracts too much from the ordinary gradient: its activation-correlation term lacks the normalization supplied by the BatchNorm Jacobian. The exact calculation for a fixed batch has a factor of 1/m in each squared projection term.","primaryUrl":"https://arxiv.org/pdf/1805.11604v5","citationCount":1763,"citationUrl":"https://www.semanticscholar.org/paper/94be567c32ae76bdaadabd4975807a94181e39b3","severity":"major","issueTypes":["false_statement","incorrect_bound","proof_gap"],"classifications":["major"]},{"slug":"review-io-1711-05101","title":"Decoupled Weight Decay Regularization","authors":["Ilya Loshchilov","Frank Hutter"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Decoupled Weight Decay Regularization","plainExplanation":"No issue was found in the checked claims. Proposition 3's displayed anisotropic penalty divides by the learning rate twice after substituting its stated coefficient. Appendix A of the same paper uses the correct coefficient and reproduces the desired decoupled shrinkage. This is a same-paper clarification, not a separate correction or a full equivalence claim for history-dependent Adam updates.","primaryUrl":"https://arxiv.org/pdf/1711.05101v3","noteUrl":"/review?paper=review-io-1711-05101","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1705-08292","title":"The Marginal Value of Adaptive Gradient Methods in Machine Learning","authors":["Ashia C. Wilson","Rebecca Roelofs","Mitchell Stern","Nathan Srebro","Benjamin Recht"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"The Marginal Value of Adaptive Gradient Methods in Machine Learning","plainExplanation":"The selected invariant-ray and test-score calculations are valid. No retained new error; unconditional convergence language is not established by the invariant-ray argument alone.","primaryUrl":"https://arxiv.org/pdf/1705.08292v2","noteUrl":"/review?paper=review-io-1705-08292","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1802-08246","title":"Characterizing Implicit Bias in Terms of Optimization Geometry","authors":["Suriya Gunasekar","Jason Lee","Daniel Soudry","Nathan Srebro"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Characterizing Implicit Bias in Terms of Optimization Geometry","plainExplanation":"No issue was found in the checked claims. Selected proofs reconstruct correctly; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1802.08246v3","noteUrl":"/review?paper=review-io-1802-08246","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1802-06509","title":"On the Optimization of Deep Networks: Implicit Acceleration by Overparameterization","authors":["Sanjeev Arora","Nadav Cohen","Elad Hazan"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"On the Optimization of Deep Networks: Implicit Acceleration by Overparameterization","plainExplanation":"No issue was found in the checked claims. Exact continuous-time identity verified; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1802.06509v2","noteUrl":"/review?paper=review-io-1802-06509","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1905-13655","title":"Implicit Regularization in Deep Matrix Factorization","authors":["Sanjeev Arora","Nadav Cohen","Wei Hu","Yuping Luo"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Implicit Regularization in Deep Matrix Factorization","plainExplanation":"No issue was found in the checked claims. The selected conditional counterexample proof is correct; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1905.13655v3","noteUrl":"/review?paper=review-io-1905-13655","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1806-00468","title":"Implicit Bias of Gradient Descent on Linear Convolutional Networks","authors":["Suriya Gunasekar","Jason Lee","Daniel Soudry","Nathan Srebro"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Implicit Bias of Gradient Descent on Linear Convolutional Networks","plainExplanation":"No issue was found in the checked claims. Induced-regulariser formulas reconstruct correctly. The final phase notation defect is locally repairable from the paper own polar definition and is not retained as a substantive unaddressed error.","primaryUrl":"https://arxiv.org/pdf/1806.00468v2","noteUrl":"/review?paper=review-io-1806-00468","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1902-02918","title":"Certified Adversarial Robustness via Randomized Smoothing","authors":["Jeremy M Cohen","Elan Rosenfeld","J. Zico Kolter"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Certified Adversarial Robustness via Randomized Smoothing","plainExplanation":"No issue was found in the checked claims. Selected certificate proof reconstructs correctly; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1902.02918v2","noteUrl":"/review?paper=review-io-1902-02918","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1805-12152","title":"Robustness May Be at Odds with Accuracy","authors":["Dimitris Tsipras","Shibani Santurkar","Logan Engstrom","Alexander Turner","Aleksander Madry"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Robustness May Be at Odds with Accuracy","plainExplanation":"No issue was found in the checked claims. Full selected theorem proof verifies; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1805.12152v5","noteUrl":"/review?paper=review-io-1805-12152","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1804-11285","title":"Adversarially Robust Generalization Requires More Data","authors":["Ludwig Schmidt","Shibani Santurkar","Dimitris Tsipras","Kunal Talwar","Aleksander Mądry"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Adversarially Robust Generalization Requires More Data","plainExplanation":"No issue was found in the checked claims. The selected robust-event and tail-bound proof works; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1804.11285v2","noteUrl":"/review?paper=review-io-1804-11285","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"trades-loss-domination","title":"Theoretically Principled Trade-off between Robustness and Accuracy","authors":["Hongyang Zhang","Yaodong Yu","Jiantao Jiao","Eric P. Xing","Laurent El Ghaoui","Michael I. Jordan"],"year":2019,"category":"Optimization & generalization","status":"finding","issueTitle":"The TRADES bound needs more than classification calibration","plainExplanation":"Classification calibration does not guarantee that a surrogate loss bounds the wrong-sign indicator. The binary TRADES theorem needs $\\phi(t)\\ge1$ for $t\\le0$; a non-increasing loss with $\\phi(0)\\ge1$ satisfies this condition. A separate zero-score example shows why the full inequality chain also needs a consistent tie convention.","primaryUrl":"https://arxiv.org/pdf/1901.08573v3","citationCount":3147,"citationUrl":"https://www.semanticscholar.org/paper/6c405d4b5dc41a86be05acd59c06ed19daf01d14","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-io-1906-04893","title":"Efficient and Accurate Estimation of Lipschitz Constants for Deep Neural Networks","authors":["Mahyar Fazlyab","Alexander Robey","Hamed Hassani","Manfred Morari","George J. Pappas"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Efficient and Accurate Estimation of Lipschitz Constants for Deep Neural Networks","plainExplanation":"No issue was found in the checked claims. Current selected proof reconstructs correctly; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1906.04893v2","noteUrl":"/review?paper=review-io-1906-04893","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1610-02132","title":"QSGD: Communication-Efficient SGD via Gradient Quantization and Encoding","authors":["Dan Alistarh","Demjan Grubic","Jerry Li","Ryota Tomioka","Milan Vojnovic"],"year":2016,"category":"Optimization & generalization","status":"no-error","issueTitle":"QSGD: Communication-Efficient SGD via Gradient Quantization and Encoding","plainExplanation":"No issue was found in the checked claims. Selected quantisation moment proof verifies; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1610.02132v4","noteUrl":"/review?paper=review-io-1610-02132","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1705-07878","title":"TernGrad: Ternary Gradients to Reduce Communication in Distributed Deep Learning","authors":["Wei Wen","Cong Xu","Feng Yan","Chunpeng Wu","Yandan Wang","Yiran Chen","Hai Li"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"TernGrad: Ternary Gradients to Reduce Communication in Distributed Deep Learning","plainExplanation":"No issue was found in the checked claims. The paper own convergence-reduction algebra works under its stated assumptions; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1705.07878v6","noteUrl":"/review?paper=review-io-1705-07878","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1802-04434","title":"signSGD: Compressed Optimisation for Non-Convex Problems","authors":["Jeremy Bernstein","Yu-Xiang Wang","Kamyar Azizzadenesheli","Anima Anandkumar"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"signSGD: Compressed Optimisation for Non-Convex Problems","plainExplanation":"No issue was found in the checked claims. Selected rate derivation checks; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1802.04434v3","noteUrl":"/review?paper=review-io-1802-04434","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1901-09847","title":"Error Feedback Fixes SignSGD and other Gradient Compression Schemes","authors":["Sai Praneeth Karimireddy","Quentin Rebjock","Sebastian U. Stich","Martin Jaggi"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Error Feedback Fixes SignSGD and other Gradient Compression Schemes","plainExplanation":"No issue was found in the checked claims. Full selected residual-bound proof works; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1901.09847v2","noteUrl":"/review?paper=review-io-1901-09847","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1809-07599","title":"Sparsified SGD with Memory","authors":["Sebastian U. Stich","Jean-Baptiste Cordonnier","Martin Jaggi"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Sparsified SGD with Memory","plainExplanation":"No issue was found in the checked claims. Selected descent inequality reconstructs correctly; no retained issue.","primaryUrl":"https://arxiv.org/pdf/1809.07599v2","noteUrl":"/review?paper=review-io-1809-07599","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1812-06127","title":"Federated Optimization in Heterogeneous Networks","authors":["Tian Li","Anit Kumar Sahu","Manzil Zaheer","Maziar Sanjabi","Ameet Talwalkar","Virginia Smith"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Federated Optimization in Heterogeneous Networks","plainExplanation":"No issue was found in the checked claims. The selected one-round proof reconstructs under its sampling and dissimilarity assumptions.","primaryUrl":"https://arxiv.org/pdf/1812.06127v5","noteUrl":"/review?paper=review-io-1812-06127","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1910-06378","title":"SCAFFOLD: Stochastic Controlled Averaging for Federated Learning","authors":["Sai Praneeth Karimireddy","Satyen Kale","Mehryar Mohri","Sashank J. Reddi","Sebastian U. Stich","Ananda Theertha Suresh"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"SCAFFOLD: Stochastic Controlled Averaging for Federated Learning","plainExplanation":"The selected control-lag lemma follows. No claim that all SCAFFOLD convergence arguments have been checked.","primaryUrl":"https://arxiv.org/pdf/1910.06378v4","noteUrl":"/review?paper=review-io-1910-06378","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-2003-00295","title":"Adaptive Federated Optimization","authors":["Sashank Reddi","Zachary Charles","Manzil Zaheer","Zachary Garrett","Keith Rush","Jakub Konečný","Sanjiv Kumar","H. Brendan McMahan"],"year":2020,"category":"Optimization & generalization","status":"no-error","issueTitle":"Adaptive Federated Optimization","plainExplanation":"No issue was found in the checked claims. The printed statement uses 5K eta²(sigma_l²+2K sigma_g²)+30K² eta²||grad f||², while its proof derives a looser 6K heterogeneity factor. I independently verified the stated stronger bound: let M=max_{j≤K} average_i E||x_i,j−x||². Splitting the martingale noise sum from the predictable gradient sum gives M≤2 eta² K sigma_l²+4 eta² K²(||grad f(x)||²+sigma_g²)+4 eta² K² L² M. Since the last coefficient≤1/16, M≤(32/15)eta²K sigma_l²+(64/15)eta²K²(||grad f||²+sigma_g²), which is below every coefficient in the displayed statement. This also handles K=1 directly and avoids the proof’s intermediate 1/(K−1).","primaryUrl":"https://arxiv.org/pdf/2003.00295v5","noteUrl":"/review?paper=review-io-2003-00295","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1907-02189","title":"On the Convergence of FedAvg on Non-IID Data","authors":["Xiang Li","Kaixuan Huang","Wenhao Yang","Shusen Wang","Zhihua Zhang"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"On the Convergence of FedAvg on Non-IID Data","plainExplanation":"No issue was found in the checked claims. The selected sampling lemmas reconstruct under the two explicit sampling schemes.","primaryUrl":"https://arxiv.org/pdf/1907.02189v4","noteUrl":"/review?paper=review-io-1907-02189","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1805-09767","title":"Local SGD Converges Fast and Communicates Little","authors":["Sebastian U. Stich"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Local SGD Converges Fast and Communicates Little","plainExplanation":"No issue was found in the checked claims. The selected disagreement bound is correct under the actual-iterate second-moment assumption.","primaryUrl":"https://arxiv.org/pdf/1805.09767v3","noteUrl":"/review?paper=review-io-1805-09767","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-io-1607-00133","title":"Deep Learning with Differential Privacy","authors":["Martín Abadi","Andy Chu","Ian Goodfellow","H. Brendan McMahan","Ilya Mironov","Kunal Talwar","Li Zhang"],"year":2016,"category":"Optimization & generalization","status":"addressed","issueTitle":"Adaptive privacy losses cannot be factored as independent variables","plainExplanation":"Independent random seeds do not make adaptive privacy losses independent. The theorem remains valid by conditioning each moment bound on the released history and iterating the tower property, as the later Rényi-composition proof does.","primaryUrl":"https://arxiv.org/pdf/1607.00133v2","citationCount":8326,"citationUrl":"https://www.semanticscholar.org/paper/e9a986c8ff6c2f381d026fe014f6aaa865f34da7","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-io-1702-07476","title":"Renyi Differential Privacy","authors":["Ilya Mironov"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Renyi Differential Privacy","plainExplanation":"No issue was found in the checked claims. The selected composition, conversion, and Gaussian identities reconstruct.","primaryUrl":"https://arxiv.org/pdf/1702.07476v3","noteUrl":"/review?paper=review-io-1702-07476","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1605-02065","title":"Concentrated Differential Privacy: Simplifications, Extensions, and Lower Bounds","authors":["Mark Bun","Thomas Steinke"],"year":2016,"category":"Optimization & generalization","status":"no-error","issueTitle":"Concentrated Differential Privacy: Simplifications, Extensions, and Lower Bounds","plainExplanation":"No issue was found in the checked claims. The selected conversion proof reconstructs, with the degenerate rho=0 case supplied explicitly.","primaryUrl":"https://arxiv.org/pdf/1605.02065v1","noteUrl":"/review?paper=review-io-1605-02065","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1808-00087","title":"Subsampled R\\'enyi Differential Privacy and Analytical Moments Accountant","authors":["Yu-Xiang Wang","Borja Balle","Shiva Kasiviswanathan"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Subsampled R\\'enyi Differential Privacy and Analytical Moments Accountant","plainExplanation":"No issue was found in the checked claims. The selected amplification proposition follows under the indexed replacement model.","primaryUrl":"https://arxiv.org/pdf/1808.00087v2","noteUrl":"/review?paper=review-io-1808-00087","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1807-01647","title":"Privacy Amplification by Subsampling: Tight Analyses via Couplings and Divergences","authors":["Borja Balle","Gilles Barthe","Marco Gaboardi"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Privacy Amplification by Subsampling: Tight Analyses via Couplings and Divergences","plainExplanation":"No issue was found in the checked claims. The selected exact identity and amplification bound reconstruct.","primaryUrl":"https://arxiv.org/pdf/1807.01647v2","noteUrl":"/review?paper=review-io-1807-01647","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1610-05755","title":"Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data","authors":["Nicolas Papernot","Martín Abadi","Úlfar Erlingsson","Ian Goodfellow","Kunal Talwar"],"year":2016,"category":"Optimization & generalization","status":"no-error","issueTitle":"Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data","plainExplanation":"No issue was found in the checked claims. The selected single-query moment and vote-tail proofs reconstruct.","primaryUrl":"https://arxiv.org/pdf/1610.05755v4","noteUrl":"/review?paper=review-io-1610-05755","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-io-1802-08908","title":"Scalable Private Learning with PATE","authors":["Nicolas Papernot","Shuang Song","Ilya Mironov","Ananth Raghunathan","Kunal Talwar","Úlfar Erlingsson"],"year":2018,"category":"Optimization & generalization","status":"addressed","issueTitle":"The refined privacy bound needs a positive denominator","plainExplanation":"The theorem can invert a negative lower bound on a probability, producing a negative upper bound on zero divergence. Requiring \\(\\widetilde q e^{\\epsilon_2}<1\\) repairs that step; the official accountant already checks the condition.","primaryUrl":"https://arxiv.org/pdf/1802.08908v1","citationCount":734,"citationUrl":"https://www.semanticscholar.org/paper/44058a625cb64c311043145655645d8206e272c2","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-io-2004-00010","title":"The Discrete Gaussian for Differential Privacy","authors":["Clément L. Canonne","Gautam Kamath","Thomas Steinke"],"year":2020,"category":"Optimization & generalization","status":"no-error","issueTitle":"The Discrete Gaussian for Differential Privacy","plainExplanation":"No issue was found in the checked claims. The selected discrete-Gaussian divergence proof reconstructs.","primaryUrl":"https://arxiv.org/pdf/2004.00010v6","noteUrl":"/review?paper=review-io-2004-00010","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1811-12469","title":"Amplification by Shuffling: From Local to Central Differential Privacy via Anonymity","authors":["Úlfar Erlingsson","Vitaly Feldman","Ilya Mironov","Ananth Raghunathan","Kunal Talwar","Abhradeep Thakurta"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Amplification by Shuffling: From Local to Central Differential Privacy via Anonymity","plainExplanation":"No issue was found in the checked claims. The selected adaptive shuffling bound reconstructs after making the relevant prefix explicit.","primaryUrl":"https://arxiv.org/pdf/1811.12469v2","noteUrl":"/review?paper=review-io-1811-12469","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1905-03871","title":"Differentially Private Learning with Adaptive Clipping","authors":["Galen Andrew","Om Thakkar","H. Brendan McMahan","Swaroop Ramaswamy"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Differentially Private Learning with Adaptive Clipping","plainExplanation":"No issue was found in the checked claims. The selected relative privacy-accounting relation reconstructs as a common conservative bound.","primaryUrl":"https://arxiv.org/pdf/1905.03871v5","noteUrl":"/review?paper=review-io-1905-03871","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1909-04746","title":"Tighter Theory for Local SGD on Identical and Heterogeneous Data","authors":["Ahmed Khaled","Konstantin Mishchenko","Peter Richtárik"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Tighter Theory for Local SGD on Identical and Heterogeneous Data","plainExplanation":"No issue was found in the checked claims. Both selected heterogeneous one-step inequalities reconstruct.","primaryUrl":"https://arxiv.org/pdf/1909.04746v4","noteUrl":"/review?paper=review-io-1909-04746","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1807-06629","title":"Parallel Restarted SGD with Faster Convergence and Less Communication: Demystifying Why Model Averaging Works for Deep Learning","authors":["Hao Yu","Sen Yang","Shenghuo Zhu"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Parallel Restarted SGD with Faster Convergence and Less Communication: Demystifying Why Model Averaging Works for Deep Learning","plainExplanation":"No issue was found in the checked claims. The selected nonconvex stationarity bound reconstructs.","primaryUrl":"https://arxiv.org/pdf/1807.06629v3","noteUrl":"/review?paper=review-io-1807-06629","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1705-09056","title":"Can Decentralized Algorithms Outperform Centralized Algorithms? A Case Study for Decentralized Parallel Stochastic Gradient Descent","authors":["Xiangru Lian","Ce Zhang","Huan Zhang","Cho-Jui Hsieh","Wei Zhang","Ji Liu"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Can Decentralized Algorithms Outperform Centralized Algorithms? A Case Study for Decentralized Parallel Stochastic Gradient Descent","plainExplanation":"No issue was found in the checked claims. The selected mixing and gradient-control lemmas reconstruct.","primaryUrl":"https://arxiv.org/pdf/1705.09056v5","noteUrl":"/review?paper=review-io-1705-09056","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1902-00340","title":"Decentralized Stochastic Optimization and Gossip Algorithms with Compressed Communication","authors":["Anastasia Koloskova","Sebastian U. Stich","Martin Jaggi"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Decentralized Stochastic Optimization and Gossip Algorithms with Compressed Communication","plainExplanation":"No issue was found in the checked claims. Theorem 2 states a step size proportional to δ²ω/D, which can be too small to attain its claimed per-step contraction. Appendix C instead uses δω/D and proves the corresponding rate. The addressing evidence is already in the same paper; it is not a separate erratum or a change to the stochastic-optimization analysis.","primaryUrl":"https://arxiv.org/pdf/1902.00340v1","noteUrl":"/review?paper=review-io-1902-00340","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1705-10467","title":"Federated Multi-Task Learning","authors":["Virginia Smith","Chao-Kai Chiang","Maziar Sanjabi","Ameet Talwalkar"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Federated Multi-Task Learning","plainExplanation":"No issue was found in the checked claims. The selected surrogate-majorisation argument reconstructs.","primaryUrl":"https://arxiv.org/pdf/1705.10467v2","noteUrl":"/review?paper=review-io-1705-10467","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1902-00146","title":"Agnostic Federated Learning","authors":["Mehryar Mohri","Gary Sivek","Ananda Theertha Suresh"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Agnostic Federated Learning","plainExplanation":"No issue was found in the checked claims. The selected constructive separation is correct.","primaryUrl":"https://arxiv.org/pdf/1902.00146v1","noteUrl":"/review?paper=review-io-1902-00146","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-2006-08848","title":"Personalized Federated Learning with Moreau Envelopes","authors":["Canh T. Dinh","Nguyen H. Tran","Tuan Dung Nguyen"],"year":2020,"category":"Optimization & generalization","status":"no-error","issueTitle":"Personalized Federated Learning with Moreau Envelopes","plainExplanation":"No issue was found in the checked claims. The selected envelope-diversity inequalities reconstruct.","primaryUrl":"https://arxiv.org/pdf/2006.08848v3","noteUrl":"/review?paper=review-io-2006-08848","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-2012-04221","title":"Ditto: Fair and Robust Federated Learning Through Personalization","authors":["Tian Li","Shengyuan Hu","Ahmad Beirami","Virginia Smith"],"year":2020,"category":"Optimization & generalization","status":"unresolved","issueTitle":"The printed squared loss and optimizer use different normalizations","plainExplanation":"Differentiating the literal objective gives a factor of two absent from the displayed optimizer. The intended half-loss convention is already reflected in the subsequent calculation and author notebook.","primaryUrl":"https://arxiv.org/pdf/2012.04221v3","noteUrl":"/review?paper=review-io-2012-04221","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-io-2111-04263","title":"Federated Learning Based on Dynamic Regularization","authors":["Durmus Alp Emre Acar","Yue Zhao","Ramon Matas Navarro","Matthew Mattina","Paul N. Whatmough","Venkatesh Saligrama"],"year":2021,"category":"Optimization & generalization","status":"no-error","issueTitle":"Federated Learning Based on Dynamic Regularization","plainExplanation":"No issue was found in the checked claims. The selected drift and second-moment identities reconstruct.","primaryUrl":"https://arxiv.org/pdf/2111.04263v2","noteUrl":"/review?paper=review-io-2111-04263","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1810-02281","title":"A Convergence Analysis of Gradient Descent for Deep Linear Neural Networks","authors":["Sanjeev Arora","Nadav Cohen","Noah Golowich","Wei Hu"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"A Convergence Analysis of Gradient Descent for Deep Linear Neural Networks","plainExplanation":"No issue was found in the checked claims. The endpoint singular-value cases in equation (18) reverse the input/output dimension inequalities, so the displayed branch can bound the wrong factor. The proof on page 19 gives the correct branch conditions. This is a same-paper case-label repair, not a refutation of the balanced-product estimates.","primaryUrl":"https://arxiv.org/pdf/1810.02281v3","noteUrl":"/review?paper=review-io-1810-02281","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1901-08584","title":"Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural Networks","authors":["Sanjeev Arora","Simon S. Du","Wei Hu","Zhiyuan Li","Ruosong Wang"],"year":2019,"category":"Optimization & generalization","status":"no-error","issueTitle":"Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural Networks","plainExplanation":"No issue was found in the checked claims. The selected target-complexity bound reconstructs.","primaryUrl":"https://arxiv.org/pdf/1901.08584v2","noteUrl":"/review?paper=review-io-1901-08584","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1706-08947","title":"Exploring Generalization in Deep Learning","authors":["Behnam Neyshabur","Srinadh Bhojanapalli","David McAllester","Nathan Srebro"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Exploring Generalization in Deep Learning","plainExplanation":"No issue was found in the checked claims. The selected fixed-activation perturbation bound reconstructs under its noncancellation assumptions.","primaryUrl":"https://arxiv.org/pdf/1706.08947v2","noteUrl":"/review?paper=review-io-1706-08947","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-io-1902-09843","title":"Adaptive Gradient Methods with Dynamic Bound of Learning Rate","authors":["Liangchen Luo","Yuanhao Xiong","Yan Liu","Xu Sun"],"year":2019,"category":"Optimization & generalization","status":"addressed","issueTitle":"Clipping does not ensure monotone inverse step sizes","plainExplanation":"The regret proof assumes that inverse adaptive step sizes are nondecreasing. The clipping conditions do not ensure this, leaving the bound on the signed telescoping terms unjustified. Savarese’s On the Convergence of AdaBound and its Connection to SGD identifies the error, gives a slow-convergence counterexample and derives stronger sufficient conditions.","primaryUrl":"https://arxiv.org/pdf/1902.09843v1","citationCount":696,"citationUrl":"https://www.semanticscholar.org/paper/03af562fb8e69677865dbe94910e464443dd4623","severity":"repairable","issueTypes":["incorrect_bound","proof_gap"],"classifications":["repairable"]},{"slug":"review-io-1904-09237","title":"On the Convergence of Adam and Beyond","authors":["Sashank J. Reddi","Satyen Kale","Sanjiv Kumar"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"On the Convergence of Adam and Beyond","plainExplanation":"No issue was found in the checked claims. The selected energy bound reconstructs.","primaryUrl":"https://arxiv.org/pdf/1904.09237v1","noteUrl":"/review?paper=review-io-1904-09237","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1908-03265","title":"On the Variance of the Adaptive Learning Rate and Beyond","authors":["Liyuan Liu","Haoming Jiang","Pengcheng He","Weizhu Chen","Xiaodong Liu","Jianfeng Gao","Jiawei Han"],"year":2019,"category":"Optimization & generalization","status":"unresolved","issueTitle":"An auxiliary gamma-function bound has an incorrect constant","plainExplanation":"The stronger inequality is false, but the preceding required bound follows directly from log-convexity of the gamma function. No change to the variance formula or monotonicity theorem is needed.","primaryUrl":"https://arxiv.org/pdf/1908.03265v4","noteUrl":"/review?paper=review-io-1908-03265","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-io-1711-00851","title":"Provable defenses against adversarial examples via the convex outer adversarial polytope","authors":["Eric Wong","J. Zico Kolter"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Provable defenses against adversarial examples via the convex outer adversarial polytope","plainExplanation":"No issue was found in the checked claims. The selected certificate/loss upper bound reconstructs from dual feasibility.","primaryUrl":"https://arxiv.org/pdf/1711.00851v3","noteUrl":"/review?paper=review-io-1711-00851","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1710-10571","title":"Certifying Some Distributional Robustness with Principled Adversarial Training","authors":["Aman Sinha","Hongseok Namkoong","Riccardo Volpi","John Duchi"],"year":2017,"category":"Statistical learning","status":"no-error","issueTitle":"Certifying Some Distributional Robustness with Principled Adversarial Training","plainExplanation":"No issue was found in the checked claims. The selected stability and smoothness constants reconstruct conditional on well-defined maximizers.","primaryUrl":"https://arxiv.org/pdf/1710.10571v5","noteUrl":"/review?paper=review-io-1710-10571","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1804-09699","title":"Towards Fast Computation of Certified Robustness for ReLU Networks","authors":["Tsui-Wei Weng","Huan Zhang","Hongge Chen","Zhao Song","Cho-Jui Hsieh","Duane Boning","Inderjit S. Dhillon","Luca Daniel"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Towards Fast Computation of Certified Robustness for ReLU Networks","plainExplanation":"No issue was found in the checked claims. The selected affine certification argument reconstructs.","primaryUrl":"https://arxiv.org/pdf/1804.09699v4","noteUrl":"/review?paper=review-io-1804-09699","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-io-1801-09344","title":"Certified Defenses against Adversarial Examples","authors":["Aditi Raghunathan","Jacob Steinhardt","Percy Liang"],"year":2018,"category":"Optimization & generalization","status":"addressed","issueTitle":"The SDP matrix places its constant block on the wrong coordinates","plainExplanation":"The printed block order represents a different quadratic form and can underestimate the required perturbation bound. Moving the constant block restores the SDP certificate. The official implementation already uses the equivalent correct construction.","primaryUrl":"https://arxiv.org/pdf/1801.09344v2","citationCount":1021,"citationUrl":"https://www.semanticscholar.org/paper/966e3c7a65ec75a6359b55c0cecaf3896d318432","severity":"repairable","issueTypes":["formula_error","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-io-1705-08475","title":"Formal Guarantees on the Robustness of a Classifier against Adversarial Manipulation","authors":["Matthias Hein","Maksym Andriushchenko"],"year":2017,"category":"Robotics & action models","status":"no-error","issueTitle":"Formal Guarantees on the Robustness of a Classifier against Adversarial Manipulation","plainExplanation":"No issue was found in the checked claims. The quantitative margin/radius inequality reconstructs.","primaryUrl":"https://arxiv.org/pdf/1705.08475v2","noteUrl":"/review?paper=review-io-1705-08475","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1801-10578","title":"Evaluating the Robustness of Neural Networks: An Extreme Value Theory Approach","authors":["Tsui-Wei Weng","Huan Zhang","Pin-Yu Chen","Jinfeng Yi","Dong Su","Yupeng Gao","Cho-Jui Hsieh","Luca Daniel"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Evaluating the Robustness of Neural Networks: An Extreme Value Theory Approach","plainExplanation":"No issue was found in the checked claims. The exact-Lipschitz quantitative bound reconstructs, with the same boundary-tie qualification as Hein–Andriushchenko.","primaryUrl":"https://arxiv.org/pdf/1801.10578v1","noteUrl":"/review?paper=review-io-1801-10578","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1802-04034","title":"Lipschitz-Margin Training: Scalable Certification of Perturbation Invariance for Deep Neural Networks","authors":["Yusuke Tsuzuku","Issei Sato","Masashi Sugiyama"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Lipschitz-Margin Training: Scalable Certification of Perturbation Invariance for Deep Neural Networks","plainExplanation":"No issue was found in the checked claims. Both selected inequalities reconstruct.","primaryUrl":"https://arxiv.org/pdf/1802.04034v3","noteUrl":"/review?paper=review-io-1802-04034","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1802-03471","title":"Certified Robustness to Adversarial Examples with Differential Privacy","authors":["Mathias Lecuyer","Vaggelis Atlidakis","Roxana Geambasu","Daniel Hsu","Suman Jana"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Certified Robustness to Adversarial Examples with Differential Privacy","plainExplanation":"No issue was found in the checked claims. The exact-expected-score certificate reconstructs.","primaryUrl":"https://arxiv.org/pdf/1802.03471v4","noteUrl":"/review?paper=review-io-1802-03471","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1811-00866","title":"Efficient Neural Network Robustness Certification with General Activation Functions","authors":["Huan Zhang","Tsui-Wei Weng","Pin-Yu Chen","Cho-Jui Hsieh","Luca Daniel"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"Efficient Neural Network Robustness Certification with General Activation Functions","plainExplanation":"No issue was found in the checked claims. The selected output bound and dual-norm optimization reconstruct.","primaryUrl":"https://arxiv.org/pdf/1811.00866v1","noteUrl":"/review?paper=review-io-1811-00866","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-io-1803-06567","title":"A Dual Approach to Scalable Verification of Deep Networks","authors":["Krishnamurthy","Dvijotham","Robert Stanforth","Sven Gowal","Timothy Mann","Pushmeet Kohli"],"year":2018,"category":"Optimization & generalization","status":"no-error","issueTitle":"A Dual Approach to Scalable Verification of Deep Networks","plainExplanation":"No issue was found in the checked claims. The chosen dual-feasibility and ReLU conjugate argument reconstructs.","primaryUrl":"https://arxiv.org/pdf/1803.06567v2","noteUrl":"/review?paper=review-io-1803-06567","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-mbio-atom-001","title":"MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields","authors":["Ilyes Batatia","Dávid Péter Kovács","Gregor N. C. Simm","Christoph Ortner","Gábor Csányi"],"year":2022,"category":"Graphs & equivariance","status":"no-error","issueTitle":"MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields","plainExplanation":"No error found in the selected density-factorization and contraction argument. Its inclusion of repeated indices is essential and explicitly stated.","primaryUrl":"https://arxiv.org/abs/2206.07697","noteUrl":"/review?paper=review-mbio-atom-001","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-mbio-atom-002","title":"TorchMD-NET: Equivariant Transformers for Neural Network based Molecular Potentials","authors":["Philipp Thölke","Gianni De Fabritiis"],"year":2022,"category":"Molecular models & equivariance","status":"addressed","issueTitle":"The vector message can remain nonzero at the distance cutoff","plainExplanation":"A biased distance filter can remain nonzero after its radial inputs vanish. The scalar branch is tapered again, but the printed vector branch is not. Applying the cutoff to the entire vector message restores continuity under continuous downstream operations; the official code exposes this as an option.","primaryUrl":"https://arxiv.org/abs/2202.02541","citationCount":303,"citationUrl":"https://www.semanticscholar.org/paper/68ab47dbf07dcff861548ee37c906b6793dfd75c","severity":"repairable","issueTypes":["algorithm_specification"],"classifications":["repairable"]},{"slug":"review-mbio-atom-003","title":"PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments, and Partial Charges.","authors":["Oliver T. Unke","Markus Meuwly"],"year":2019,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments, and Partial Charges.","plainExplanation":"No error found in the selected charge correction and damped-electrostatics derivation.","primaryUrl":"https://arxiv.org/abs/1902.08408","noteUrl":"/review?paper=review-mbio-atom-003","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-mbio-atom-004","title":"SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects","authors":["Oliver T. Unke","Stefan Chmiela","Michael Gastegger","Kristof T. Schütt","Huziel E. Sauceda","Klaus-Robert Müller"],"year":2021,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects","plainExplanation":"No error found in the selected invariance and square-versus-tetrahedron expressivity argument.","primaryUrl":"https://arxiv.org/abs/2105.00304","noteUrl":"/review?paper=review-mbio-atom-004","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-mbio-atom-005","title":"Moment Tensor Potentials: A Class of Systematically Improvable Interatomic Potentials","authors":["Alexander V. Shapeev"],"year":2015,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Moment Tensor Potentials: A Class of Systematically Improvable Interatomic Potentials","plainExplanation":"No error found in the selected exact contraction lemma and its invariance consequence.","primaryUrl":"https://arxiv.org/abs/1512.06054","noteUrl":"/review?paper=review-mbio-atom-005","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-mbio-atom-006","title":"End-to-end Symmetry Preserving Inter-atomic Potential Energy Model for Finite and Extended Systems","authors":["Linfeng Zhang","Jiequn Han","Han Wang","Wissam A. Saidi","Roberto Car","Weinan E"],"year":2018,"category":"Graphs & equivariance","status":"no-error","issueTitle":"End-to-end Symmetry Preserving Inter-atomic Potential Energy Model for Finite and Extended Systems","plainExplanation":"No error found in Proposition1 in the specified domain. Its statement concerns invariance, which is distinct from injectivity of the compressed descriptor.","primaryUrl":"https://arxiv.org/abs/1805.09003","noteUrl":"/review?paper=review-mbio-atom-006","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-mbio-atom-007","title":"Quantum-chemical insights from deep tensor neural networks","authors":["Kristof T. Schütt","Farhad Arbabzadah","Stefan Chmiela","Klaus R. Müller","Alexandre Tkatchenko"],"year":2016,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Quantum-chemical insights from deep tensor neural networks","plainExplanation":"No error found in the selected low-rank factorization and symmetry/dependency arguments.","primaryUrl":"https://arxiv.org/abs/1609.08259","noteUrl":"/review?paper=review-mbio-atom-007","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-bio-gen-01","title":"Equivariant Diffusion for Molecule Generation in 3D","authors":["Emiel Hoogeboom","Victor Garcia Satorras","Clément Vignac","Max Welling"],"year":2022,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Equivariant Diffusion for Molecule Generation in 3D","plainExplanation":"No error found in the selected equal-variance constrained-Gaussian KL argument or its orthogonal invariance consequence.","primaryUrl":"https://arxiv.org/pdf/2203.17003","noteUrl":"/review?paper=review-bio-gen-01","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-bio-gen-02","title":"GeoDiff: a Geometric Diffusion Model for Molecular Conformation Generation","authors":["Minkai Xu","Lantao Yu","Yang Song","Chence Shi","Stefano Ermon","Jian Tang"],"year":2022,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"GeoDiff: a Geometric Diffusion Model for Molecular Conformation Generation","plainExplanation":"No error found in Proposition2 on noncoincident interacting atoms with the vector-field group action used in the proof.","primaryUrl":"https://arxiv.org/pdf/2203.02923","noteUrl":"/review?paper=review-bio-gen-02","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-bio-gen-03","title":"Torsional Diffusion for Molecular Conformer Generation","authors":["Bowen Jing","Gabriele Corso","Jeffrey Chang","Regina Barzilay","Tommi Jaakkola"],"year":2022,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Torsional Diffusion for Molecular Conformer Generation","plainExplanation":"No error found in the selected torsion-update and parity-score results with the stated geometric and density assumptions.","primaryUrl":"https://arxiv.org/pdf/2206.01729","noteUrl":"/review?paper=review-bio-gen-03","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-bio-gen-04","title":"GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation","authors":["Chence Shi","Minkai Xu","Zhaocheng Zhu","Weinan Zhang","Ming Zhang","Jian Tang"],"year":2020,"category":"Molecular generation","status":"addressed","issueTitle":"A BFS edge can span more positions than one level contains","plainExplanation":"A BFS edge can span more positions than the largest single level contains. The authors’ linked TorchDrug implementation instead measures actual endpoint gaps for the orderings processed. This addresses that window calculation without establishing a bound for every later random ordering or a failure on the reported datasets.","primaryUrl":"https://arxiv.org/pdf/2001.09382","citationCount":568,"citationUrl":"https://www.semanticscholar.org/paper/036d743c7ca1e513adf0a91594fc8111e03dc30c","severity":"repairable","issueTypes":["algorithm_specification"],"classifications":["repairable"]},{"slug":"review-bio-gen-05","title":"MoFlow: An Invertible Flow Model for Generating Molecular Graphs","authors":["Chengxi Zang","Fei Wang"],"year":2020,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"MoFlow: An Invertible Flow Model for Generating Molecular Graphs","plainExplanation":"No error found in the selected conditional coupling inverse and determinant calculation.","primaryUrl":"https://arxiv.org/pdf/2006.10137","noteUrl":"/review?paper=review-bio-gen-05","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-bio-gen-06","title":"DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking","authors":["Gabriele Corso","Hannes Stärk","Bowen Jing","Regina Barzilay","Tommi Jaakkola"],"year":2022,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking","plainExplanation":"No error found in the selected infinitesimal alignment argument and the stated nondegenerate interpretation of the bijection proof.","primaryUrl":"https://arxiv.org/pdf/2210.01776","noteUrl":"/review?paper=review-bio-gen-06","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-bio-gen-07","title":"Junction Tree Variational Autoencoder for Molecular Graph Generation","authors":["Wengong Jin","Regina Barzilay","Tommi Jaakkola"],"year":2018,"category":"Molecular generation","status":"addressed","issueTitle":"Singleton insertion can break the junction-tree property","plainExplanation":"For two fused rings and a pendant bond, the printed singleton construction separates the rings while omitting a shared atom. The released code distinguishes this bond/ring case and uses overlap weights, avoiding the example. A general junction-tree guarantee still requires connectedness of every atom’s cluster set.","primaryUrl":"https://arxiv.org/pdf/1802.04364","citationCount":1716,"citationUrl":"https://www.semanticscholar.org/paper/fd17bd9a5dc24a081ad9743570f50dd6750f54b2","severity":"repairable","issueTypes":["algorithm_specification"],"classifications":["repairable"]},{"slug":"review-bio-gen-08","title":"Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning","authors":["Frank Noé","Simon Olsson","Jonas Köhler","Hao Wu"],"year":2019,"category":"Generative models","status":"no-error","issueTitle":"Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning","plainExplanation":"No error found in the selected latent-KL/free-energy and importance-reweighting derivations under the change-of-variables assumptions.","primaryUrl":"https://aiichironakano.github.io/phys516/Noe-BoltzmannGenerator-Science19.pdf","noteUrl":"/review?paper=review-bio-gen-08","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-bio-ml-01","title":"Pre-training Molecular Graph Representation with 3D Geometry","authors":["Shengchao Liu","Hanchen Wang","Weiyang Liu","Joan Lasenby","Hongyu Guo","Jian Tang"],"year":2022,"category":"Molecular representation learning","status":"addressed","issueTitle":"The entropy and KL terms change which bounds hold","plainExplanation":"The mutual-information derivation needs marginal entropy offsets for continuous variables. A beta-weighted reconstruction objective is not automatically an evidence lower bound below unit KL weight. The authors’ later MoleculeSDE appendix supplies the entropy constants, a unit-weight ELBO and a separately labeled beta-weighted proxy; it is a related correct formulation rather than an erratum to this version.","primaryUrl":"https://arxiv.org/abs/2110.07728v2","citationCount":489,"citationUrl":"https://www.semanticscholar.org/paper/b8f816e23ff40d6afabccca2ee4770087ef0ef57","severity":"repairable","issueTypes":["formula_error","interpretation_error"],"classifications":["repairable"]},{"slug":"review-bio-ml-02","title":"InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization","authors":["Fan-Yun Sun","Jordan Hoffmann","Vikas Verma","Jian Tang"],"year":2020,"category":"Graphs & equivariance","status":"no-error","issueTitle":"InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization","plainExplanation":"No error found in the selected logistic variational objective under its joint/product sampling interpretation.","primaryUrl":"https://arxiv.org/abs/1908.01000v3","noteUrl":"/review?paper=review-bio-ml-02","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"addressed-bio-ml-03","title":"Graph Contrastive Learning with Augmentations","authors":["Yuning You","Tianlong Chen","Yongduo Sui","Ting Chen","Zhangyang Wang","Yang Shen"],"year":2020,"category":"Molecular representation learning","status":"addressed","issueTitle":"The original loss formula included the positive pair in its denominator","plainExplanation":"The positive-pair denominator mismatch in v1 was acknowledged and corrected in v2. Current v3 retains the negative-pair-only expression. This was already documented in the addressed note; it does not establish that a finite-batch mutual-information theorem was repaired.","primaryUrl":"https://arxiv.org/pdf/2010.13902v1","citationCount":2978,"citationUrl":"https://www.semanticscholar.org/paper/76c124786ccf4263e6403a15a8e350ac28be4e65","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-bio-scvi","title":"Deep Generative Modeling for Single-cell Transcriptomics","authors":["Romain Lopez","Jeffrey Regier","Michael B. Cole","Michael I. Jordan","Nir Yosef"],"year":2018,"category":"Generative models","status":"no-error","issueTitle":"Deep Generative Modeling for Single-cell Transcriptomics","plainExplanation":"No issue was found in the checked claims. The collapsed likelihood and Eq. (2) check out under the stated parameterisation. Approximate posterior probabilities are not exact frequentist error guarantees.","primaryUrl":"https://www.ebi.ac.uk/europepmc/webservices/rest/PMC6289068/fullTextXML","noteUrl":"/review?paper=review-bio-scvi","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-bio-totalvi","title":"Joint probabilistic modeling of single-cell multi-omic data with totalVI","authors":["Adam Gayoso","Zoë Steier","Romain Lopez","Jeffrey Regier","Kristopher L Nazor","Aaron Streets","Nir Yosef"],"year":2021,"category":"Statistical learning","status":"no-error","issueTitle":"Joint probabilistic modeling of single-cell multi-omic data with totalVI","plainExplanation":"No issue was found in the checked claims. The selected likelihood, factorisation and observed-data marginalisation arguments are consistent. Background/foreground mean ordering is weaker than perfect assignment or biological identifiability.","primaryUrl":"https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7954949/fullTextXML","noteUrl":"/review?paper=review-bio-totalvi","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-bio-scanvi","title":"Probabilistic harmonization and annotation of single‐cell transcriptomics data with deep generative models","authors":["Xu, Chenling","Lopez, Romain","Mehlman, Edouard","Regier, Jeffrey","Jordan, Michael I","Yosef, Nir"],"year":2021,"category":"Generative models","status":"unresolved","issueTitle":"Exact collapsing tightens the fixed-family ELBO","plainExplanation":"Exact marginalization removes a nonnegative conditional KL penalty. At the same retained-variable distribution, it therefore gives a tighter or equal bound, not a looser one.","primaryUrl":"https://d-nb.info/1226443095/34","noteUrl":"/review?paper=review-bio-scanvi","severity":"minor_cosmetic","issueTypes":["interpretation_error"],"classifications":["minor_cosmetic"]},{"slug":"review-bio-cellot","title":"Learning single-cell perturbation responses using neural optimal transport","authors":["Bunne, Charlotte","Stark, Stefan G.","Gut, Gabriele","del Castillo, Jacobo Sarabia","Levesque, Mitch","Lehmann, Kjong-Van","Pelkmans, Lucas","Krause, Andreas","Rätsch, Gunnar"],"year":2023,"category":"Statistical learning","status":"unresolved","issueTitle":"The transport value and coupling need consistent conventions","plainExplanation":"The dual computes half the value assigned by the primal. The displayed plan places the target marginal before the source marginal. These are local convention errors; the transport direction and subsequent saddle construction are consistent.","primaryUrl":"https://www.nature.com/articles/s41592-023-01969-x.pdf","noteUrl":"/review?paper=review-bio-cellot","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-bio-icnn","title":"Input Convex Neural Networks","authors":["Brandon Amos","Lei Xu","J. Zico Kolter"],"year":2017,"category":"Optimization & generalization","status":"no-error","issueTitle":"Input Convex Neural Networks","plainExplanation":"No issue was found in the checked claims. The main convexity and representation arguments hold under the stated conditions. The LP optimal-value argument is sound with that final-layer qualification; its informal assertion that all optimiser activations must be exact is stronger than needed when downstream weights vanish.","primaryUrl":"https://proceedings.mlr.press/v70/amos17b/amos17b.pdf","noteUrl":"/review?paper=review-bio-icnn","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-arxiv-2405-15943","title":"Transformers Represent Belief State Geometry in Their Residual Stream","authors":["Adam S. Shai","Sarah E. Marzen","Lucas Teixeira","Alexander Gietelink Oldenziel","Paul M. Riechers"],"year":2024,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Transformers Represent Belief State Geometry in Their Residual Stream","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/abs/2405.15943","noteUrl":"/review?paper=review-arxiv-2405-15943","citationCount":76,"citationUrl":"https://www.semanticscholar.org/paper/f481cf18598b8e4c069e6aced38b71b654db5dc3","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"constrained-belief-updates-normalisation","title":"Constrained Belief Updates Explain Geometric Structures in Transformer Representations","authors":["Mateusz Piotrowski","Paul M. Riechers","Daniel Filan","Adam S. Shai"],"year":2025,"category":"Transformers & in-context learning","status":"finding","issueTitle":"Averaging row-conditioned transitions does not give the posterior","plainExplanation":"An observed symbol changes which predecessor states are plausible as well as where each state can transition next. The disputed expression conditions the transitions but continues averaging them with the original predecessor weights. For the valid three-state example below, it gives (0.5224, 0.2388, 0.2388) instead of the Bayesian posterior (0.6, 0.2, 0.2). In this symmetric model, the centered vectors differ only by a common scale, which a fitted linear map can absorb.","primaryUrl":"https://proceedings.mlr.press/v267/piotrowski25a.html","citationCount":16,"citationUrl":"https://www.semanticscholar.org/paper/decab8df62433e06a202da0d85761e78dae1d4c8","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138873","title":"SciHazard: A Benchmark for Measuring Scientific Safety Risks with Decomposed Harm Scoring","authors":["Chunxiao Li","Yuan Xiong","Lijun Li","Tianyi Du","Wenlong Zhang","LEI BAI","Jing Shao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SciHazard: A Benchmark for Measuring Scientific Safety Risks with Decomposed Harm Scoring","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.18665","noteUrl":"/review?paper=review-neurips-2026-138873","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138874","title":"ColdDDI: Evaluating Knowledge Utilization in Cold-Start Drug-Drug Interaction Prediction","authors":["Jiheng Liang","Chen Zhao","Di Wu","Chenyang Bu","Yunpeng Hong","Xingquan Zhu","Yi He"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"ColdDDI: Evaluating Knowledge Utilization in Cold-Start Drug-Drug Interaction Prediction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.05590","noteUrl":"/review?paper=review-neurips-2026-138874","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138876","title":"ScrapeGraphAI-100k: Dataset for Schema-Constrained LLM Generation","authors":["William Brach","Francesco Zuppichini","Lorenzo Padoan","Marco Vinciguerra"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"ScrapeGraphAI-100k: Dataset for Schema-Constrained LLM Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.15189","noteUrl":"/review?paper=review-neurips-2026-138876","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138878","title":"Re:Cognize - A Framework for Open-Set Sequential Character Re-Identification","authors":["Aaditya Baranwal","Madhav Kataria","Yogesh Rawat","Shruti Vyas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Re:Cognize - A Framework for Open-Set Sequential Character Re-Identification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34032","noteUrl":"/review?paper=review-neurips-2026-138878","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138879","title":"AI GAMESTORE: Scalable, Open-Ended Evaluation of Machine General Intelligence with Human Games","authors":["Lance Ying","Ryan Truong","Prafull Sharma","Kaiya Zhao","Nathan Cloos","Kelsey Allen","Tom Griffiths","Katie Collins","Jose Hernandez-Orallo","Phillip Isola","Samuel J Gershman","Josh Tenenbaum"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AI GAMESTORE: Scalable, Open-Ended Evaluation of Machine General Intelligence with Human Games","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2602.17594v1","noteUrl":"/review?paper=review-neurips-2026-138879","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138880","title":"VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models","authors":["Qijia He","Xunmei Liu","Hammaad Memon","Ziang Li","Zixian Ma","Jaemin Cho","Jason Ren","Daniel Weld","Ranjay Krishna"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.24575","noteUrl":"/review?paper=review-neurips-2026-138880","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138881","title":"CapTrack: Multifaceted Evaluation of Forgetting in LLM Post-Training","authors":["Lukas Thede","Stefan Winzeck","Zeynep Akata","Jonathan Richard Schwarz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CapTrack: Multifaceted Evaluation of Forgetting in LLM Post-Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.06610","noteUrl":"/review?paper=review-neurips-2026-138881","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138882","title":"GitInject: Real-World Prompt Injection Attacks in AI-Powered CI/CD Pipelines","authors":["Jafar Isbarov","Umid Suleymanov","I Shumailov","Murat Kantarcioglu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"GitInject: Real-World Prompt Injection Attacks in AI-Powered CI/CD Pipelines","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.09935","noteUrl":"/review?paper=review-neurips-2026-138882","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138884","title":"Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR","authors":["Plawan Kumar Rath"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Soundness needs both scope and type invariants","plainExplanation":"Name membership does not imply type compatibility. The mask restricts SSA names, but its non-SSA branch does not constrain type annotations. The proof’s final identification of name scope with full validator acceptance therefore omits one required invariant.","primaryUrl":"https://arxiv.org/html/2607.18254","noteUrl":"/review?paper=review-neurips-2026-138884","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138887","title":"AVENUE: Audio-Video Editing Understanding and Evaluation","authors":["Hayeon Kim","Yoojin Jang","Jaejun Yoo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AVENUE: Audio-Video Editing Understanding and Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.04253","noteUrl":"/review?paper=review-neurips-2026-138887","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138890","title":"Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Models","authors":["Julian Skirzynski","Harry Cheon","Shreyas Kadekodi","Meredith Stewart","Berk Ustun"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Models","plainExplanation":"The formula counts the event that at least one concept prediction is wrong. That event is identical to a wrong AND label on a valid board. On an invalid board, a false positive requires every false concept to flip to true while every true concept stays true. Most concept errors therefore leave the invalid label unchanged.","primaryUrl":"https://arxiv.org/html/2606.04326","noteUrl":"/review?paper=review-neurips-2026-138890","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138891","title":"CoffeeBench: A Benchmark for Long-Horizon Strategic Decision-Making in Multi-Agent Economies","authors":["Issa Sugiura","Daichi Hattori","Kazuo Araragi","Keita Ogawa","Shota Onose","Taro Makino","Teppei Usuki","Takashi Ishida"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Zero attractiveness leaves market shares undefined","plainExplanation":"The shares are then \\(0/0\\).","primaryUrl":"https://arxiv.org/html/2606.16613","noteUrl":"/review?paper=review-neurips-2026-138891","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138892","title":"HelpBench: Assessing the Ability of LLMs to Provide Privacy, Safety, and Security Advice","authors":["Sarah Meiklejohn","Sunny Consolvo","Patrick G Kelley","Tara Matthews","Sai Teja Peddinti","Renee Shelby","Lenin Simicich","Kurt Thomas"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"HelpBench: Assessing the Ability of LLMs to Provide Privacy, Safety, and Security Advice","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.24819","noteUrl":"/review?paper=review-neurips-2026-138892","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138893","title":"MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image","authors":["Alan Arazi","Eilam Shapira","Shoham Grunblat","Mor Ventura","Elad Hoffer","Gioia Blayer","David Holzmüller","Lennart Purucker","Gael Varoquaux","Frank Hutter","Roi Reichart"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10616","noteUrl":"/review?paper=review-neurips-2026-138893","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138894","title":"Contextualized Evaluation of Vision Language Models through Dynamic Interviews","authors":["Yijiang Li","Huiqi Zou","Bingyang Wang","Ziang Xiao"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Contextualized Evaluation of Vision Language Models through Dynamic Interviews","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.14499","noteUrl":"/review?paper=review-neurips-2026-138894","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138896","title":"Welfare, Improvability, and Variance: A Principal-Agent Approach to Optimal Benchmark Item Aggregation","authors":["Andreas Haupt","Justin Hartenstein","Anka Reuel-Lamparth","Mykel J Kochenderfer","Sanmi Koyejo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Pooling within checkpoints does not give a checkpoint-average noise floor","plainExplanation":"The estimator pools within-checkpoint variances rather than averaging checkpoints first, so it does not acquire the claimed inverse-checkpoint variance factor. The printed reference and excess-noise interpretation are affected. The exact calculation concerns \\(Z^2\\), not an assertion that \\(\\mathbb E Z=1\\).","primaryUrl":"https://arxiv.org/html/2605.30916","noteUrl":"/review?paper=review-neurips-2026-138896","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138897","title":"ViMU: Benchmarking Video Metaphorical Understanding","authors":["Qi Li","Xinchao Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ViMU: Benchmarking Video Metaphorical Understanding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14607","noteUrl":"/review?paper=review-neurips-2026-138897","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138899","title":"MyoChallenge 2025: A New Benchmark for Human Athletic Intelligence","authors":["Cheryl Wang","Chun Kwang Tan","Balint Hodossy","Shirui Lyu","Jun Guo","Wentao Zhao","Huaping Liu","Chengkun Li","Merkourios Simos","Bianca Ziliotto","Alexander Mathis","Siyuan Liu","Jiahao Chen","Shanlin Zhong","Bo Jiang","Ci Song","Yaoye Zhu","Chenhui Zuo","Yanan Sui","Irfan Refai","Massimo Sartori","Guillaume Durandau","Vikash Kumar","Vittorio Caggiano"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MyoChallenge 2025: A New Benchmark for Human Athletic Intelligence","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15650","noteUrl":"/review?paper=review-neurips-2026-138899","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138900","title":"Build-and-Find: An Effort-Aware Protocol for Evaluating Agent-Managed Codebases","authors":["JHEN-KE LIN"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Build-and-Find: An Effort-Aware Protocol for Evaluating Agent-Managed Codebases","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06136","noteUrl":"/review?paper=review-neurips-2026-138900","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138901","title":"Rethinking XAI Evaluation: A Human-Centered Audit of Shapley Benchmarks in High-Stakes Settings","authors":["Inês Oliveira e Silva","Sérgio Jesus","Iker Perez","Rita P. Ribeiro","Carlos Soares","Hugo Ferreira","Pedro Bizarro"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The normalized entropy increment is undefined for equal endpoints","plainExplanation":"The admitted logistic model with all slopes zero predicts probability \\(0.5\\) at every deletion step.","primaryUrl":"https://arxiv.org/html/2604.22662","noteUrl":"/review?paper=review-neurips-2026-138901","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138902","title":"Re-evaluating Continual Learning with Few-Shot Adaptation","authors":["Amogh Inamdar","Matthew So","Vici I Milenia","Richard Zemel"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The SAUCE normalization can exceed one","plainExplanation":"The loss-offset subtraction omits the shot-interval length, so the normalized area can exceed one and change under a constant shift of all losses. The defect concerns the central metric formula and its range and offset invariance. It does not establish that the implementation used this expression or that reported rankings change.","primaryUrl":"https://arxiv.org/html/2606.03843","noteUrl":"/review?paper=review-neurips-2026-138902","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138903","title":"SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks","authors":["Gabriel Orlanski","Devjeet R Roy","Alexander Yun","Changho Shin","Alex Gu","Albert Ge","Dyah Adila","Nicholas Roberts","Frederic Sala","Aws Albarghouthi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.24755","noteUrl":"/review?paper=review-neurips-2026-138903","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138907","title":"DDBench: A Benchmark for Agentic Debugging on Distributed Systems","authors":["Yibo Yan","Huijuan Wang","Junzhou He","Yizhuo Liang","Shaoyu Wang","Huanchen Sun","Seo Jin Park"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DDBench: A Benchmark for Agentic Debugging on Distributed Systems","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.14863","noteUrl":"/review?paper=review-neurips-2026-138907","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138908","title":"A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts","authors":["Claudio César Claros-Olivares","Austin Brockmeier"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Global centering is incompatible with the reported within-class collapse metric","plainExplanation":"Set regret takes the best detector in a returned shortlist and compares it with the best detector in the full side inventory. Four NC-free singleton baselines are described as bounding every selector that ignores NC features. The NC-free selector returning the entire inventory has regret \\(\\min_{m\\in M_i}\\mathrm{AUGRC}_{im}-\\min_{m\\in M_i}\\mathrm{AUGRC}_{im}=0\\) on every case.","primaryUrl":"https://arxiv.org/html/2511.11934","noteUrl":"/review?paper=review-neurips-2026-138908","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138909","title":"VEX-Bench: Benchmarking Verification Complexity of LLM-Generated Misinformation","authors":["Hanxun Huang","Oscar W","Qizhou Wang","Silvia Montaña-Niño","Yige Li","Xiang Zheng","Elif B Doyuran","Phoebe Matich","Xiao Liu","Xingjun Ma","Sarah Erfani","Christopher Leckie"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"VEX-Bench: Benchmarking Verification Complexity of LLM-Generated Misinformation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35028","noteUrl":"/review?paper=review-neurips-2026-138909","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138912","title":"Benchmarking Sensor-Fault Robustness in Forecasting","authors":["Alexander Windmann","Philipp Wittenberg","Gianluca Manca","Marcel Dix","Jens U. Brandt","Oliver Niggemann"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Benchmarking Sensor-Fault Robustness in Forecasting","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10822","noteUrl":"/review?paper=review-neurips-2026-138912","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138913","title":"CausalSpatial: A Benchmark for Object-Centric Causal Spatial Reasoning","authors":["Wenxin (Wendy) Ma","Chenlong Wang","Ruisheng Yuan","Hao Chen","Nanru Dai","Chengxin Qian","Yijun Yang","Qi Chen","Zhaoyang Wang","S. Kevin Zhou","Jianwen Xie","Alan Yuille","Jieneng Chen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"CausalSpatial: A Benchmark for Object-Centric Causal Spatial Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.13304","noteUrl":"/review?paper=review-neurips-2026-138913","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138914","title":"Harbor Adapters and Harbor-Mix: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation","authors":["Lin Shi","Haowei Lin","Zixuan Zhu","Xiaoyue Zhou","Xiang Li","Xiangning Lin","Yaxuan Deng","Han Xu","Yuangang Li","Shanda Li","Zizhao Chen","Hanwen Xing","Harsh Raj","Bo Chen","Quan Shi","Steven Dillmann","Yipeng Gao","Puneesh Khanna","Ruofan Lu","Chao Zhou","Michael Yang","Robert Zhang","Siyuan Chai","Jiayu Chang","Yizhao Chen","Xiaokun Chen","Yiwei Dai","Wenting Yang","Hange Liu","Minghao Liu","Zihan Wang","Adnan E Assadi","Benedikt Stroebl","Estefany Kelly Buchanan","Han Meng","Junwei He","Longxuan YU","Radin Shayanfar","Yukyung Lee","Zhikang Dong","Allen G Hart","Anjiang Wei","Anurag Kashyap","Arpandeep Khatua","Audrey J Zheng","Chengrui Ma","David Heineman","dubing","Hai-Anh Trinh","Haishuo Fang","Hefan Zhang","Hui Shen","Issa Sugiura","Jiankai Sun","Jiechao Gao","Junhong Lin","Junnan Li","Kai Yang","Lei Hsiung","Maoyu Wang","Mengze Tang","Nabil Omi","Negin Raoof","Nicholas Edwards","Ruohao Guo","Orfeas Menis Mastromichalakis","Ryan Ji","Przemysław Hejman","Qi Qi","Qunshu Lin","Richard Zhuang","Rui Yang","Ruichen Zheng","Ryan Marten","Shaghayegh Fazliani","Shizheng Hou","Sicong Jiang","Sijie Li","Song Bian","Terry Yue Zhuo","Tianqing Wu","Tom Tang","Wanjia Zhao","Weihao Xuan","Wenhua Liang","Xian Liu","Xin Lan","Ruby Zhang","Xuandong Zhao","Yanchuan Tang","Yifan Jiang","Yijiang Li","Yitong Guan","Yizhi Li","YONGHUI LIU","Yuheng Tang","Yujun Mao","Yunfei Zhao","Yuxin Wang","Yuxuan Tang","Zhenheng Tang","Zhifei Li","Ziruo Wang","Ziyu She","Kaiyuan Liu","Iheb Chaabane","Yuxin Tang","Xiangyi Li","Soroush Vosoughi","Sanmi Koyejo","Di He","Etash Guha","Benjamin Feuer","Andy Konwinski","Boxuan Li","Leon L Chen","Alex Dimakis","Nicholas Carlini","Mike Merrill","Ludwig Schmidt","Alex Shaw"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Harbor Adapters and Harbor-Mix: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.04298","noteUrl":"/review?paper=review-neurips-2026-138914","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138915","title":"MEME: Multi-Entity & Evolving Memory Evaluation","authors":["Seokwon Jung","Alexander Rubinstein","Arnas Uselis","Sangdoo Yun","Seong Joon Oh"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MEME: Multi-Entity & Evolving Memory Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12477","noteUrl":"/review?paper=review-neurips-2026-138915","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138916","title":"PCEval: A Benchmark for Evaluating Physical Computing Capabilities of Large Language Models","authors":["Inpyo Song","Eunji Chon","Jangwon Lee"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"PCEval: A Benchmark for Evaluating Physical Computing Capabilities of Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.02404","noteUrl":"/review?paper=review-neurips-2026-138916","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138917","title":"WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors","authors":["Keming Wu","Yijing Cui","Wenhan Xue","Qijie Wang","Xuan Luo","ZhiYuan Feng","Zuhao Yang","Sudong Wang","Sicong Jiang","Haowei Zhu","Zihan Wang","Ping Nie","Wenhu Chen","Bin Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10434","noteUrl":"/review?paper=review-neurips-2026-138917","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138918","title":"BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments","authors":["Jiahao Zhang","Pengbin Feng","Chunlei Meng","Hang He","Tianyong Hao"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Shared and per-rater latent likelihoods differ","plainExplanation":"The shared-latent likelihood is \\(\\sum_k g_k\\prod_r C_r[k,y_r]\\), consistent with Equations (4) and (5) and Algorithm 3. The multi-rater likelihood definitions are inconsistent. The text does not determine which definition produced the reported table.","primaryUrl":"https://arxiv.org/html/2610.11116","noteUrl":"/review?paper=review-neurips-2026-138918","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138919","title":"BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces","authors":["Liangwei Yang","Jielin Qiu","Zixiang Chen","Ming Zhu","Juntao Tan","Zhiwei Liu","Wenting Zhao","Zhujun Lan","Akshara Prabhakar","Silvio Savarese","Huan Wang","Shelby Heinecke"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02798","noteUrl":"/review?paper=review-neurips-2026-138919","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138920","title":"PBT-Bench: Benchmarking AI Agents on Property-Based Testing","authors":["Guohao Jing","Xinqi Wang","Liao Zhang","Simon Du"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PBT-Bench: Benchmarking AI Agents on Property-Based Testing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15229","noteUrl":"/review?paper=review-neurips-2026-138920","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138921","title":"OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling","authors":["Adam Bawatneh","Sagar Sapkota","Amrit Singh Bedi","Santu Karmaker","Mubarak Shah"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26322","noteUrl":"/review?paper=review-neurips-2026-138921","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138923","title":"Auditing Cross-Lingual Fairness in Language Model Watermarking","authors":["Alexander Nemecek","Osama Zafar","Debargha Ganguly","Vikash Singh","Vipin Chaudhary","Erman Ayday"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Auditing Cross-Lingual Fairness in Language Model Watermarking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.20047","noteUrl":"/review?paper=review-neurips-2026-138923","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138927","title":"DPrivBench: Benchmarking LLMs' Reasoning for Differential Privacy","authors":["Erchi Wang","Pengrun Huang","Eli Chien","Om Thakkar","Kamalika Chaudhuri","Yu-Xiang Wang","Ruihan Wu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"DPrivBench: Benchmarking LLMs' Reasoning for Differential Privacy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.15851","noteUrl":"/review?paper=review-neurips-2026-138927","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138929","title":"Reporting Practice Matters: The Impact of Reference Choice on Chest X-ray Report Evaluation","authors":["Daniel P Jeong","Charles Q Li","Hossein Hosseiny","Nitya M Bhalla","Fatma U Morency","Pradeep Ravikumar","Zachary Lipton","Michael Oberst"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Reporting Practice Matters: The Impact of Reference Choice on Chest X-ray Report Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.19093","noteUrl":"/review?paper=review-neurips-2026-138929","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138930","title":"JobBench: Aligning Agent Work With Human Will","authors":["Yuetai Li","Yichen Feng","Zhangchen Xu","Zixian Ma","Kaiyuan Zheng","Fengqing Jiang","Xinghua Sun","Rulin Shao","Zichen Chen","Yue Huang","Xinyang Han","Brian Lee","Kayla Xu","Shenglai Zeng","Hang Hua","Xiangliang Zhang","Basel Alomair","Ranjay Krishna","Luke Zettlemoyer","Pang Wei Koh","Bhaskar Ramasubramanian","Luyao Niu","Xiang Yue","Radha Poovendran"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"JobBench: Aligning Agent Work With Human Will","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.26329","noteUrl":"/review?paper=review-neurips-2026-138930","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138931","title":"CellularS2-Bench: A Staged, Evidence-Grounded Benchmark for Cellular Network Security Reasoning","authors":["Ke Xie","Xingyi Zhao","Yiwen Hu","Haotian Lin","Shuhan Yuan","Wen Li","Tian Xie"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"CellularS2-Bench: A Staged, Evidence-Grounded Benchmark for Cellular Network Security Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.12716","noteUrl":"/review?paper=review-neurips-2026-138931","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138932","title":"Animation2Code: Evaluating Temporal Visual Reasoning in Video-to-Code Generation","authors":["Anya Ji","Abhijith Varma Mudunuri","David Chan","Alane Suhr"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Zero displacement makes the motion similarity undefined","plainExplanation":"Pairing a moving track with an absent track produces a zero cosine denominator.","primaryUrl":"https://arxiv.org/html/2606.28593","noteUrl":"/review?paper=review-neurips-2026-138932","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138933","title":"LPS-Bench: Benchmarking Safety Awareness of Computer-Use Agents in Long-Horizon Planning under Benign and Adversarial Scenarios","authors":["Tianyu Chen","Chujia Hu","Dongrui Liu","Xia Hu","Wenjie Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LPS-Bench: Benchmarking Safety Awareness of Computer-Use Agents in Long-Horizon Planning under Benign and Adversarial Scenarios","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.03255","noteUrl":"/review?paper=review-neurips-2026-138933","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138935","title":"Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces","authors":["Jiawei Chen","Ruoxi Xu","Boxi Cao","Ruotong Pan","Zhang yunfei","Yifei Hu","Yong Du","Tingting Gao","Han Li","Yaojie Lu","Yingfei Sun","Xianpei Han","Le Sun","Xiangyu Wu","Hongyu Lin"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.08362","noteUrl":"/review?paper=review-neurips-2026-138935","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138936","title":"Rich Insights from Cheap Signals: Efficient Evaluations via Tensor Factorization","authors":["Felipe Maia Polo","Aida Nematzadeh","Virginia Aglietti","Adam Fisch","Isabela Albuquerque"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An interval containing zero does not establish no performance loss","plainExplanation":"The replacement claim adds prompts with positive confidence lower bounds to prompts whose intervals contain zero, calling the union safe to substitute without performance loss. A zero-containing interval may also contain negative capability gaps.","primaryUrl":"https://arxiv.org/html/2603.02029","noteUrl":"/review?paper=review-neurips-2026-138936","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138939","title":"U-Bench: A Comprehensive Understanding of U-Net through 100-Variant Benchmarking","authors":["Fenghe Tang","Chengqi Dong","Wenxin (Wendy) Ma","Zikang Xu","Heqin Zhu","Zihang Jiang","Rongsheng Wang","Yuhao Wang","Chenxu Wu","Yingtai Li","S. Kevin Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"U-Bench: A Comprehensive Understanding of U-Net through 100-Variant Benchmarking","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.07041","noteUrl":"/review?paper=review-neurips-2026-138939","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138940","title":"LinuxArena: A Control Setting for AI Agents in Live Production Software Environments","authors":["Tyler Tracy","Ram Potham","Nikolas Kuhn","Myles Heller","Anshul Khandelwal","Cody Rushing","Henri Lemoine","Miguel Brandão","Tomáš Turlík","Adam Hanson","Josh Hills","Amy D Ngo","Ram Rachum","Nik Mitchell","Falko Galperin","Oscar Sykes","Pip Arnott","Samuel P Lima","Carlos R Giudice","Matt Goldwater","Daniel J Popp","Drew de Wet","Ruben Castaing","Qi Guo","Douw Marx","Benjamin Shaffrey","Justin Shenk","Martin Milbradt","Hannah Meagher","Shaheen Ahmed-Chowdhury","Daniel O'Connell","Christopher W Canal","Buck Shlegeris","Aryan Bhatt"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LinuxArena: A Control Setting for AI Agents in Live Production Software Environments","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.15384","noteUrl":"/review?paper=review-neurips-2026-138940","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138941","title":"Neptuna: A Comprehensive Machine Learning Framework for Benchmarking Complex Multiphase Flows","authors":["Harish Ramachandran","Björn Kimpel","Thomas Paula","Josef Winter","Steffen Schmidt","Nikolaus Adams"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Neptuna: A Comprehensive Machine Learning Framework for Benchmarking Complex Multiphase Flows","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.22280","noteUrl":"/review?paper=review-neurips-2026-138941","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138946","title":"SLAyiNG: A Diverse and Community-validated Dataset of Queer Slang","authors":["Leonor Veloso","Lea Hirlimann","Lucija M Zidar","Philipp Wicke","Valentin Hofmann","Hinrich Schuetze"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SLAyiNG: A Diverse and Community-validated Dataset of Queer Slang","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.17449","noteUrl":"/review?paper=review-neurips-2026-138946","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138948","title":"Agentick: A Unified Benchmark for General Sequential Decision-Making Agents","authors":["Roger Creus Castanyer","Pablo Samuel Castro","Glen Berseth"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Agentick: A Unified Benchmark for General Sequential Decision-Making Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06869","noteUrl":"/review?paper=review-neurips-2026-138948","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138950","title":"MindGuard: Guardrail Classifiers for Multi-Turn Mental Health Support","authors":["António Farinhas","Nuno M Guerreiro","José Pombal","Pedro Henrique Martins","Laura Melton","Alexandra Conway","Cara Dochat","Maya D&amp;#x27;Eon","Ricardo Rei"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MindGuard: Guardrail Classifiers for Multi-Turn Mental Health Support","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.00950","noteUrl":"/review?paper=review-neurips-2026-138950","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138951","title":"AgentLens: Revealing The Lucky Pass Problem in SWE-Agent Evaluation","authors":["Priyam Sahoo","Gaurav Mittal","Xiaomin Li","Shengjie Ma","Benjamin Steenhoek","Pingping Lin","Yu Hu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AgentLens: Revealing The Lucky Pass Problem in SWE-Agent Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12925","noteUrl":"/review?paper=review-neurips-2026-138951","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138952","title":"BioXArena: Benchmarking LLM Agents on Multi-Modal Biomedical Machine Learning Tasks","authors":["Loka Li","Duzhen Zhang","Xingbo Du","Leonard Song","Zixiao Wang","Assanali Aukenov","Noel Thomas","Shakhnazar Sailaukan","Yonghan Yang","Feilong Chen","Jiahua Dong","Kun Zhang","Bin Zhang","Le Song"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BioXArena: Benchmarking LLM Agents on Multi-Modal Biomedical Machine Learning Tasks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15766","noteUrl":"/review?paper=review-neurips-2026-138952","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138953","title":"TSNBench: Benchmarking LLM Proficiency in Time-Sensitive Networking","authors":["Rubi Debnath","Daniel Bujosa","Luxi Zhao","Silviu S. Craciunas","Paul Pop","Sebastian Steinhorst"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Signed send slope requires a subtraction in reserved bandwidth","plainExplanation":"Signed negative send slope enters the credit-balance denominator by subtraction. The printed plus denominator can imply reserved bandwidth above link capacity.","primaryUrl":"https://arxiv.org/html/2605.09481","noteUrl":"/review?paper=review-neurips-2026-138953","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138954","title":"Fine-Grained Benchmark Generation for Comprehensive Evaluation of Foundation Models","authors":["Mohammed Saidul Islam","Arash Afkanpour","Negin Baghbanzadeh","Farnaz Kohankhaki","Afshin Cheraghi","Ali Kore","Shayaan Mehdi","Elham Dolatabadi"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Fine-Grained Benchmark Generation for Comprehensive Evaluation of Foundation Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.18824","noteUrl":"/review?paper=review-neurips-2026-138954","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138955","title":"SWE Atlas: Benchmarking Coding Agents Beyond Issue Resolution","authors":["Mohit Raghavendra","Soham Dan","Miguel Romero","Yannis He","Johannes B Mols","Gautam Anand","Cole McCollum","Edgar Arakelyan","Vijay Bharadwaj","Andrew Park","Jeff Da","MohammadHossein Rezaei","Bing Liu","Brad Kenstler","Yunzhong He"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SWE Atlas: Benchmarking Coding Agents Beyond Issue Resolution","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08366","noteUrl":"/review?paper=review-neurips-2026-138955","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138957","title":"The Labeling Problem in Hallucination Detection Benchmarks: An Empirical Evaluation","authors":["Jorma Valjakka","Juhani Kivimäki","Juha Mylläri","Jukka K Nurminen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Labeling Problem in Hallucination Detection Benchmarks: An Empirical Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.08026","noteUrl":"/review?paper=review-neurips-2026-138957","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138958","title":"Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents","authors":["Wei-Chieh Huang","Weizhi Zhang","Yuchen Wu","Yankai Chen","Eric Hanchen Jiang","Wooseong Yang","Yiwei Yang","Henry P Zou","Hanrong Zhang","Ying Nian Wu","Haolun Wu","Kai-Wei Chang","Philip S Yu","Steve Liu","Aylin Caliskan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.15008","noteUrl":"/review?paper=review-neurips-2026-138958","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138960","title":"Benchmark Health Index: A Systematic Framework for Benchmarking the Benchmarks of LLMs","authors":["Longyuan Zhu","Hairan Hua","Linlin Miao","Wei Wang","必应 赵"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Capability weighting has the opposite marginal effect","plainExplanation":"For capability weights \\(\\theta_i>0\\), the derivative with respect to score \\(s_i\\) is \\(-\\theta_i/\\sum_j\\theta_j\\).","primaryUrl":"https://arxiv.org/html/2602.11674","noteUrl":"/review?paper=review-neurips-2026-138960","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138963","title":"TwinRouterBench: Fast Static and Live Dynamic Evaluation for Realistic Agentic LLM Routing","authors":["Pei Yang","Wanyi Chen","Tongyun Yang","Pengbin Feng","Jiarong Xing","Wentao Guo","Yuhang Yao","Yuhang Han","Hanchen Li","Xu Wang","Jie Xiao","Anjie Yang","Lynn Ai","Eric Yang","TIANYU SHI"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TwinRouterBench: Fast Static and Live Dynamic Evaluation for Realistic Agentic LLM Routing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18859","noteUrl":"/review?paper=review-neurips-2026-138963","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138965","title":"CompleteRXN: Toward Completing Open Chemical Reaction Databases","authors":["Gabriel Vogel","Minouk Noordsij","Evgeny A Pidko","Jana M. Weber"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"CompleteRXN: Toward Completing Open Chemical Reaction Databases","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.00222","noteUrl":"/review?paper=review-neurips-2026-138965","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138971","title":"ER-Reason: A Benchmark Dataset for LLM Clinical Reasoning in the Emergency Room","authors":["Nikita Mehandru","Niloufar Golchini","Namrata Garg","Kathy T LeSaint","Christopher J Nash","Anu Ramachandran","Travis Zack","Liam McCoy","Adam Rodman","David Bamman","Melanie Molina","Ahmed Alaa"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ER-Reason: A Benchmark Dataset for LLM Clinical Reasoning in the Emergency Room","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2505.22919","noteUrl":"/review?paper=review-neurips-2026-138971","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138973","title":"Verify0: Can AI Agents Build Formally Verified Software Repositories?","authors":["Zhe Ye","Hantao Lou","Yuechun Sun","Peiyang Song","Zhengxu Yan","Timothe Kasriel","Qingyang Zhang","Kaiyu Yang","Soonho Kong","Jingxuan He","Dawn Song"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Verify0: Can AI Agents Build Formally Verified Software Repositories?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.13522","noteUrl":"/review?paper=review-neurips-2026-138973","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138975","title":"MedKIT: Evaluating Knowledge Integration and Generalization in Large Language Models","authors":["Lukas Thede","Yash Kumar","David Chen","Danielle Bitterman","Matthias Bethge","Tom Hartvigsen","Zeynep Akata"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"MedKIT: Evaluating Knowledge Integration and Generalization in Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.38543","noteUrl":"/review?paper=review-neurips-2026-138975","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138979","title":"FakeParts: a New Family of AI-Generated Video Forgeries","authors":["Ziyi LIU","Firas Gabetni","Awais H SANI","Xi Wang","Soobash Daiboo","Gaëtan Brison","Gianni Franchi","Vicky Kalogeiton"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FakeParts: a New Family of AI-Generated Video Forgeries","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2508.21052v2","noteUrl":"/review?paper=review-neurips-2026-138979","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138980","title":"Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems","authors":["Mason Nakamura","Abhinav Kumar","Saswat Das","Sahar Abdelnabi","Saaduddin Mahmud","Ferdinando Fioretto","Shlomo Zilberstein","Eugene Bagdasarian"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The achievability display names the collusive range instead of the nominal range","plainExplanation":"Proposition C.1 bounds nominal degradation by the nominal objective’s range, but its achievability display replaces that range by the collusive objective’s range. The achievability equality names the collusive objective’s range instead of the nominal range bounded in part one; the following paragraph gives the valid intended reading.","primaryUrl":"https://arxiv.org/html/2602.15198","noteUrl":"/review?paper=review-neurips-2026-138980","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-138984","title":"RealICU: Do LLM Agents Understand Long-Context ICU Data? A Benchmark Beyond Behavior Imitation","authors":["Chengzhi Shen","Weixiang Shen","Tobias Susetzky","Chen Chen","Jun Li","Yuyuan Liu","zhang xuepeng","Zhenyu Gong","Daniel Rueckert","Jiazhen Pan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RealICU: Do LLM Agents Understand Long-Context ICU Data? A Benchmark Beyond Behavior Imitation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13542","noteUrl":"/review?paper=review-neurips-2026-138984","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138985","title":"PieArena: Ranking and Profiling Language Agents in Realistic Negotiation Scenarios","authors":["Chris Zhu","Sasha Cui","Will S Dufallo","Runzhi Jin","Zhen Xu","Linjun Zhang","Daylian M Cain"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Inverse log-likelihood Hessians give the wrong covariance sign","plainExplanation":"The analytic gradient is that of log likelihood \\(\\ell\\), while the covariance step inverts its Jacobian \\(H\\) at a maximum. At a regular maximum, \\(\\nabla^2\\ell\\) is negative definite, so \\(H^{-1}\\) is not a covariance matrix.","primaryUrl":"https://arxiv.org/html/2602.05302","noteUrl":"/review?paper=review-neurips-2026-138985","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-138987","title":"CurveBench: A Benchmark for Exact Topological Reasoning over Nested Jordan Curves","authors":["Amirreza Mohseni","Mona Mohammadi","Morteza Saghafian","Naser Talebizadeh Sardari"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"CurveBench: A Benchmark for Exact Topological Reasoning over Nested Jordan Curves","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14068","noteUrl":"/review?paper=review-neurips-2026-138987","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138990","title":"Towards Understanding and Measuring Cognitive Atrophy in LLM Behaviour","authors":["Abeer Badawi","Moyosoreoluwa Olatosi","Negin Baghbanzadeh","Laleh Seyyed-Kalantari","Frank Rudzicz","R. Shayna Rosenbaum","Sara Pishdadian","Elham Dolatabadi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Towards Understanding and Measuring Cognitive Atrophy in LLM Behaviour","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.18129","noteUrl":"/review?paper=review-neurips-2026-138990","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138991","title":"DataComp-VLM: Improved open datasets for Vision-Language Models","authors":["Matteo Farina","Vishaal Udandarao","Thao Nguyen","Selim Kuzucu","Maximilian Böther","Andreas Hochlehnert","Adhiraj Ghosh","Marianna Nezhurina","Karsten Roth","Joschka Strüber","Yuhui Zhang","Sebastian Dziadzio","Elaine Sui","Soumya S Jahagirdar","Dhruba Ghosh","Hasan Hammoud","Thomas De Min","Simone Caldarella","Muhammad Jehanzeb Mirza","Sedrick Keh","Mehdi Cherti","Hilde Kuehne","Bernt Schiele","Serena Yeung-Levy","Muhammad Ferjad Naeem","Federico Tombari","Ana Klimovic","Elisa Ricci","Matthias Bethge","Sewoong Oh","Ameya Prabhu","Alessio Tonioni","Jenia Jitsev","Massimiliano Mancini","Ludwig Schmidt","Nikhil Parthasarathy"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"DataComp-VLM: Improved open datasets for Vision-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.28551","noteUrl":"/review?paper=review-neurips-2026-138991","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138992","title":"MaSC: A Masked Similarity Metric for Evaluating Concept-Driven Generation","authors":["Patryk Bartkowiak","Lennart Petersen","Bartosz Kotrys","Dominik L Michels","Soeren Pirk","Wojtek Palubicki"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MaSC: A Masked Similarity Metric for Evaluating Concept-Driven Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22469","noteUrl":"/review?paper=review-neurips-2026-138992","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-138995","title":"SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?","authors":["Kean Shi","Zihang Li","Tianyi Ma","Zengji Tu","Jialong Wu","Xinbo Xu","Qingyao Yang","Weichu Xie","Ruoyu Wu","Ming Wu","Jason Zeng","Michael Heinrich","Liang Chen","Kuan Li","Baobao Chang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15777","noteUrl":"/review?paper=review-neurips-2026-138995","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-138997","title":"VSRo-200: A Romanian Visual Speech Recognition Dataset for Studying Supervision and Multimodal Robustness","authors":["Iulia-Maria Udrea","Alexandra Diaconu","Bogdan Alexe"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"VSRo-200: A Romanian Visual Speech Recognition Dataset for Studying Supervision and Multimodal Robustness","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.08112","noteUrl":"/review?paper=review-neurips-2026-138997","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139003","title":"HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing","authors":["Zhenjie Yang","Xingyu Jiao","Guopeng Zhong","Shuzhe Yang","Shi Che","Chao Wu","Chenyu Jiang","Dongjie Zhang","Yideng Zhang","Zheng Zhang","Muyun Jiang","Haisheng Su","Shuang Jin","Donghang Zhang","Chao Yang","Li Chen","Hongyang Li","Zuxuan Wu","Junchi Yan","Xiaosong Jia","Yu-Gang Jiang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.12122","noteUrl":"/review?paper=review-neurips-2026-139003","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139006","title":"DiscoverPhysics: Benchmarking LLMs for out-of-the-box scientific thinking","authors":["Lindsay Smith","Matt Sampson","Siddharth Mishra-Sharma","Peter Melchior","Andrew Wilson","Pavel Izmailov","Carolina Cuesta Lazaro"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DiscoverPhysics: Benchmarking LLMs for out-of-the-box scientific thinking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26087","noteUrl":"/review?paper=review-neurips-2026-139006","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139007","title":"HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning","authors":["Taowen Wang","Zikang Xie","Bin Yang","Yunheng Wang","Zizhao Yuan","Yuetong Fang","Yixiao Feng","Yichi Wang","Xingyu Chen","Haodong Chen","Qiwei Wu","Weisheng Xu","Lihan Chen","Lusong Li","zeng zecui","Renjing Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.17833","noteUrl":"/review?paper=review-neurips-2026-139007","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139008","title":"VI-Bench: Benchmarking Prompt Inversion from AIGC Videos","authors":["Wulin Xie","Rui Zhao","Kecen Li","Xiujin Liu","Bokang Zhang","Zheng Liu","Xinwen Hou","Chen GONG"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"VI-Bench: Benchmarking Prompt Inversion from AIGC Videos","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.08079","noteUrl":"/review?paper=review-neurips-2026-139008","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139011","title":"DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents","authors":["Yun-Shiuan Chuang","Ruixuan Tu","Chengtao Dai","You Li","Binwei Yao","Michael Tessler","Sijia Yang","Dhavan Shah","Robert Hawkins","Junjie Hu","Timothy T Rogers"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.25110","noteUrl":"/review?paper=review-neurips-2026-139011","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139012","title":"Get a GRIP, this will be a long TRIP: A Quantifiable Long-Range Framework for Verifying Over-squashing","authors":["Ferran Hernandez Caralt","Simon Heilig","Adrián Bazaga","Asja Fischer","Moshe Eliasof","Pietro Lió"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Get a GRIP, this will be a long TRIP: A Quantifiable Long-Range Framework for Verifying Over-squashing","plainExplanation":"For squared loss, the best predictor is the conditional mean. Replacing unseen inputs by their modes can give a larger error, so the proposed oracle does not certify optimal use of information at a given radius. Combining stable aggregates from several ranges can also reveal shorter-range information, contrary to the universal strict-range claim. The remaining mathematical corrections specify nonempty shells, nondegenerate target laws, the appropriate entropy convention, and independence for product stability. A separate result identifies the positive error floor imposed by the default bounded regression output; it is not evidence about the historical plotted models.","primaryUrl":"https://arxiv.org/html/2610.03556","noteUrl":"/review?paper=review-neurips-2026-139012","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference","mathematical_error","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-139013","title":"X-Palm: Paired Multispectral-to-Smartphone Dataset for Cross-Domain Palmprint Authentication","authors":["Seyed Jamal Seyedmohammadi","Pai Chet Ng","Angelo Genovese","Zhixiang Chi","Jeannie Lee","Konstantinos N Plataniotis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"X-Palm: Paired Multispectral-to-Smartphone Dataset for Cross-Domain Palmprint Authentication","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.08437","noteUrl":"/review?paper=review-neurips-2026-139013","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139014","title":"Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models","authors":["Nilaksh","Saurav Jha","Artem Zholus","Sarath Chandar"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Reconstruction or Semantics? What Makes a Latent Space Useful for Robotic World Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06388","noteUrl":"/review?paper=review-neurips-2026-139014","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139016","title":"SceneFactory: GPU-Accelerated Multi-Agent Driving Simulation with Physics-Based Vehicle Dynamics","authors":["Yicheng Zhu","Yang Chen","Tao Li","Zilin Bian"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Minimum friction combination clips coefficients above one","plainExplanation":"The specified effective friction is \\(\\min(\\mu,1)\\), equal to \\(\\mu\\) only for \\(\\mu\\le1\\).","primaryUrl":"https://arxiv.org/html/2605.08528","noteUrl":"/review?paper=review-neurips-2026-139016","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139017","title":"Precise Debugging Benchmark: Is Your Model Debugging or Regenerating?","authors":["Miaosen Chai","Wang Bill Zhu","Shangshang Wang","Yejia Liu","Song Bian","Honghua Dong","Willie Neiswanger","Robin Jia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Contiguous sub-edit search does not find a minimum arbitrary subset","plainExplanation":"Equation (3) defines essential edit size using the smallest successful subset, but Algorithm 2 searches only contiguous sub-edits. A three-edit block changes \\((a,z,c)=(0,0,0)\\) to \\((1,7,1)\\) for test \\(a+c=2\\).","primaryUrl":"https://arxiv.org/html/2604.17338","noteUrl":"/review?paper=review-neurips-2026-139017","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139018","title":"Bridging Academia and Industry: A Comprehensive Benchmark for Attributed Graph Clustering","authors":["Yunhui Liu","Pengyu Qiu","Yu Xing","Peng Du","Yongchao Liu","Chuntao Hong","Jiajun Zheng","Tao Zheng","Tieke He"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Bridging Academia and Industry: A Comprehensive Benchmark for Attributed Graph Clustering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.08519","noteUrl":"/review?paper=review-neurips-2026-139018","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139020","title":"CTE-Bench: Audited Counterfactual Trace Evaluation for Stateful Software Simulators","authors":["Xinran Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CTE-Bench: Audited Counterfactual Trace Evaluation for Stateful Software Simulators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36647","noteUrl":"/review?paper=review-neurips-2026-139020","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139021","title":"Towards Error-Free EHRs: Reasoning-Intensive Consistency Verification Between Clinical Notes and Structured Tables in Electronic Health Records","authors":["Yeonsu Kwon","Jiho Kim","Junseong Choi","Paloma Rabaey","Minseo Kim","Sujeong Im","Jeewon Yang","Jun-Min Lee","Sangji Lee","Jiwon Kim","Hangyul Yoon","Hyunwook Kwon","Edward Choi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Towards Error-Free EHRs: Reasoning-Intensive Consistency Verification Between Clinical Notes and Structured Tables in Electronic Health Records","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.26463","noteUrl":"/review?paper=review-neurips-2026-139021","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139024","title":"Redefining Instance Matching: A Unified Framework for Part-Aware Matching in Panoptic Segmentation Evaluation","authors":["Erik Großkopf","Soumya Snigdha Kundu","Hendrik Möller","Nicolas Münster","Mehdi Astaraki","Paula T Buzduga","Kerstin Ritter","Benedikt Wiestler","Jan Kirschke","Jonathan Shapey","Tom Vercauteren","Florian Kofler"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"One-to-Many matching can also separate maximum weight from maximum PQ","plainExplanation":"Appendix B says only the Many-to-One strategy can have a max-weight/max-PQ discrepancy, despite later acknowledging low-overlap discrepancies for One-to-Many. One-to-Many prediction usage couples the PQ denominator, so independently maximizing edge weight can also fail to maximize PQ, contrary to the exclusive Many-to-One wording. This is a cosmetic explanatory overstatement already partly acknowledged in the next paragraph. The high-threshold equivalence and the operational surrogate remain unaffected.","primaryUrl":"https://arxiv.org/html/2605.31094","noteUrl":"/review?paper=review-neurips-2026-139024","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-139025","title":"TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning","authors":["Ruochi Li","peter lin","Haoxuan Zhang","Haihua Chen","Junhua Ding","Edward Gehringer","Yang Zhang"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.09360","noteUrl":"/review?paper=review-neurips-2026-139025","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139026","title":"SFT-then-RL Outperforms Mixed-Policy Methods for LLM Reasoning","authors":["Alexis Limozin","Eduard F Ďurech","Torsten Hoefler","Imanol Schlag","Valentina Pyatkin"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Rank normalization alone does not normalize an accumulation window","plainExplanation":"Normalizing across data-parallel ranks leaves unequal accumulation microbatches incorrectly weighted unless the denominator covers the whole window.","primaryUrl":"https://arxiv.org/html/2604.23747","noteUrl":"/review?paper=review-neurips-2026-139026","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139027","title":"Croissant Baker: Metadata Generation for Discoverable, Governable, and Reusable ML Datasets","authors":["Rafi Al Attrach","Rajna Fani","Sebastian Lobentanzer","Joan Giner-Miguelez","DEBANSHU DAS","Varuni H K","Nobin Sarwar","Rajat Ghosh","Anwai Archit","Surbhi Motghare","Christina C Parry","Luis Oala","Lara Grosso","Joaquin Vanschoren","Steffen Vogler","Sujata Goswami","Eric Rosenthal","Marzyeh Ghassemi","Matthew McDermott","Tom Pollard"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Croissant Baker: Metadata Generation for Discoverable, Governable, and Reusable ML Datasets","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15079","noteUrl":"/review?paper=review-neurips-2026-139027","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139028","title":"SWE-Chain: Benchmarking Coding Agents on Chained Release-Level Package Upgrades","authors":["Man Ho Lam","Chaozheng Wang","Hange Liu","Jingyu Xiao","Haau-Sing Li","Jen-Tse Huang","Terry Yue Zhuo","Michael R Lyu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SWE-Chain: Benchmarking Coding Agents on Chained Release-Level Package Upgrades","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14415","noteUrl":"/review?paper=review-neurips-2026-139028","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139030","title":"JMed48k: A Multi-Profession Japanese Medical Licensing Benchmark for Vision-Language Model Evaluation","authors":["Yue Xun","Junyu Liu","Qian Niu","Xinyi Wang","Zheng Yuan","Zirui Li","Zequn Zhang","Bowen Zhao","Emma, Shujun Wang","Zihui Li","Kan Hatakeyama-Sato","Yusuke Iwasawa","Yutaka Matsuo"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Above random guessing does not certify image use","plainExplanation":"The positive-control interpretation treats \\(1/k\\) chance accuracy as a distribution-free upper bound for image-blind predictors on image-as-options questions. Uniform independent guessing has expectation \\(1/k\\), but a constant-label image-blind predictor scores the prevalence of that label.","primaryUrl":"https://arxiv.org/html/2605.22080","noteUrl":"/review?paper=review-neurips-2026-139030","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139032","title":"OmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered Scenes","authors":["Naiyu Fang","Zhongjin Luo","Yuxin Mo","Siyuan Huang","Jianbo Liu","Yufei Liu","Zheyuan Zhou","Chengkai Jin","Xiaogang Wang","Hongsheng Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered Scenes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.11194","noteUrl":"/review?paper=review-neurips-2026-139032","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139033","title":"ORBIT: A Framework for Multi-Agent Security Evaluations","authors":["Ben Hagag","William L Anderson","Srija Chakraborty","Christian Schroeder de Witt"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ORBIT: A Framework for Multi-Agent Security Evaluations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33102","noteUrl":"/review?paper=review-neurips-2026-139033","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139035","title":"Who&When Pro: Can LLMs Really Attribute Failures in AI Agents?","authors":["Jiale Liu","Huajun Xi","Shaokun Zhang","Yifan Zeng","Tianwei Yue","Chi Wang","Jian Kang","Huazheng Wang","Qingyun Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Who&When Pro: Can LLMs Really Attribute Failures in AI Agents?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.09996","noteUrl":"/review?paper=review-neurips-2026-139035","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139037","title":"GraphInstruct: A Progressive Benchmark for Diagnosing Capability Gaps in LLM Graph Generation","authors":["Zihe Wei","Sheng Xiang","Ying Zhang","changjun jiang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"GraphInstruct: A Progressive Benchmark for Diagnosing Capability Gaps in LLM Graph Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.09997","noteUrl":"/review?paper=review-neurips-2026-139037","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139039","title":"ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins","authors":["Yichen Zhou","Jonathan L Golob","Amir Mohammad Karimi Mamaghan","Stefan Bauer","Patrick Schwab"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.06740","noteUrl":"/review?paper=review-neurips-2026-139039","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139043","title":"Understanding the Interplay between Memorization and Learning in Large Language Models","authors":["Bishwamittra Ghosh","Soumi Das","Qinyuan Wu","Mohammad Aflah Khan","Krishna Gummadi","Evimaria Terzi","Deepak Garg"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A reference model does not identify the target’s leave-one-string counterfactual","plainExplanation":"The proxy treats a disjoint-data reference model matching or exceeding target recollection as certifying absence of target contextual memorization. Contextual memorization compares the target with its own counterfactual training without the string.","primaryUrl":"https://arxiv.org/html/2507.14777","noteUrl":"/review?paper=review-neurips-2026-139043","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139044","title":"Benchmarking Multimodal Mathematical Reasoning with Explicit Visual Dependency","authors":["Zhikai Wang","Jiashuo Sun","Wenqi Zhang","Zhiqiang Hu","Xin Li","Deli Zhao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Benchmarking Multimodal Mathematical Reasoning with Explicit Visual Dependency","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2504.18589","noteUrl":"/review?paper=review-neurips-2026-139044","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139046","title":"LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary Goal Navigation","authors":["Bo Miao","Weijia Liu","Jun Luo","Lachlan Shinnick","Jian Liu","Thomas Hamilton-Smith","Yuhe Yang","Zijie Wu","Vanja Videnovic","Feras Dayoub","Anton van den Hengel"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary Goal Navigation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.02220","noteUrl":"/review?paper=review-neurips-2026-139046","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139047","title":"A Benchmark for Omni-Modal Reasoning in Long Videos","authors":["Mohammed Irfan Kurpath","Jaseel M Kaithakkodan","Jinxing Zhou","Sahal Shaji Mullappilly","Mohammad Almansoori","Noor Ahsan","Beknur Kalmakhanbet","sambal shikhar","Rishabh Lalla","Jean Lahoud","Mariette Awad","Fahad Shahbaz Khan","Salman Khan","Rao Anwer","Hisham Cholakkal"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A Benchmark for Omni-Modal Reasoning in Long Videos","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.16978","noteUrl":"/review?paper=review-neurips-2026-139047","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139049","title":"IPIBench: Evaluating Interactive Proactive Intelligence of MLLMs under Continuous Streams","authors":["Jinzhao Li","Yinuo Chen","Wenxuan Song","Yijia Lei","Yichi Zhang","Honglei Yan","Panwang Pan","Miao Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"IPIBench: Evaluating Interactive Proactive Intelligence of MLLMs under Continuous Streams","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27074","noteUrl":"/review?paper=review-neurips-2026-139049","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139051","title":"WorldMemArena: Evaluating Multimodal Agent Memory Through Action–World Interaction","authors":["Chengzhi Liu","Yuzhe YANG","Sophia Xiao Pu","Yepeng Liu","Lin Long","Yichen Guo","Nuo Chen","Zhaotian Weng","Yiheng Zhong","Angxiao Zong","Elena Kochkina","Simerjot Kaur","Charese Smiley","Xiaomo Liu","James Zou","Sheng Liu","Yuheng Bu","Songyou Peng","Xin Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"WorldMemArena: Evaluating Multimodal Agent Memory Through Action–World Interaction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.29341","noteUrl":"/review?paper=review-neurips-2026-139051","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139052","title":"Cracks in the Foundation: A Civil Infrastructure Dataset to Challenge Vision Foundation Models","authors":["Nicola Farronato","Niccolò Avogaro","Thomas Frick","Mattia Rigotti","Rizwan U Khan","Michele Magno","Konrad Schindler","A. Cristiano I. Malossi","Florian Scheidegger"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Cracks in the Foundation: A Civil Infrastructure Dataset to Challenge Vision Foundation Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.18413","noteUrl":"/review?paper=review-neurips-2026-139052","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139053","title":"Apple-$\\pi$: Benchmarking Thinking with Video Towards Law-Grounded Physical Intelligence","authors":["Runmao Yao","Kairui Hu","Yukang Cao","Ruisi Wang","Shulin Tian","Ziang Cao","Weichen Fan","Ziqi Huang","Yuhao Dong","Hao Li","Zhaoxi Chen","Zhongang Cai","Lei Yang","Ziwei Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Apple-$\\pi$: Benchmarking Thinking with Video Towards Law-Grounded Physical Intelligence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.16401","noteUrl":"/review?paper=review-neurips-2026-139053","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139055","title":"IndustryCode: A Benchmark for Industry Code Generation","authors":["Puyu Zeng","Zhaoxi Wang","Zhixu Duan","Liang Feng","Shaobo Wang","Cunxiang Wang","jinghang Wang","必应 赵","HU WEI","Qibing Ren","Linfeng Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"IndustryCode: A Benchmark for Industry Code Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.02729","noteUrl":"/review?paper=review-neurips-2026-139055","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139057","title":"MAEB: Massive Audio Embedding Benchmark","authors":["Adnan E Assadi","Isaac Chung","Chenghao Xiao","Roman Solomatin","Animesh Jha","Rahul Chand","Silky Singh","Kaitlyn Wang","Ali Sartaz Khan","Marc M Nasser","Sufen Fong","Pengfei He","Alan Xiao","Ayush Sunil Munot","Aditya Shrivastava","Artem Gazizov","Niklas Muennighoff","Kenneth Enevoldsen"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"MAEB: Massive Audio Embedding Benchmark","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.16008","noteUrl":"/review?paper=review-neurips-2026-139057","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139059","title":"Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?","authors":["Anton Korznikov","Andrey V Galichin","Alexey Dontsov","Oleg Rogov","Ivan Oseledets","Elena Tutubalina"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.14111","noteUrl":"/review?paper=review-neurips-2026-139059","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139060","title":"UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks","authors":["Zhekai Chen","CHENGQI DUAN","Kaiyue Sun","Bohao Li","Yuqing Wang","Manyuan Zhang","Xihui Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.08768","noteUrl":"/review?paper=review-neurips-2026-139060","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139061","title":"No Triangulation Without Representation: Generalization in Topological Deep Learning","authors":["Johannes S. Schmidt","Martin Carrasco","Ernst Röell","Guy Wolf","Nello Blaser","Bastian Rieck"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Surface classification needs connectedness","plainExplanation":"The sphere \\(S^2\\) and disjoint union \\(S^2\\sqcup T^2\\) are orientable with Euler characteristic two, but have different numbers of connected components and cannot be homeomorphic. The omitted connectedness hypothesis is the defect; connected-surface applications are unaffected.","primaryUrl":"https://arxiv.org/html/2605.06467","noteUrl":"/review?paper=review-neurips-2026-139061","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139062","title":"XL-DocBench: Benchmarking Evidence-Grounded Extra-Long Document Understanding","authors":["Hongchen Wei","Yuanzhe Wang","Bei Liu","Yifan Yang","Qi Dai","Ruichun Ma","Kai Qiu","Yunsheng Li","Dongdong Chen","Chong Luo","Zhenzhong Chen","Baining Guo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"XL-DocBench: Benchmarking Evidence-Grounded Extra-Long Document Understanding","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.00036","noteUrl":"/review?paper=review-neurips-2026-139062","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139063","title":"RxnOptBench: Benchmarking LLMs for Reaction-Condition Optimization in Organic Methodology","authors":["Lingli Ge","Junyuan Gao","Jiahe Song","Jiaxing Sun","Yubin Wang","Boyu Zhu","Haote Yang","Jingchao Wang","Lixin Ma","Jiang Wu","Yuqiang Li","Conghui He"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"RxnOptBench: Benchmarking LLMs for Reaction-Condition Optimization in Organic Methodology","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02242","noteUrl":"/review?paper=review-neurips-2026-139063","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139065","title":"TabPrep: Closing the Feature Engineering Gap in Tabular Benchmarks","authors":["Andrej Tschalzev","Nick Erickson","Yuyang (Bernie) Wang","Huzefa Rangwala","Stefan Lüdtke","Christian Bartelt","Heiner Stuckenschmidt"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TabPrep: Closing the Feature Engineering Gap in Tabular Benchmarks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.02384","noteUrl":"/review?paper=review-neurips-2026-139065","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139066","title":"TROPT: An Open Framework for Unifying and Advancing Discrete Text Optimization","authors":["Matan Ben-Tov","Mahmood Sharif"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"TROPT: An Open Framework for Unifying and Advancing Discrete Text Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.23496","noteUrl":"/review?paper=review-neurips-2026-139066","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139067","title":"PerturbReason: A Knowledge-Grounded Benchmark and Framework for Cell-State–Conditioned Mechanistic Reasoning of Perturbation Effects","authors":["Dongkwan Kim","Yiming Gao","Yining Yang","Yang Shen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"PerturbReason: A Knowledge-Grounded Benchmark and Framework for Cell-State–Conditioned Mechanistic Reasoning of Perturbation Effects","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.18777","noteUrl":"/review?paper=review-neurips-2026-139067","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139070","title":"ProactBench: Beyond What The User Asked For","authors":["Sepehr Harfi Moridani","Ahmad Salimi","Dongming Shen","Alexander Smola"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ProactBench: Beyond What The User Asked For","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.09228","noteUrl":"/review?paper=review-neurips-2026-139070","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139071","title":"TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots","authors":["Fangrui Huang","Souhad Chbeir","Arpandeep Khatua","Sheng Wang","Sijun Tan","Kenan Ye","Lillian A Bailey","Merryn Daniel","Ryan Louie","Sanmi Koyejo","Ehsan Adeli"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.18008","noteUrl":"/review?paper=review-neurips-2026-139071","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139072","title":"SynthForensics: Benchmarking and Evaluating People-Centric Synthetic Video Deepfakes","authors":["Roberto Leotta","Salvatore A Sambataro","Claudio V Ragaglia","Mirko Casu","Yuri Petralia","Francesco Guarnera","Luca Guarnera","Sebastiano Battiato"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SynthForensics: Benchmarking and Evaluating People-Centric Synthetic Video Deepfakes","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.04939","noteUrl":"/review?paper=review-neurips-2026-139072","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139073","title":"CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets","authors":["Berke Arda","Mubashara Akhtar","Ahmetcan Yavuz","Paul Gerry","Sebastian Lobentanzer","Nobin Sarwar","Joan Giner-Miguelez","Kongtao Chen","Luyao Zhang","Mrinmaya Sachan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.07132","noteUrl":"/review?paper=review-neurips-2026-139073","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139074","title":"Large Language Model Selection with Limited Annotations","authors":["Yavuz Durmazkeser","Patrik Okanovic","Andreas Kirsch","Torsten Hoefler","Nezihe Merve Gürel"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Two derivation steps need qualification","plainExplanation":"A common normalizer does not follow from a general similarity function. Normalization requires \\(Z(y,\\tau)=\\sum_r\\exp(s(r,y)/\\tau)\\), which can depend on the output \\(y\\). Omitting that dependence changes the Bayesian posterior.","primaryUrl":"https://arxiv.org/abs/2510.09418v2","noteUrl":"/review?paper=review-neurips-2026-139074","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139075","title":"How Does Personalized Memory Shape LLM Behavior? Benchmarking Rational Preference Utilization in Personalized Assistants","authors":["Xueyang Feng","Weinan Gan","Xu Chen","Quanyu Dai","Yong Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"How Does Personalized Memory Shape LLM Behavior? Benchmarking Rational Preference Utilization in Personalized Assistants","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.16621","noteUrl":"/review?paper=review-neurips-2026-139075","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139077","title":"Can AI Agents Synthesize Scientific Conclusions?","authors":["Hayoung Jung","Pedro V Diniz","José R Roveda","Abner F da Silva","Haeun Jung","Enoch Tsai","Aleksandra Korolova","Manoel Ribeiro"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Can AI Agents Synthesize Scientific Conclusions?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11337","noteUrl":"/review?paper=review-neurips-2026-139077","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139079","title":"Are Agents Ready to Teach? A Multi-Stage Benchmark for Real-World Teaching Workflows","authors":["Zixin (Steven) CHEN","Peng Liu","Rui SHENG","Haobo Li","Jianhong Tu","Deng Xiaodong","KaShun SHUM","Dayiheng Liu","Huamin Qu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Are Agents Ready to Teach? A Multi-Stage Benchmark for Real-World Teaching Workflows","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.14322","noteUrl":"/review?paper=review-neurips-2026-139079","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139081","title":"CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage","authors":["Liu Jiale","Jungang Li","Jieming Yu","Yu Xinglin","Zihao Dongfang","Shi Keyu","Zongjian Ding","家焕 张","Shunwen Bai","Haoran Huang","Yurun Wang","Yanxi Wu","Ningzhe Yu","Yudong Gao","Mingjun Cheng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15597","noteUrl":"/review?paper=review-neurips-2026-139081","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139084","title":"Safe, or Simply Incapable? Rethinking Safety Evaluation for Phone-Use Agents","authors":["Zhengyang Tang","Yi Zhang","Chenxin Li","Xin Lai","Pengyuan Lyu","Yiduo Guo","Weinong Wang","Junyi Li","Yang Ding","Huawen Shen","Zhengyao Fang","Xingran Zhou","Liang Wu","Fei Tang","Sunqi Fan","Shangpin Peng","Zheng Ruan","anran zhang","Benyou Wang","Chengquan Zhang","Han Hu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Safe, or Simply Incapable? Rethinking Safety Evaluation for Phone-Use Agents","plainExplanation":"No issue was found in the checked claims. The v1 coefficient is inconsistent with its displayed ranks. V2 instead reports Pearson correlation \\(r=0.239\\) between general success rate and conditional safety on Clean-684. The population, safety variable and correlation statistic have all changed, so the v1 calculation does not evaluate the v2 claim. The exact historical discrepancy has been removed from the current version.","primaryUrl":"https://arxiv.org/html/2605.07630","noteUrl":"/review?paper=review-neurips-2026-139084","collection":"neurips-2026","latestUploadedAt":"2026-05-08T11:58:57Z","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139085","title":"SchemeArena: Factorized Stress Testing of Scheming in LLM Agents","authors":["Jie Ruan","Inderjeet Nair","Amy Liu","Muhammad Khalifa","Yusheng Zhou","Lu Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SchemeArena: Factorized Stress Testing of Scheming in LLM Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.08126","noteUrl":"/review?paper=review-neurips-2026-139085","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139087","title":"Sparkle: Realizing Lively Instruction-Guided Video Background Replacement via Decoupled Guidance","authors":["Ziyun Zeng","Yiqi Lin","Guoqiang Liang","Mike Zheng Shou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Sparkle: Realizing Lively Instruction-Guided Video Background Replacement via Decoupled Guidance","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06535","noteUrl":"/review?paper=review-neurips-2026-139087","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139088","title":"MemLeak: Diagnosing Information Leaks in Multimodal Agent Memory","authors":["Kuan Wang","Chao Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MemLeak: Diagnosing Information Leaks in Multimodal Agent Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.29788","noteUrl":"/review?paper=review-neurips-2026-139088","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139090","title":"ProgramBench: Can Language Models Rebuild Programs From Scratch?","authors":["John Yang","Kilian Lieret","Jeffrey Ma","Parth Thakkar","Dmitrii Pedchenko","Sten Sootla","Emily McMilin","Pengcheng Yin","Rui Hou","Gabriel Synnaeve","Diyi Yang","Ofir Press"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ProgramBench: Can Language Models Rebuild Programs From Scratch?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.03546","noteUrl":"/review?paper=review-neurips-2026-139090","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139091","title":"LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM Compression","authors":["Zishan Shao","Lixun Zhang","Kangning Cui","Wenhao Wu","Jinhee Kim","Yixiao Wang","Ting Jiang","Hancheng Ye","Qinsi Wang","Fan Yang","Danyang Zhuo","Yiran Chen","Hai Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM Compression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.26389","noteUrl":"/review?paper=review-neurips-2026-139091","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139094","title":"BankerToolBench: Evaluating AI Agents in End-to-End Investment Banking Workflows","authors":["Elaine Lau","Markus Dücker","Ronak Chaudhary","Hui Wen Goh","Rosemary Wei","Vaibhav Kumar","Saed Qunbar","Guram Gogia","Yi Liu","Scott Millslagle","Nasim Borazjanizadeh","Ulyana Tkachenko","Samuel E Danquah","Collin Schweiker","Vijay Karumathil","Asrith Devalaraju","Andrew P Martin","Varsha Sandadi","Haemi Nam","Punit Arani","Ray Epps","Abdullah Arif","Sahil Bhaiwala","Curtis Northcutt","Skyler Wang","Anish Athalye","Jonas Mueller","Francisco Guzmán"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BankerToolBench: Evaluating AI Agents in End-to-End Investment Banking Workflows","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.11304","noteUrl":"/review?paper=review-neurips-2026-139094","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139096","title":"FrontierOR: Benchmarking LLMs' Capacity for Efficient Algorithm Design in Large-Scale Optimization","authors":["Minwei Kong","Chonghe Jiang","Ao Qu","Wenbin Ouyang","Zhaoming Zeng","Xiaotong Guo","Zhekai Li","Junyi Li","Yi Fan","Xinshou Zheng","Xi Jing","Yikai Zhang","Zhiwei Liang","Seonghoo Kim","Runqing Yang","Sirui Li","Han Zheng","Wangyang Ying","Ou Zheng","Chonghuan Wang","Jinglong Zhao","Paul Liang","Jinhua Zhao","Hai Wang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Max-absolute normalization permits a gap of magnitude two","plainExplanation":"The signed objective gap divides the difference of two objectives by their maximum absolute value and claims range \\([-1,1]\\). Opposite objectives \\(a\\) and \\(-a\\), with \\(a\\ne0\\), give a normalized difference of magnitude \\(2\\).","primaryUrl":"https://arxiv.org/html/2605.25246","noteUrl":"/review?paper=review-neurips-2026-139096","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139102","title":"ClinMAS: A Knowledge-Grounded Multi-Agent Simulation Framework for Evaluating Clinical Reasoning in LLMs","authors":["Yuqi Tang","Jing Yu","Zichang Su","Kehua Feng","Zhihui Zhu","Libin Wang","Lei Liang","Qiang Zhang","Keyan Ding","Huajun Chen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ClinMAS: A Knowledge-Grounded Multi-Agent Simulation Framework for Evaluating Clinical Reasoning in LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.23440v1","noteUrl":"/review?paper=review-neurips-2026-139102","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139104","title":"No More, No Less: Task Alignment in Terminal Agents","authors":["Sina Mavali","David Pape","Jonathan Evertz","Samira Abedini","Devansh Srivastav","Thorsten Eisenhofer","Sahar Abdelnabi","Lea Schönherr"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"No More, No Less: Task Alignment in Terminal Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12233","noteUrl":"/review?paper=review-neurips-2026-139104","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139106","title":"Measuring and Mitigating the Distributional Gap Between Real and Simulated User Behaviors","authors":["Shuhaib Mehri","Philippe Laban","Sumuk Shashidhar","Marwa Abdulhai","Sergey Levine","Michel Galley","Dilek Hakkani-Tur"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Measuring and Mitigating the Distributional Gap Between Real and Simulated User Behaviors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07847","noteUrl":"/review?paper=review-neurips-2026-139106","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139107","title":"MM-IssueLoc: A Controlled Benchmark for Evaluating Visual Evidence in Multimodal Repository-Level Issue Localization","authors":["Shaoxiong Zhan","Shi Hu","Hai Lin","BoyuFeng","Andrew Gong","Zhengda Zhou","Jiaying Zhou","Yunyun Hou","Hao Su","Hai-Tao Zheng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strict top-three success is possible with two gold functions","plainExplanation":"The all-gold retrieval metric is said to force both top-one and top-three accuracy to zero whenever more than one gold function is present. The all-gold top-cutoff rule is forced to fail only when the gold-set size exceeds that cutoff, so two gold functions do not force top-three accuracy to zero.","primaryUrl":"https://arxiv.org/html/2607.15205","noteUrl":"/review?paper=review-neurips-2026-139107","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-139108","title":"Benchmarking Multi-Modal Graph-based Social Media Popularity Prediction","authors":["Utkarsh Sahu","Zhisheng Qi","Li Zhu","Yizhao Yang","Jun Li","Ryan Rossi","Yu Wang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Benchmarking Multi-Modal Graph-based Social Media Popularity Prediction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.27539","noteUrl":"/review?paper=review-neurips-2026-139108","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139111","title":"SynBench: A Benchmark for Differentially Private Text Generation","authors":["Yidan Sun","Viktor Schlegel","Srinivasan Nandakumar","Iqra Zahid","Yuping Wu","Yulong Wu","Hao Li","Jie Zhang","Warren Del-Pinto","Goran Nenadic","Siew Kei Lam","Anil A Bharath"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"SynBench: A Benchmark for Differentially Private Text Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.14594","noteUrl":"/review?paper=review-neurips-2026-139111","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139112","title":"SaaSBench: Exploring the Boundaries of Coding Agents in Long-Horizon Enterprise SaaS Engineering","authors":["Qingnan Ren","Shun Zou","Shiting Huang","Ziao Zhang","Kou Shi","Zhen Fang","Yiming Zhao","Yu Zeng","Qisheng Su","Lin Chen","Yong Wang","Zehui Chen","Xiangxiang Chu","Feng Zhao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SaaSBench: Exploring the Boundaries of Coding Agents in Long-Horizon Enterprise SaaS Engineering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17526","noteUrl":"/review?paper=review-neurips-2026-139112","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139116","title":"THEIA: A Multimodal Dataset and Benchmark for Vision-Language Analysis of Layout","authors":["Giuseppe Chiari","Michele Piccoli","Federico Viola","Davide Zoni"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"THEIA: A Multimodal Dataset and Benchmark for Vision-Language Analysis of Layout","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.35035","noteUrl":"/review?paper=review-neurips-2026-139116","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139117","title":"PixFoundation 2.0: Do Video Multi-Modal LLMs Use Motion in Visual Grounding?","authors":["Mennatullah Siam"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PixFoundation 2.0: Do Video Multi-Modal LLMs Use Motion in Visual Grounding?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2509.02807","noteUrl":"/review?paper=review-neurips-2026-139117","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139118","title":"OccStress: Stress-Testing the 4D Occupancy Forecasting Chain","authors":["Yu Zheng","Jie Hu","Jiaqi Xiong","Ruiping Liu","Junwei Zheng","Kailun Yang","Jiaming Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OccStress: Stress-Testing the 4D Occupancy Forecasting Chain","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.15621","noteUrl":"/review?paper=review-neurips-2026-139118","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139119","title":"Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability.","authors":["Roberto Stanzione","Jules Barbe","Magali Parrino","Jérémie Fourmann","Paul Boniol"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability.","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.01168","noteUrl":"/review?paper=review-neurips-2026-139119","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139123","title":"Cephalonauts One: A deep fMRI dataset for decoding naturalistic speech in the human brain","authors":["Antoine Collas","Louis Jalouzot","Géraud Ilinca","Corentin Caris","Romain Valabregue","Ahmed Hassayoune","David Gonçalves","Madeleine Hueber","Thaddée Delebarre","Savatovsky Julien","Clara Fonteneau","Charles Maussion","Bertrand Thirion","Alexis Thual"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Cephalonauts One: A deep fMRI dataset for decoding naturalistic speech in the human brain","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03558","noteUrl":"/review?paper=review-neurips-2026-139123","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139128","title":"What Drives Test-Time Adaptation for CLIP? A Controlled Empirical Study from an Update Perspective","authors":["Jiazhen Huang","Xiao Chen","Zhiming Liu","Yaru Sun","Jingyan Jiang","Zhi Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"What Drives Test-Time Adaptation for CLIP? A Controlled Empirical Study from an Update Perspective","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.14299","noteUrl":"/review?paper=review-neurips-2026-139128","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139129","title":"OTel: Open Telco AI Datasets, Benchmarks, and Models","authors":["Farbod Tavakkoli","Gregory Diamos","Kenneth Church","David Kanter","Mark Austin","Imtiaz Karim","Mirza M Rahman","Merouane DEBBAH","Zeinab Nezami","Ali Maatouk","Leandros Tassiulas","Rex Ying","Nick Sorros","Louis Powell","Nikolaos Vasiloglou","Ashish Vaswani","Somanshu Singla","Adarsh Chaluvaraju"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"OTel: Open Telco AI Datasets, Benchmarks, and Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.07766","noteUrl":"/review?paper=review-neurips-2026-139129","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139130","title":"EVALUATION CARDS: An Interpretive Layer for AI Evaluation Reporting","authors":["Avijit Ghosh","Anka Reuel-Lamparth","Wm. Matthew Kennedy","Jenny Chim","Yilin Huang","Srishti Yadav","Anastassia Kornilova","Damian Stachura","Kevin Klyman","Jennifer Mickel","Felix Friedrich","Max Lamparth","Jan Batzner","Anoop Mishra","Jeba Sania","Eliya Habba","Nathan Heath","Shalaleh Rismani","Usman Gohar","Andrea Loehr","David Manheim","Ruchira Dhar","Sree Harsha Nelaturu","Aarush Sinha","Leshem Choshen","Yixiong Hao","Yanan Long","Andrew Tran","Drishti Sharma","Ishan Khire","Amit Saha","Subramanyam Sahoo","Michael Hardy","Michael A. Riegler","Kabir Manghnani","Michelle Lin","Nancy Jiang","Asaf Yehudai","Nuno Moniz","Yacine Jernite","Zeerak Talat","Stella Biderman","Mubashara Akhtar","Jessica Ji","Aris Hofmann","Mykel J Kochenderfer","Sanmi Koyejo","Irene Solaiman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EVALUATION CARDS: An Interpretive Layer for AI Evaluation Reporting","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.09809","noteUrl":"/review?paper=review-neurips-2026-139130","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139131","title":"EvalAwareBench: A Benchmark for Measuring Evaluation Awareness in Frontier Language Models","authors":["Xinning Li","Kemunto Ochwang&amp;#x27;i","Ram Bharadwaj Aryasomayajula","Alexandra Souly","Robert Kirk"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"EvalAwareBench: A Benchmark for Measuring Evaluation Awareness in Frontier Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.01611","noteUrl":"/review?paper=review-neurips-2026-139131","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139132","title":"The PokeAgent Challenge: Competitive and Long-Context Learning at Scale","authors":["Seth Karten","Jake Grigsby","Tersoo Upaa","Junik Bae","Seonghun Hong","Hyunyoung Jeong","Jaeyoon Jung","Kun Kerdthaisong","Gyungbo Kim","Hyeokgi Kim","Yujin Kim","Eunju Kwon","Dongyu Liu","Patrick Mariglia","Sangyeon Park","Benedikt Schink","Xianwei Shi","Anthony sistilli","Joseph Twin","Arian Urdu","Matin Urdu","Qiao Wang","Ling Wu","Wenli Zhang","Kunsheng Zhou","Stephanie Milani","Kiran Vodrahalli","Amy Zhang","Fei Fang","Yuke Zhu","Chi Jin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The PokeAgent Challenge: Competitive and Long-Context Learning at Scale","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.15563","noteUrl":"/review?paper=review-neurips-2026-139132","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139133","title":"Workspace-Bench 1.0: Benchmarking AI Agents on Workspace Tasks with Large-Scale File Dependencies","authors":["Tang Zirui","Xuanhe Zhou","Yumou Liu","Linchun Li","Wang Weizheng","Hongzhang Huang","Jun Zhou","Jiachen Song","Shaoli Yu","Wang Jinqi","Zihang Zhou","Hongyi Zhou","YutingLv","Jinyang Li","Jiashuo Liu","Ruoyu Chen","Chunwei Liu","Guoliang Li","Jihua Kang","Fan Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Workspace-Bench 1.0: Benchmarking AI Agents on Workspace Tasks with Large-Scale File Dependencies","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.03596","noteUrl":"/review?paper=review-neurips-2026-139133","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139134","title":"GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations","authors":["Jingbo Yang","Henry Lai","Xiaowen Wang","Shiyu Chang","Yaar Harari","Evgeniy Gabrilovich"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.14498","noteUrl":"/review?paper=review-neurips-2026-139134","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139137","title":"Do Composed Image Retrieval Benchmarks Require Multimodal Composition?","authors":["Matteo Attimonelli","Alessandro De Bellis","Aryo Gema","Rohit Saxena","Monica Sekoyan","Wai-Chung Kwan","Claudio Pomo","Alessandro Suglia","Dietmar Jannach","Tommaso Di Noia","Pasquale Minervini"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Do Composed Image Retrieval Benchmarks Require Multimodal Composition?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14787","noteUrl":"/review?paper=review-neurips-2026-139137","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139139","title":"SALART-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images","authors":["Xiaoxiao Sun","Ruotian Zhang","Junzhe Huang","James Burgess","Serena Yeung-Levy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SALART-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.12671","noteUrl":"/review?paper=review-neurips-2026-139139","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139141","title":"Do Joint Audio-Video Generation Models Understand Physics?","authors":["Zijun Cui","Xiulong Liu","Hao Fang","Mingwei Xu","Jiageng Liu","Zexin Xu","Weiguo Pian","Shijian Deng","Feiyu Du","Chenming Ge","Yapeng Tian"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Do Joint Audio-Video Generation Models Understand Physics?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07061","noteUrl":"/review?paper=review-neurips-2026-139141","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139145","title":"MedMemoryBench: Benchmarking Agent Memory in Personalized Healthcare","authors":["Yihao Wang","Haoran Xu","Renjie Gu","Yixuan Ye","Xinyi Chen","Xinyu Mu","Yuan Gao","guo chunxiao","Peng Wei","Jinjie GU","Huan Li","Ke Chen","Lidan Shou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MedMemoryBench: Benchmarking Agent Memory in Personalized Healthcare","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11814","noteUrl":"/review?paper=review-neurips-2026-139145","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139146","title":"Theory-Scale Auto-Formalization of Logics for Computer Science","authors":["Yuming Feng","Frederick Pu","One An","Osbert Bastani","Li Zhang","Jiani Huang","Xujie Si","Ziyang Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Theory-Scale Auto-Formalization of Logics for Computer Science","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2606.26525","noteUrl":"/review?paper=review-neurips-2026-139146","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139148","title":"PatchBench: Measuring Collateral Damage in Activation Patching","authors":["Alexi Canesse","Mathis Le Bail","Maël JENNY","Clément Elliker","Mahammed El Sharkawy","Sonia Vanier"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Small covariance eigenvalues do not give exact annihilation","plainExplanation":"Small eigenvalues need not be zero.","primaryUrl":"https://arxiv.org/html/2610.10276","noteUrl":"/review?paper=review-neurips-2026-139148","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139149","title":"MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome","authors":["Fangda Ye","Yuxin Hu","Pengxiang Zhu","Yibo Li","Ziqi Jin","Yao Xiao","Yibo Wang","Lei Wang","Zhen Zhang","Lu Wang","Yue Deng","Bin Wang","Yifan Zhang","Liangcai Su","Xinyu Wang","He ZHAO","CHEN WEI","Qiang Ren","Bryan Hooi","Bo An","Shuicheng Yan","Lidong Bing"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.28407","noteUrl":"/review?paper=review-neurips-2026-139149","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139150","title":"MMOU: A Massive Multi-Task Omni Understanding and Reasoning Benchmark for Long and Complex Real-World Videos","authors":["Arushi Goel","Sreyan Ghosh","Vatsal Agarwal","Nishit Anand","Kaousheik Jayakumar","Lasha Koroshinadze","Yao Xu","Katie Lyons","James Case","Siddharth Gururani","Karan Sapra","Kevin Shih","Abhinav Shrivastava","Ramani Duraiswami","Dinesh Manocha","Andrew Tao","Bryan Catanzaro","Mohammad Shoeybi","Wei Ping"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"MMOU: A Massive Multi-Task Omni Understanding and Reasoning Benchmark for Long and Complex Real-World Videos","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.14145","noteUrl":"/review?paper=review-neurips-2026-139150","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139151","title":"Conditional Multi-Event Temporal Grounding in Long-Form Video","authors":["Yuanhao Zou","Arthad Kulkarni","Lucas Toñanez","Lincoln Spencer","Guangyu Sun","Tianxingjian Ding","Andong Deng","Yi Li","Shuangjun Liu","Yuan Li","Dashan Gao","Ning Bi","TAOTAO JING","Shuai Zhang","Chen Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stated thresholds rule out every nonself traversal edge","plainExplanation":"The temporal graph weights multiply cosine similarity by \\(e^{-d/\\sigma}\\), using discrete hop distance and the listed Short/Medium thresholds. Cosine similarity is at most one and a nonself edge has \\(d\\geq1\\), so \\(w\\leq e^{-1/\\sigma}\\).","primaryUrl":"https://arxiv.org/html/2606.15320","noteUrl":"/review?paper=review-neurips-2026-139151","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139154","title":"APM: Evaluating Style Personalization in LLMs with Arbitrary Preference Mappings","authors":["Philipp Spohn","Leander Girrbach","Zeynep Akata"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Adherence margins omit the user preference direction","plainExplanation":"For a negative preference weight, adherence scores nine and two give a positive routing margin even though user reward prefers the second response.","primaryUrl":"https://arxiv.org/html/2605.21063","noteUrl":"/review?paper=review-neurips-2026-139154","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139155","title":"SpreadsheetBench 2: Evaluating Agents on End-to-End Business Spreadsheet Workflows","authors":["Jian Zhu","Yuzheng Zhang","Zeyao Ma","Bohan Zhang","Armin Schoepf","Daniel Woloch","Peter Wang","Guangyu Robert Yang","Samuel Jacob","Siddharth Nagisetty","Abhiram Chundru","Jean Lin","Spencer Mateega","Jing Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SpreadsheetBench 2: Evaluating Agents on End-to-End Business Spreadsheet Workflows","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.29955","noteUrl":"/review?paper=review-neurips-2026-139155","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139156","title":"OpenVTON-Bench: A Large-Scale High-Resolution Benchmark for Controllable Virtual Try-On Evaluation","authors":["Jin Li","Tao Chen","Kai Wen","Siqi Yin","Shuai Jiang","Weijie Wang","Jingwen Luo","Chenhui Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OpenVTON-Bench: A Large-Scale High-Resolution Benchmark for Controllable Virtual Try-On Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.22725","noteUrl":"/review?paper=review-neurips-2026-139156","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139157","title":"Freeing the Law with LOCUS: A Local Ordinance Corpus for the United States","authors":["Denis Peskoff","Joe Barrow","Christopher Vu","Diag Davenport"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Freeing the Law with LOCUS: A Local Ordinance Corpus for the United States","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.19334","noteUrl":"/review?paper=review-neurips-2026-139157","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139158","title":"PosteriorBench: From Point Estimates to Posterior Matching in Evaluating Generative Inverse Solvers","authors":["Jiachen Yao","Zi-Siang Hsu","Xi Deng","Aditi Gupta","Xin Ju","Animashree Anandkumar"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Averaging residual squares tempers the Gaussian likelihood","plainExplanation":"Dividing residual energy by the observation count changes the likelihood temperature and effective Gaussian noise scale.","primaryUrl":"https://arxiv.org/html/2609.20794","noteUrl":"/review?paper=review-neurips-2026-139158","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139159","title":"TimeWarp: Evaluating Web Agents by Revisiting the Past","authors":["Farhan Ishmam","Kenneth Marino"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TimeWarp: Evaluating Web Agents by Revisiting the Past","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2603.04949","noteUrl":"/review?paper=review-neurips-2026-139159","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139160","title":"BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems","authors":["Yutaro Yamada","Kei Hiroshima","Nozomu Yoshinari","Kento Uchida","Shinichi Shirakawa"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.02612","noteUrl":"/review?paper=review-neurips-2026-139160","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139161","title":"CausalDriveBench: Evaluating Causal Reasoning in Vision-Language-Action Models for Autonomous Driving","authors":["Narendiran Chembu","Navvrat Rao","Shreedhar Kodate","Gayatri S Banda","Arko Sarkar","Abhinav Khanna","Rajarshee Das","Umesh Kanala","Siddharth Khandelwal","Kumar Aman","Aish Dubey","Kaustubh Beedkar","Arjun Jain"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"CausalDriveBench: Evaluating Causal Reasoning in Vision-Language-Action Models for Autonomous Driving","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.32157","noteUrl":"/review?paper=review-neurips-2026-139161","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139162","title":"BrainTRACE: Tracing Longitudinal, Multimodal, and Volumetric Evidence in Brain MRI Clinical Reasoning","authors":["Qizhen Lan","Mengchen Fan","Hang Zhang","Jingwei Duan","Moule Lin","Jialin Chen","Baocheng Geng","Xiaoqian Jiang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"BrainTRACE: Tracing Longitudinal, Multimodal, and Volumetric Evidence in Brain MRI Clinical Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.06571","noteUrl":"/review?paper=review-neurips-2026-139162","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139163","title":"Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering with Generative Optimization","authors":["Dapeng Jiang","Yizhe Chi","Kaisen Yang","Tianwei Luo","Boshi Zhang","Deyao Hong","Dianqiao Lei","Yifan Zhou","Weiyang Jin","Xiaoyan Fan","Han Hao","Zhe Cao","Qian Houde","Qingle Liu","Bohan Lyu","Bowen Wang","Situ Wang","Youjie Zheng","Bingxiang He","Eren Cai","Calvin Xiao","Qinhuai Na"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering with Generative Optimization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.12290","noteUrl":"/review?paper=review-neurips-2026-139163","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139164","title":"EEG Benchmarking Needs a Task Specification Layer: NeuroDoc for Rulebook-Guided, Executable Benchmark Construction","authors":["Chengxuan Qin","致格 陈","Pengshu","Rui Yang","JiPing Cui","Yikai Dong","Jun Li","Liu Peng","Zhida Shang","Mingze Tang","KC Tan","Jibin Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EEG Benchmarking Needs a Task Specification Layer: NeuroDoc for Rulebook-Guided, Executable Benchmark Construction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.22925","noteUrl":"/review?paper=review-neurips-2026-139164","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139165","title":"SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery","authors":["Jiajun Jiang","Chunliang Hua","Zichun Chen","Yanxing Wu","Zeyuan YANG","Jie Song","Xiao Hu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31507","noteUrl":"/review?paper=review-neurips-2026-139165","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139166","title":"MITO: A Millimeter-Wave Dataset and Simulator for Non-Line-of-Sight Perception","authors":["Tara Boroushaki","Laura Dodds","Cusuh Ham","Fadel Adib"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MITO: A Millimeter-Wave Dataset and Simulator for Non-Line-of-Sight Perception","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2502.10259","noteUrl":"/review?paper=review-neurips-2026-139166","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139167","title":"VeriTrip: A Verifiable Benchmark for Travel Planning Agents over Unstructured Web Corpora","authors":["Yuting Xu","Jiayi Tian","Jian Liang","Xin Xiong","Hang Zhang","Mu Xu","Xiao-Yu Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"VeriTrip: A Verifiable Benchmark for Travel Planning Agents over Unstructured Web Corpora","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.28683","noteUrl":"/review?paper=review-neurips-2026-139167","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139169","title":"When Simulation Lies: A Sim-to-Real Benchmark and Domain-Randomized RL Recipe for Tool-Use Agents","authors":["Xiaolin Zhou","Aojie Yuan","Zheng Luo","Zipeng Ling","Xixiao Pan","Yicheng Gao","Haiyue Zhang","Jiate Li","Shuli Jiang","Prince Z Wang","Zixuan Zhu","Jinbo Liu","Ryan Rossi","Hua Wei","Xiyang Hu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"When Simulation Lies: A Sim-to-Real Benchmark and Domain-Randomized RL Recipe for Tool-Use Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11928","noteUrl":"/review?paper=review-neurips-2026-139169","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139172","title":"PathView-Bench: Can Multimodal Large Language Models Achieve Fine-grained Multiscale Understanding of Pathology Images?","authors":["Zongyi Chen","Yu Liang","Jie Lin","Liansheng Wang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"PathView-Bench: Can Multimodal Large Language Models Achieve Fine-grained Multiscale Understanding of Pathology Images?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.28318","noteUrl":"/review?paper=review-neurips-2026-139172","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139174","title":"GraphIP–Bench: How Hard Is It to Steal a Graph Neural Network, and Can We Stop It?","authors":["Kaixiang Zhao","Bolin Shen","Yuyang Dai","Shayok Chakraborty","Yushun Dong"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"GraphIP–Bench: How Hard Is It to Steal a Graph Neural Network, and Can We Stop It?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12827","noteUrl":"/review?paper=review-neurips-2026-139174","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139175","title":"CruxBench: A Benchmark of Information Discovery","authors":["Lina Piao","Amelia Hui Dai","Nick Merrill","Nadja Flechner","Ezra Karger","Haifeng Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Full revelation does not make realized KL constant","plainExplanation":"The realized KL is \\(-\\log_2p\\) when \\(U=1\\), and \\(-\\log_2(1-p)\\) when \\(U=0\\).","primaryUrl":"https://arxiv.org/html/2609.35879","noteUrl":"/review?paper=review-neurips-2026-139175","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139176","title":"AbsoluteDegradation: A Physics-Inspired Synthetic Film-Degradation Pipeline and Archival Film Restoration Benchmark","authors":["Mikołaj Jastrzębski","Dawid Glinkowski","Dawid Zieliński","Daniel Borkowski","Wojciech M Kozłowski","Kamil Adamczewski"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Gaussian driving noise can make the displacement nonreal","plainExplanation":"A nondegenerate Gaussian update assigns positive probability to \\(\\theta<0\\), even from a positive initial value such as \\(0.25\\). The real-valued generative specification needs a positivity rule.","primaryUrl":"https://arxiv.org/html/2607.02131","noteUrl":"/review?paper=review-neurips-2026-139176","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139177","title":"Multi-site PPG: An In-the-Wild Physiological Dataset from Emerging Multi-Site Wearables","authors":["Jiayi Shao","Jiaying Ye","ShengYao Liu","Zachary Englhardt","Girish Narayanswamy","Vikram Iyer","Qiuyue (Shirley) Xue"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Multi-site PPG: An In-the-Wild Physiological Dataset from Emerging Multi-Site Wearables","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17859","noteUrl":"/review?paper=review-neurips-2026-139177","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139179","title":"MOSAIC-Bench: Measuring Compositional Vulnerability Induction in Coding Agents","authors":["Jonathan Steinberg","Oren Gal"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MOSAIC-Bench: Measuring Compositional Vulnerability Induction in Coding Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.03952","noteUrl":"/review?paper=review-neurips-2026-139179","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139182","title":"C3VD-DEFCOL: A Deformable Colonoscopy Dataset with Time-Resolved 3D Ground Truth and Realistic Appearance","authors":["Ethan Luk","Mayank Golhar","Anthony Song","Raúl Iranzo","Victor M. Batlle","Lalithkumar Seenivasan","JMM Montiel","Nicholas Durr"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"C3VD-DEFCOL: A Deformable Colonoscopy Dataset with Time-Resolved 3D Ground Truth and Realistic Appearance","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.07891","noteUrl":"/review?paper=review-neurips-2026-139182","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139183","title":"Are Multimodal Benchmarks Really Useful? Item-Level Multimodal Benchmark Diagnosis via Structure-Response Co-Calibration","authors":["Shiqi Zhang","Weixin Zeng","Ziheng Zhang","Jiuyang Tang","Xiang Zhao"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Are Multimodal Benchmarks Really Useful? Item-Level Multimodal Benchmark Diagnosis via Structure-Response Co-Calibration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.14542","noteUrl":"/review?paper=review-neurips-2026-139183","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139184","title":"ViTeX-Bench: Benchmarking High-Fidelity Video Scene Text Editing","authors":["Xinghao Chen","Xiangbo Gao","Jiongze Yu","Yuheng Wu","Zhengzhong Tu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ViTeX-Bench: Benchmarking High-Fidelity Video Scene Text Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.40356","noteUrl":"/review?paper=review-neurips-2026-139184","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139185","title":"SANEval: Open-Vocabulary Compositional Benchmarks with Failure-mode Diagnosis","authors":["Rishav Pramanik","Ian Nielsen","Jeffrey Smith","Saurav Pandit","Ravi P Ramachandran","Zhaozheng Yin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SANEval: Open-Vocabulary Compositional Benchmarks with Failure-mode Diagnosis","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.00249","noteUrl":"/review?paper=review-neurips-2026-139185","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139192","title":"Seahorse: A Unified Benchmarking Framework for Spatiotemporal Event Modeling","authors":["Yahya Aalaila","Sebastian Vollmer","Gerrit Großmann"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Seahorse: A Unified Benchmarking Framework for Spatiotemporal Event Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.01022","noteUrl":"/review?paper=review-neurips-2026-139192","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139193","title":"Crossing the Validation crisis: Cross-validation reduces benchmarking variance surprisingly well","authors":["Célestin Eve","Gael Varoquaux","Thomas Moreau"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Crossing the Validation crisis: Cross-validation reduces benchmarking variance surprisingly well","plainExplanation":"Increasing $K$ averages over partitions of the same observations. It removes randomness from the choice of partition but retains the error caused by drawing a finite study set. Distinct population scores and paired comparisons do not remove that error. The paper already identifies the corresponding nonzero variance floor after Eq. (9).","primaryUrl":"https://arxiv.org/html/2606.12552","noteUrl":"/review?paper=review-neurips-2026-139193","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","interpretation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139194","title":"A Large-Scale Multi-Source Dataset Linking Hacker Community Discourse to the CVE Vulnerability Lifecycle","authors":["Benjamin M Ampel","Sagar Samtani"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Weighted score fusion is not reciprocal rank fusion","plainExplanation":"Weighted score fusion depends on score magnitudes, whereas reciprocal rank fusion depends only on ranks. The two descriptions specify different rankings.","primaryUrl":"https://arxiv.org/html/2605.03158","noteUrl":"/review?paper=review-neurips-2026-139194","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139197","title":"InterLV-Search: Benchmarking Interleaved Multimodal Agentic Search","authors":["Bohan Hou","Jiuning Gu","Jiayan Guo","Ronghao Dang","Sicong Leng","Xin Li","Xuemeng Song","Jianfei Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"InterLV-Search: Benchmarking Interleaved Multimodal Agentic Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07510","noteUrl":"/review?paper=review-neurips-2026-139197","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139198","title":"PhysElite:  How Far Are LLMs from Solving Olympiad-Level Physics Problems?","authors":["Ruoran Xu","Wending Gao","Liyunfeng Chen","Aixin Shi","Haoyu Cheng","Zixiang Fang","Yiqiang Zou","Qiufeng Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"PhysElite:  How Far Are LLMs from Solving Olympiad-Level Physics Problems?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.25097","noteUrl":"/review?paper=review-neurips-2026-139198","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139201","title":"Attack Selection In Agentic AI Control Evaluations Meaningfully Decreases Safety","authors":["Catherine Ge-Wang","Tyler Crosse","Benjamin Hadad","Joachim Schaeffer","Ram Potham","Tyler Tracy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Attack Selection In Agentic AI Control Evaluations Meaningfully Decreases Safety","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06529","noteUrl":"/review?paper=review-neurips-2026-139201","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139204","title":"HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models","authors":["Emmy Liu","Varun Gangal","Michael Yu","Zhuofu Tao","Karan Singh","Sachin Kumar","Steven Y. Feng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.19341","noteUrl":"/review?paper=review-neurips-2026-139204","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139207","title":"CyBiasBench: Benchmarking Bias in LLM Agents for Cyber-Attack Scenarios","authors":["Taein Lim","Seongyong Ju","Munhyeok Kim","Hyunjun Kim","Hoki Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CyBiasBench: Benchmarking Bias in LLM Agents for Cyber-Attack Scenarios","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07830","noteUrl":"/review?paper=review-neurips-2026-139207","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139208","title":"LL-Bench: Rethinking Low-Level Vision Evaluation in the Era of Large-Scale Generative Models","authors":["Lu Liu","Huiyu Duan","Chenxin Zhu","Jintong Lu","Haoyun Jiang","Liu Yang","Qiang Hu","Guangtao Zhai","Xiaoyun Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"LL-Bench: Rethinking Low-Level Vision Evaluation in the Era of Large-Scale Generative Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.02535","noteUrl":"/review?paper=review-neurips-2026-139208","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139210","title":"RealityTest: How People Probe AI Identity and Whether Models Disclose It","authors":["Anna Gausen","Sarenne Wallbridge","Bessie O&amp;#x27;Dell","Christopher Summerfield","Hannah Rose Kirk"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RealityTest: How People Probe AI Identity and Whether Models Disclose It","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.00168","noteUrl":"/review?paper=review-neurips-2026-139210","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139213","title":"SynDORBench: Evaluating LVLM Perceptual Robustness Under Physically Constrained Visibility Conditions","authors":["Jeremy S G Yee","Daniel Zhengkui Wang","Zhiyuan Zhang","Avinash Anand","Timothy Liu","Benedict Chan","Aik B Ng","Simon See"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SynDORBench: Evaluating LVLM Perceptual Robustness Under Physically Constrained Visibility Conditions","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.31823","noteUrl":"/review?paper=review-neurips-2026-139213","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139214","title":"scShapeBench: Discovering geometry from high dimensional scRNAseq data","authors":["Andrew J Steindl","Joao Felipe Rocha","Brian T Di Bassinga","Zachary Warren","Matthew Scicluna","César Miguel Valdez Córdova","Shabarni Gupta","Leire Torices","Daniel Neumann","Timothy J Mann","Ihuan Gunawan","Dhananjay Bhaskar","John G Lock","Christine L Chaffer","Guy Wolf","Smita Krishnaswamy"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"scShapeBench: Discovering geometry from high dimensional scRNAseq data","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12662","noteUrl":"/review?paper=review-neurips-2026-139214","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139215","title":"Are LLM Safety Judges Policy-Invariant? A Three-Principle Stress-Test","authors":["Shihao Weng","Yang Feng","Xiaofei Xie"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Conditional independence does not give the stated ensemble error bound","plainExplanation":"Negative excess flip rates can make the score exceed one. The unit-cube theorem is valid, but its input restriction is not automatic for the signed statistic used in the definition.","primaryUrl":"https://arxiv.org/html/2605.06161","noteUrl":"/review?paper=review-neurips-2026-139215","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139217","title":"SourceBench: Can AI Answers Reference Quality Web Sources?","authors":["Hexi Jin","Xurui Liu","Yuheng Li","Simran Malik","Yiying Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SourceBench: Can AI Answers Reference Quality Web Sources?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.16942","noteUrl":"/review?paper=review-neurips-2026-139217","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139220","title":"FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics","authors":["Qiran Zou","Hou Hei Lam","Wenhao Zhao","Tingting Chen","Yiming Tang","Samson Yu","Yingtao Zhu","srinivas anumasa","Zufeng Zhang","Tianyi Zhang","Chang Liu","zhengyao Jiang","Anirudh Goyal","Dianbo Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Baseline-normalized improvement remains bounded for nonnegative loss","plainExplanation":"Appendix E calls normalized improvement unbounded above when it sets worst value equal to a positive baseline for nonnegative lower-is-better metrics. Nonnegative lower-is-better loss keeps the stated positive-baseline normalized improvement at most one, even when raw loss has no upper bound.","primaryUrl":"https://arxiv.org/html/2605.17373","noteUrl":"/review?paper=review-neurips-2026-139220","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-139222","title":"AffordSim: A Scalable Data Generator and Benchmark for Affordance-Aware Robotic Manipulation","authors":["Mingyang Li","Haofan Xu","Haowen Sun","Xinzhe Chen","Sihua Ren","Liqi Huang","Chenyang Miao","Xinyang Sui","jiawei Ye","Qiongjie Cui","Zeyang Liu","Xingyu Chen","Xuguang Lan"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"AffordSim: A Scalable Data Generator and Benchmark for Affordance-Aware Robotic Manipulation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.11674","noteUrl":"/review?paper=review-neurips-2026-139222","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139224","title":"TerminalWorld: Benchmarking Agents on Real-World Terminal Tasks","authors":["Zhaoyang Chu","Jiarui Hu","Xingyu Jiang","Pengyu Zou","Han Li","Chao Peng","Peter O&amp;#x27;Hearn","Earl Barr","Mark Harman","Federica Sarro","He Ye"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TerminalWorld: Benchmarking Agents on Real-World Terminal Tasks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22535","noteUrl":"/review?paper=review-neurips-2026-139224","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139226","title":"PitchBench: Measuring Pitch Hearing in Audio-Language Models","authors":["Milan Liessens Dujardin","Song-Ze Yu","Craver C Thomas-Smith","David Chan","Karina Nguyen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"PitchBench: Measuring Pitch Hearing in Audio-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.26176","noteUrl":"/review?paper=review-neurips-2026-139226","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139227","title":"HiGraph: A Large-Scale Hierarchical Graph Dataset for Malware Analysis","authors":["Han Chen","Hanchen Wang","Hongmei Chen","Ying Zhang","Lu Qin","Wenjie Zhang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The normalized time area has one extra trapezoid","plainExplanation":"The AUT expression uses \\(N\\) monthly values but sums \\(N\\) adjacent trapezoids and normalizes by \\(N-1\\). There are only \\(N-1\\) adjacent intervals.","primaryUrl":"https://arxiv.org/html/2509.02113","noteUrl":"/review?paper=review-neurips-2026-139227","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139229","title":"PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction","authors":["Zhiwen Yang","Pan Liu","yifan Li","Yunhua Zhong","Jun Xia"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.01945","noteUrl":"/review?paper=review-neurips-2026-139229","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139231","title":"GT-Free OCR Metrics: A Reference-Free Evaluation Framework for Document OCR Systems","authors":["Kshitij Singh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"GT-Free OCR Metrics: A Reference-Free Evaluation Framework for Document OCR Systems","plainExplanation":"","primaryUrl":"https://zenodo.org/api/records/20070291/files/main.pdf/content","noteUrl":"/review?paper=review-neurips-2026-139231","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139233","title":"EcoGym: Evaluating LLMs for Long-Horizon Plan-and-Execute in Interactive Economies","authors":["Xueyu Hu","Jinxiang Xia","Shengze Xu","Kangqi Song","Yishuo Yuan","Guibin Zhang","JinCheng Ren","Boyu Feng","Li Lu","Tieyong Zeng","Jiaheng Liu","Minghao Liu","He Zhu","Eleanor Jiang","Wei Wang","Wangchunshu Zhou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EcoGym: Evaluating LLMs for Long-Horizon Plan-and-Execute in Interactive Economies","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09514","noteUrl":"/review?paper=review-neurips-2026-139233","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139235","title":"MEMAUDIT: An Exact Package-Oracle Evaluation Protocol for Budgeted Long-Term LLM Memory Writing","authors":["Nishant Bhargava","Rodrigo S Barrento"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MEMAUDIT: An Exact Package-Oracle Evaluation Protocol for Budgeted Long-Term LLM Memory Writing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.02199","noteUrl":"/review?paper=review-neurips-2026-139235","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139237","title":"RescueBench: Can Embodied Agents Save Lives in the Wild?","authors":["Kui Wu","Beiyu Guo","Hao Chen","Shuhang Xu","Yuling Li","Yongdan Zeng","Zhoujun Li","Yizhou Wang","Fangwei Zhong"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"RescueBench: Can Embodied Agents Save Lives in the Wild?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.01848","noteUrl":"/review?paper=review-neurips-2026-139237","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139238","title":"SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs","authors":["Huizhe Zhang","Jintang Li","Yuchang Zhu","Huazhen Zhong","Liang Chen"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2509.21342","noteUrl":"/review?paper=review-neurips-2026-139238","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139240","title":"AgentHop: A Diagnostic Benchmark for Agentic Multi-Hop Scientific Question Answering","authors":["Chanhee Park","Jeongho Yoon","Sungbin Han","Hyeonseok Moon","Heuiseok Lim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AgentHop: A Diagnostic Benchmark for Agentic Multi-Hop Scientific Question Answering","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.34428","noteUrl":"/review?paper=review-neurips-2026-139240","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139241","title":"Benchmarking Vietnamese Legal Knowledge of Large Language Models","authors":["Dong Nguyen Tien","Nguyen Minh-Anh","Thanh D Hoang","Nguyen T Ngoc","Quang M Xuan","Phan P Hai","Nguyen T Anh","Binh Vu","Dung Le"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Benchmarking Vietnamese Legal Knowledge of Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.14554v5","noteUrl":"/review?paper=review-neurips-2026-139241","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139242","title":"ConnectomeBench2: A Unified Benchmark for Automated Connectomic Proofreading","authors":["Jeff Brown","Tim Farkas","Gleb Razgar","Edward Boyden"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ConnectomeBench2: A Unified Benchmark for Automated Connectomic Proofreading","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.21116","noteUrl":"/review?paper=review-neurips-2026-139242","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139244","title":"CUDABeaver: Benchmarking LLM-Based Automated CUDA Debugging","authors":["Shiyang Li","Haoyang Chen","Mattia Fazzini","Caiwen Ding"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A repeated coarse error signature does not establish a fixed point","plainExplanation":"The no-progress rule interprets three equal error-category/signature tuples as a fixed point of the repair process. The signature keeps only the first matched error pattern and discards identifiers and code state.","primaryUrl":"https://arxiv.org/html/2605.08455","noteUrl":"/review?paper=review-neurips-2026-139244","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139245","title":"AnyHand: A Large-Scale Synthetic Dataset for RGB(-D) Hand Pose Estimation","authors":["Chen Si","Yulin Liu","Bo Ai","Jianwen Xie","Rolandos Alexandros Potamias","Chuanxia Zheng","Hao Su"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AnyHand: A Large-Scale Synthetic Dataset for RGB(-D) Hand Pose Estimation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.25726","noteUrl":"/review?paper=review-neurips-2026-139245","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139246","title":"VocalCoachBench: Benchmarking Audio-Language Models on Expert Feedback for Singing","authors":["Hayeon Bang","Hounsu Kim","Wonil Kim","Juhan Nam"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"VocalCoachBench: Benchmarking Audio-Language Models on Expert Feedback for Singing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.04241","noteUrl":"/review?paper=review-neurips-2026-139246","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139247","title":"HSCO-Bench: An Agent-Driven End-to-End Hardware-Software Co-design Benchmark for Systems-on-Chip","authors":["Pei-Huan Tsai","Kuan-Lin Chiu","William Baisi","Pin-Yu Chen","Luca P Carloni"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"HSCO-Bench: An Agent-Driven End-to-End Hardware-Software Co-design Benchmark for Systems-on-Chip","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.19399","noteUrl":"/review?paper=review-neurips-2026-139247","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139248","title":"OverLay++: Dense-Overlap Layout-to-Image Generation Dataset","authors":["Shivansh Aggarwal","Shresth Grover","Divyansh Srivastava","Haiyang Xu","Bingnan Li","Xiang Zhang","Ethan Armand","Chuan Li","Jianwen Xie","Zhuowen Tu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"OverLay++: Dense-Overlap Layout-to-Image Generation Dataset","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.09071","noteUrl":"/review?paper=review-neurips-2026-139248","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139250","title":"PHMForge: Evaluating LLM Agents on Industrial Prognostics through MCP-Native, Algorithm-Grounded Tools","authors":["Yusheng Li","Tianjun Feng","Yunfeng Chen","Chun-Yi Tsai","Yihan Sun","Dhaval Patel","Kaoutar El Maghraoui","Shuxin Lin","Ayan Das"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PHMForge: Evaluating LLM Agents on Industrial Prognostics through MCP-Native, Algorithm-Grounded Tools","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.01532","noteUrl":"/review?paper=review-neurips-2026-139250","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139252","title":"Signature-Kernel Evaluation Metrics for Robust Probabilistic and Tail-Event Forecasting","authors":["Benjamin R Redhead","Thomas L Lee","Peng Gu","Víctor Elvira","Amos Storkey"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Signature-Kernel Evaluation Metrics for Robust Probabilistic and Tail-Event Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.10182","noteUrl":"/review?paper=review-neurips-2026-139252","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139253","title":"RustMizan: A Compilable, Contamination-Aware Benchmarking Framework for Rust Vulnerabilities","authors":["Tarek Elsayed","Shiping Yang","Eunsong Koh","Sanika Goyal","Vincent Huang","Paul Ngo","Nathan Young","Mohammad Omidvar Tehrani","Alvyn Kang","Arnell Kang","Zeyu Chen","Angelica Moreira","Xuan Feng","Angel Chang","Nick Sumner","Steve Ko"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RustMizan: A Compilable, Contamination-Aware Benchmarking Framework for Rust Vulnerabilities","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.04729","noteUrl":"/review?paper=review-neurips-2026-139253","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139254","title":"Predicting Nothing Beats SAM 3: Revisiting Evaluation in Video Object Segmentation","authors":["Jihwan Hong","Woohyeon Park","Jaeik Kim","Jaeyoung Do"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Predicting Nothing Beats SAM 3: Revisiting Evaluation in Video Object Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02946","noteUrl":"/review?paper=review-neurips-2026-139254","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139255","title":"Exploring the Limits of Compositional Generalization in Vision-Language-Action Manipulation","authors":["Xupeng Zhang","Yantai Yang","Chang Guo","Zhaokai Yin","Zhipeng Zhang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Exploring the Limits of Compositional Generalization in Vision-Language-Action Manipulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.15425","noteUrl":"/review?paper=review-neurips-2026-139255","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139256","title":"Auditing Attention Head Masking for Out-of-Distribution Detection: Cross-Architecture Wins, Failures, and Polarity Inversions","authors":["Vishnu Bindu Balachandran"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The significance test does not establish polarity inversion","plainExplanation":"A significant difference between two methods does not establish that either is below chance. The tested paired differences and the claimed below-chance contrasts have different null hypotheses.","primaryUrl":"https://zenodo.org/api/records/20281393/files/Vishnu.pdf/content","noteUrl":"/review?paper=review-neurips-2026-139256","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139260","title":"How Private is Private? A Comparative Study for Face De-Identification","authors":["Hui Wei","Hao Yu","Hui Kuurila-Zhang","Guoying Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The regularized score can exceed its stated upper endpoint","plainExplanation":"The geometric mean has the stated zero-limit property too. Its weaker suppression at finite component values is a different criterion.","primaryUrl":"https://arxiv.org/pdf/2610.10334","noteUrl":"/review?paper=review-neurips-2026-139260","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_ambiguity","formula_error","missing_assumption"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-139261","title":"Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks","authors":["David Schmotz","Luca Beurer-Kellner","Sahar Abdelnabi","Maksym Andriushchenko"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.20156","noteUrl":"/review?paper=review-neurips-2026-139261","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139262","title":"ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?","authors":["Zhun Wang","Nico Schiller","Hongwei Li","Srijiith S Narayana","Milad Nasr","Nicholas Carlini","Xiangyu Qi","Eric Wallace","Elie Bursztein","Luca Invernizzi","Kurt Thomas","Yan Shoshitaishvili","Wenbo Guo","Jingxuan He","Thorsten Holz","Dawn Song"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11086","noteUrl":"/review?paper=review-neurips-2026-139262","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139263","title":"Chess-World-Model: A 10M-Game Benchmark for Exact State Tracking from Chess Move Sequences","authors":["Benjamin Walker","Terry Lyons"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Chess-World-Model: A 10M-Game Benchmark for Exact State Tracking from Chess Move Sequences","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30100","noteUrl":"/review?paper=review-neurips-2026-139263","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139264","title":"L-FAME: Longitudinal Focused Attention Meditation EEG Dataset and Benchmark","authors":["Angqi Li","Basit R Syed","Hamzeh Alzweri","Taosheng Liu","Barry H Cohen","Saiprasad Ravishankar"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Baseline balance cannot establish MCAR","plainExplanation":"MCAR requires missingness to be independent of unobserved follow-up outcomes as well as baseline variables.","primaryUrl":"https://arxiv.org/html/2605.22893","noteUrl":"/review?paper=review-neurips-2026-139264","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139267","title":"Benchmarks Are Not Atomic: Composition-Aware LLM Evaluation using BenchHub","authors":["Eunsu Kim","Haneul Yoo","Guijin Son","Hitesh Patel","Amit Agarwal","Alice Oh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Benchmarks Are Not Atomic: Composition-Aware LLM Evaluation using BenchHub","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2506.00482","noteUrl":"/review?paper=review-neurips-2026-139267","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139268","title":"ManiTaskGen: A Comprehensive Task Generator for Benchmarking and Improving Vision-Language Agents on Embodied Decision-Making","authors":["Liu Dai","Haina Wang","Weikang Wan","Hao Su"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"ManiTaskGen: A Comprehensive Task Generator for Benchmarking and Improving Vision-Language Agents on Embodied Decision-Making","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2505.20726","noteUrl":"/review?paper=review-neurips-2026-139268","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139269","title":"On Data Engineering for Scaling LLM Terminal Capabilities","authors":["Renjie Pi","Grace Lam","Mohammad Shoeybi","Pooya Jannaty","Bryan Catanzaro","Wei Ping"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"On Data Engineering for Scaling LLM Terminal Capabilities","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.21193","noteUrl":"/review?paper=review-neurips-2026-139269","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139270","title":"Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments","authors":["Parth Asawa","Christopher M Glaze","Gabriel Orlanski","Ramya Ramakrishnan","Benji Xu","Asim Biswal","Vincent Chen","Frederic Sala","Matei A Zaharia","Joseph Gonzalez"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05661","noteUrl":"/review?paper=review-neurips-2026-139270","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139271","title":"OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data","authors":["Louisa Cornelis","Johan Mathe","Louis Van Langendonck","Guillermo Bernárdez","Nina Miolane"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15511","noteUrl":"/review?paper=review-neurips-2026-139271","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139272","title":"KNN Implementation Details Can Dramatically Change Performance: An Example from Cover Trees","authors":["Amol Khanna","Edward Raff"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"KNN Implementation Details Can Dramatically Change Performance: An Example from Cover Trees","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16940","noteUrl":"/review?paper=review-neurips-2026-139272","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139273","title":"TasteBench: multimodal benchmark for sensory prediction, from molecules to sustainable foods","authors":["Anna Thomas","Sohum Patnaik","Caroline Cotto","Benjamin Sanchez-Lengeling"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"TasteBench: multimodal benchmark for sensory prediction, from molecules to sustainable foods","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02599","noteUrl":"/review?paper=review-neurips-2026-139273","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139275","title":"OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations","authors":["Faadil Mustun","Chiara Semenzin","Roberto Dessi","Pablo Robin Guerrero","pierre orhan","Alexis Emanuelli","Emanuele Rossi","Yair Lakretz","Gonzalo de Polavieja","Germán Sumbre"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.34839","noteUrl":"/review?paper=review-neurips-2026-139275","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139277","title":"Is Agentic AI Ready for Real-World Hardware Engineering? A Deep Dive with Phoenix-bench","authors":["Qingyun Zou","Feng Yu","Hongshi Tan","Jiahao Cui","Bingsheng He","Weng-Fai Wong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Is Agentic AI Ready for Real-World Hardware Engineering? A Deep Dive with Phoenix-bench","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15226","noteUrl":"/review?paper=review-neurips-2026-139277","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139280","title":"TRACE: Tourism Recommendation with Accountable Citation Evidence","authors":["Zixu Zhao","SIJIN WANG","Yu Hou","YUANYUAN XU","Yufan Sheng","Xike Xie","Wenjie Zhang","Won-Yong Shin","Xin Cao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Near-zero correlation does not rule out score inflation","plainExplanation":"Adding valid quotes raises that fraction whenever an unmatched quote remains: one matching and one unmatched quote give \\(1/2\\), while nine matching and one unmatched give \\(9/10\\). The inference of no mechanical inflation fails; the descriptive correlation remains a descriptive result.","primaryUrl":"https://arxiv.org/html/2605.07677","noteUrl":"/review?paper=review-neurips-2026-139280","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139281","title":"Not All Proofs Are Equal: Evaluating LLM Proof Quality Beyond Correctness","authors":["Ivo Petrov","Jasper Dekoninck","Dimitar I. Dimitrov","Martin Vechev"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Larger compressibility ratios mean more removable text","plainExplanation":"Section 3.2 defines \\(C=|p|/|\\tilde p|\\), then says smaller ratios correspond to a larger removable fraction. The removable fraction increases with the original-to-shortened length ratio, so the sentence reverses its direction while smaller ratios can still mean more concise original proofs.","primaryUrl":"https://arxiv.org/html/2605.10379","noteUrl":"/review?paper=review-neurips-2026-139281","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-139282","title":"Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty","authors":["Sarina Penquitt","Jonathan Klees","Antonia van Betteray","Parssa Jashnieh","Peter Stehr","Matthias Rottmann","Lars Schmarje"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Hard annotations alone do not recover average precision","plainExplanation":"Hard annotations do not make soft predicted-label similarity binary, so the stated conditions do not ensure an exact AP reduction. The special-case conditions are underspecified. Under the stronger hard-prediction reading, the binary reduction can hold; AUSRC itself is not refuted.","primaryUrl":"https://arxiv.org/html/2609.21822","noteUrl":"/review?paper=review-neurips-2026-139282","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139283","title":"SLVMBench: Skill Learning from Video Memory","authors":["Yudong Yang","Guangzhi Sun","Yixuan Li","Wei Li","Zejun MA","Chao Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SLVMBench: Skill Learning from Video Memory","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.11312","noteUrl":"/review?paper=review-neurips-2026-139283","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139284","title":"Brittlebench: Quantifying LLM robustness via prompt sensitivity","authors":["Angelika Romanou","Mark Ibrahim","Candace Ross","Kerem Oktar","Chantal Shaib","Samuel J Bell","Anaelia Ovalle","Antoine Bosselut","Jesse Dodge","Koustuv Sinha","Adina Williams"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Brittlebench: Quantifying LLM robustness via prompt sensitivity","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.13285","noteUrl":"/review?paper=review-neurips-2026-139284","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139287","title":"Can We Trust Item Response Theory for AI Evaluation?","authors":["Han Jiang","Sunbeom Kwon","Jinwen Luo","Ziang Xiao","Susu Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Can We Trust Item Response Theory for AI Evaluation?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.15190","noteUrl":"/review?paper=review-neurips-2026-139287","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139288","title":"DynT2I-Eval: A Dynamic Evaluation Framework for Text-to-Image Models","authors":["Wang Juntong","Wang Jiarui","Huiyu Duan","Lewei Li","Guangtao Zhai","Xiongkuo Min"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DynT2I-Eval: A Dynamic Evaluation Framework for Text-to-Image Models","plainExplanation":"No issue was found in the checked claims. No_established_error; limited to the claim and evidence checked here.","primaryUrl":"https://arxiv.org/html/2605.06170","noteUrl":"/review?paper=review-neurips-2026-139288","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139289","title":"The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset","authors":["Richard Schwarzkopf","Fabian Immel","Alexander Blumberg","Jonas Merkert","Nils A Rack","Kaiwen Wang","Fabian Konstantinidis","Julian Truetsch","Carlos Fernandez","Annika Bätz","Kevin Rösch","Marlon Steiner","Willi Poh","Yinzhe Shen","Felix Hauser","Dominik Strutz","Jaime Villa","Gleb Stepanov","Royden Wagner","Omer Sahin Tas","Frank Bieder","Holger Caesar","Jan-Hendrik Pauls","Christoph Stiller"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.02956","noteUrl":"/review?paper=review-neurips-2026-139289","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139291","title":"ViDiC: Video Difference Captioning","authors":["Jiangtao Wu","Shihao Li","Zhaozhou Bian","Jialu Chen","Yiwen He","Runzhe Wen","An Ping","JiakaiWang","Yuanxing Zhang","Jiaheng Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ViDiC: Video Difference Captioning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.03405","noteUrl":"/review?paper=review-neurips-2026-139291","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139292","title":"A systematic evaluation of vision-language models for observational astronomical reasoning tasks","authors":["Wenke Ren","Hengxiao Guo","Wenwen Zuo","Xiaoman Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A systematic evaluation of vision-language models for observational astronomical reasoning tasks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.24589","noteUrl":"/review?paper=review-neurips-2026-139292","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139294","title":"MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods","authors":["Honglin Lin","Zheng Liu","Yun Zhu","Chonghan Qin","Juekai Lin","Xiaoran Shang","Conghui He","Bin CUI","Wentao Zhang","Lijun Wu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.21821","noteUrl":"/review?paper=review-neurips-2026-139294","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139295","title":"Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving","authors":["Aron Distelzweig","Faris Janjos","Andreas Look","Anna Rothenhäusler","Daniel Jost","Oliver Scheel","Raghu Rajan","Daphne Cornelisse","Eugene Vinitsky","Joschka Boedecker"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10034","noteUrl":"/review?paper=review-neurips-2026-139295","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139296","title":"RAIL: Rethinking Auditory Intelligence in Large Audio-Language Models with a CHC-Grounded Benchmark","authors":["Hongyu Jin","Siyi Wang","Yang Xiao","Jiaheng Dong","Shihong Tan","kaiyuan peng","Georgiana Juravle","Shanquan Chen","Gongping Huang","Hong Jia","Eun-Jung Holden","James Bailey","Ting Dang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"RAIL: Rethinking Auditory Intelligence in Large Audio-Language Models with a CHC-Grounded Benchmark","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.11260","noteUrl":"/review?paper=review-neurips-2026-139296","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139297","title":"An Embarrassingly Simple Graph Heuristic Reveals Shortcut-Solvable Benchmarks for Sequential Recommendation","authors":["Haoyu Han","Li Ma","Hanbing Wang","Bingheng Li","Daochen Zha","Chun How Tan","Huiji Gao","Xin Liu","Stephanie Moyerman","Sanjeev Katariya","Hui Liu","Jiliang Tang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"An Embarrassingly Simple Graph Heuristic Reveals Shortcut-Solvable Benchmarks for Sequential Recommendation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07125","noteUrl":"/review?paper=review-neurips-2026-139297","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139299","title":"Formal Conjectures: An Open and Evolving Benchmark for Verified Discovery in Mathematics","authors":["Moritz Firsching","Paul Lezeau","Salvatore Mercuri","Miklós Z. Horváth","Yaël Dillies","Calle Sönne","Eric Wieser","Fred Zhang","Thomas Hubert","Blaise Aguera y Arcas","Pushmeet Kohli"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Formal Conjectures: An Open and Evolving Benchmark for Verified Discovery in Mathematics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13171","noteUrl":"/review?paper=review-neurips-2026-139299","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139304","title":"Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests","authors":["Patrick Rim","Tom Long","Ekta Prashnani","Ruth Rosenholtz","Ben Boudaoud","Peter Xenopoulos","Alex Wong","Joohwan Kim","Jae-Hyun Jung"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests","plainExplanation":"The specified RL task reward depends on answer format and final-count error, so it cannot distinguish correct and incorrect intermediate decompositions with the same final count. Subtask supervision during earlier supervised fine-tuning is a separate objective.","primaryUrl":"https://arxiv.org/html/2607.22864","noteUrl":"/review?paper=review-neurips-2026-139304","collection":"neurips-2026","severity":"repairable","issueTypes":["implementation_mismatch","reporting_overstatement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139307","title":"HumanScore: Benchmarking Human Motions in Generated Videos","authors":["Tiange Xiang","Yusu Fang","Tian Tan","Narayan Schuetz","Scott Delp","Fei-Fei Li","Ehsan Adeli"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The RoM denominator can amplify unflagged severity","plainExplanation":"One hundred frames at \\(m_t=0.049\\) have no flags but summed severity \\(4.9\\), producing a negative final score under the printed transformation.","primaryUrl":"https://arxiv.org/html/2604.20157","noteUrl":"/review?paper=review-neurips-2026-139307","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139308","title":"Do Image Editing Models Understand Lighting?","authors":["Tim Küchler","Johann-Friedrich Feiden","Matthias Niessner","Carsten Rother"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The masks are corrected; the standardization domain still needs a condition","plainExplanation":"The cancellation in the proof is valid on the nonzero-MAD domain. Positive intensity scaling does not itself ensure that domain, and even a nonconstant finite array can have zero MAD.","primaryUrl":"https://arxiv.org/html/2606.26738","noteUrl":"/review?paper=review-neurips-2026-139308","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139309","title":"Ask-E: An Environment for Calibrated Question Generation","authors":["Sarah Pratt","Jae Sung Park","Scott Geng","Ali Farhadi"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Ask-E: An Environment for Calibrated Question Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.06933","noteUrl":"/review?paper=review-neurips-2026-139309","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139310","title":"MMLongCite: A Benchmark for Evaluating Faithfulness of Long-Context Vision-Language Models","authors":["Keyan Zhou","Zecheng Tang","Lingfeng Ming","Qiguang Chen","WangJie You","Guanghao Zhou","Dan Qiao","Zheming Yang","Libo Qin","Minghui Qiu","Juntao Li","Min zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"MMLongCite: A Benchmark for Evaluating Faithfulness of Long-Context Vision-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.13276","noteUrl":"/review?paper=review-neurips-2026-139310","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139312","title":"POLAR-Bench: A Diagnostic Benchmark for Privacy-Utility Trade-offs in LLM Agents","authors":["Qiaoyuan ZHENG","yiqu yang","Qi Gao","Imanol Schlag"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"POLAR-Bench: A Diagnostic Benchmark for Privacy-Utility Trade-offs in LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19127","noteUrl":"/review?paper=review-neurips-2026-139312","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139314","title":"Honeyval: A Comprehensive Evaluation Framework for LLM-powered HTTP Honeypots","authors":["Mark Vero","Fabian Kaczmarczyck","Ivan Petrov","I Shumailov","Niels Heinen","Jamie Hayes","Tianqi Fan","Luca Invernizzi","Martin Vechev"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Honeyval: A Comprehensive Evaluation Framework for LLM-powered HTTP Honeypots","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.29963","noteUrl":"/review?paper=review-neurips-2026-139314","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139315","title":"Beyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models","authors":["Sunny Rai","Jinyi Kuang","Reyhan Jamalova","Niyati Malhotra","Cristina Bicchieri","Victor Orozco-Olvera","Ana Maria Munoz Boudet","Annie Lou","Lyle Ungar","Sharath Chandra Guntuku"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Beyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models","plainExplanation":"No issue was found in the checked claims. The released initial stimuli file contains $451$ rows with distinct scenario IDs and both male and female variants. The experiment notebook constructs one request for each variant of every row. This supports the reported initial total of $902$. Applying $2\\times450$ to that initial bank uses the wrong population size.","primaryUrl":"https://osf.io/download/nz48j/","noteUrl":"/review?paper=review-neurips-2026-139315","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139316","title":"VEHBench: A Stage-Local Diagnostic Benchmark for LLM-Assisted Vibration Energy Harvester Design","authors":["Depeng Su","Yuyu Luo","Guobiao Hu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"VEHBench: A Stage-Local Diagnostic Benchmark for LLM-Assisted Vibration Energy Harvester Design","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.18181","noteUrl":"/review?paper=review-neurips-2026-139316","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139319","title":"PRISM: A Benchmark for Programmatic Spatial-Temporal Reasoning","authors":["Qiran Zhang","Yuheng Wang","Runde Yang","Lin Wu","Jingru Fan","Shu Yao","Jie Zhang","Tianle Zhou","Huatao Li","Ruijie Shi","Yihan Li","Chen Qian"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"PRISM: A Benchmark for Programmatic Spatial-Temporal Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19382","noteUrl":"/review?paper=review-neurips-2026-139319","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139320","title":"RAM-H1200: A Unified Evaluation and Dataset on Hand Radiographs for Rheumatoid Arthritis","authors":["YANG SONGXIAO","Haolin Wang","Yao Fu","Junmu Peng","Lin Fan","Hongruixuan Chen","JIAN SONG","Masayuki Ikebe","Shinya Takamaeda-Yamazaki","Masatoshi Okutomi","Tamostu Kamishima","yafei ou"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"RAM-H1200: A Unified Evaluation and Dataset on Hand Radiographs for Rheumatoid Arthritis","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.05616","noteUrl":"/review?paper=review-neurips-2026-139320","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139321","title":"From Pixels to Concepts: Do Segmentation Models Understand What They Segment?","authors":["Shuang Liang","Zeqing Wang","Yuxian LI","Xihui Liu","Han Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"From Pixels to Concepts: Do Segmentation Models Understand What They Segment?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.09591","noteUrl":"/review?paper=review-neurips-2026-139321","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139324","title":"EgoBench: An Interactive Egocentric Multimodal Benchmark for Tool-Using Agents","authors":["Yunqi Liu","Tong Niu","Zitong Wang","Zhenlong Dai","Yuqi Qing","Weiqiang Wang","Jian liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Completed-trajectory sums omit the declared benchmark stratum weights","plainExplanation":"The reporting claim uses fixed scenario–mode task-count weights, but the displayed success and efficiency formulas sum completed trajectories without stratum factors. When completion proportions differ by stratum, an unweighted sum over completed runs weights strata by their completed counts.","primaryUrl":"https://arxiv.org/html/2605.27820","noteUrl":"/review?paper=review-neurips-2026-139324","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139326","title":"JRDB-AVR: An Active Visual Reasoning Benchmark for Real-World Embodied Environments","authors":["Zhixi Cai","Fucai Ke","Sukai Huang","Maria Garcia de la Banda","Peter Stuckey","Reza Haffari","Hamid Rezatofighi"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"JRDB-AVR: An Active Visual Reasoning Benchmark for Real-World Embodied Environments","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.35032","noteUrl":"/review?paper=review-neurips-2026-139326","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139329","title":"CAREBench: Evaluating LLMs' Emotion Understanding by Assessing Cognitive Appraisal Reasoning","authors":["ZHAOYUE SUN","Hainiu Xu","Andero Uusberg","James Gross","Petr Slovak","Yulan He"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"CAREBench: Evaluating LLMs' Emotion Understanding by Assessing Cognitive Appraisal Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.17176","noteUrl":"/review?paper=review-neurips-2026-139329","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139331","title":"OmniCapBench: A Deep-Structured Evaluation Framework for Fine-Grained Audio-Visual Captioning","authors":["Zhongyu Yang","Jiale Tao","Ruitao Chen","Zuhao Yang","Yingfang Yuan","Xueliang Zhao","Auden","Kai Wang","Shuai Shao","Biao Wang","Steve Yves","Qinglin Lu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OmniCapBench: A Deep-Structured Evaluation Framework for Fine-Grained Audio-Visual Captioning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2610.12458","noteUrl":"/review?paper=review-neurips-2026-139331","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139333","title":"COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems","authors":["Hanqing Yang","Narjes Nourzad","Shiyu Chen","Marie Siew","Jingdi Chen","Carlee Joe-Wong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.00349","noteUrl":"/review?paper=review-neurips-2026-139333","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139335","title":"nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving","authors":["Zhiyu Huang","Johnson Liu","Rui Song","Zewei Zhou","Ruining Yang","Yun Zhang","Tianhui Cai","Hanyin Zhang","Mingxuan Gao","Valeria Xu","Jiali Chen","Yishan Shen","Yiluan Guo","Tony Qi","Jiaqi Ma"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31572","noteUrl":"/review?paper=review-neurips-2026-139335","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139336","title":"WorldSpeech: A Multilingual Speech Corpus from Around the World","authors":["Antonis Asonitis","Luca Lanzendörfer","Frédéric Berdoz","Roger Wattenhofer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"WorldSpeech: A Multilingual Speech Corpus from Around the World","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09167","noteUrl":"/review?paper=review-neurips-2026-139336","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139338","title":"FLARE: Full-Modality Long-Video Audiovisual Retrieval Benchmark with User-Simulated Queries","authors":["QiJie You","Hao Liang","Mingrui Chen","Bohan Zeng","Meiyi Qiang","Zhen H Wong","Wentao Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FLARE: Full-Modality Long-Video Audiovisual Retrieval Benchmark with User-Simulated Queries","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10228","noteUrl":"/review?paper=review-neurips-2026-139338","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139345","title":"AmaraSpatial-10K: A Spatially and Semantically Aligned 3D Dataset for Spatial Computing and Embodied AI","authors":["Mohammadsadegh Salehi","Alexander J Perkins","Igor P Maurell","Ashkan Dabbagh","Raymond Wong"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"AmaraSpatial-10K: A Spatially and Semantically Aligned 3D Dataset for Spatial Computing and Embodied AI","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.23018","noteUrl":"/review?paper=review-neurips-2026-139345","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139346","title":"MedVIGIL: Evaluating Trustworthy Medical VLMs Under Broken Visual Evidence","authors":["Hanqi Jiang","Junhao Chen","Yi Pan","Lifeng Chen","Weihang You","Haozhen Gong","Ruiyu Yan","Jinglei Lv","Lin Zhao","Hui Ren","Quanzheng Li","Tianming Liu","Xiang Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The grounding axis cancels its masked-refusal input","plainExplanation":"The grounding score averages clipped visual contrast \\(a-b+50\\) with masked-image accuracy \\(b\\), where \\(a,b\\in[0,100]\\). On the unclipped region, the score is \\((a-b+50+b)/2=25+a/2\\), so changing \\(b\\) has no effect.","primaryUrl":"https://arxiv.org/html/2605.07919","noteUrl":"/review?paper=review-neurips-2026-139346","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","interpretation_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139347","title":"FutureSim: Replaying World Events to Evaluate Adaptive Agents","authors":["Shashwat Goel","Nikhil Chandak","Arvindh Arun","Ameya Prabhu","Steffen Staab","Moritz Hardt","Maksym Andriushchenko","Jonas Geiping"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FutureSim: Replaying World Events to Evaluate Adaptive Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15188","noteUrl":"/review?paper=review-neurips-2026-139347","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139349","title":"DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document Understanding","authors":["Xiang Feng","Jiawei Zhou","Zhangfeng Huang","Kewei Wang","Shanshan Ye","Jinxin Hu","Zulong Chen","Yong Luo","Jing Zhang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document Understanding","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08888","noteUrl":"/review?paper=review-neurips-2026-139349","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139350","title":"quanda: An Interpretability Toolkit for Training Data Attribution Evaluation","authors":["Dilyara Bareeva","Galip Ümit Yolcu","Anna Hedström","Niklas Schmolenski","Thomas Wiegand","Wojciech Samek","Sebastian Lapuschkin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"quanda: An Interpretability Toolkit for Training Data Attribution Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2410.07158","noteUrl":"/review?paper=review-neurips-2026-139350","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139353","title":"SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks","authors":["Xiangyi Li","Yimin Liu","Wenbo Chen","Shenghan Zheng","Yifeng He","Xiaokun Chen","Bingran You","Haotian Shen","Yubo Li","Kyoung Whan Choe","Binxu Li","Shuyi Wang","Chris Kong","Xinyi Liu","Jiankai Sun","Runhui Wang","Xuanqing Liu","Jiachen Li","Xin Lan","Yuanli Wang","Xuandong Zhao","Yueqian Lin","Wengao Ye","Junwei He","Di Wang","Roey B Chaim","Songlin Li","Hanwen Xing","Yue Zhang","Yipeng Gao","Yijiang Li","Ze Ma","Liqiang Jing","Qunhong Zeng","Tianyu Wang","Kaixin Li","Yiqi Xue","Zonglin Di","Yizhuo He","Yuchen Tian","Yixuan Gao","Hejia Geng","Bo Chen","Litong Liu","Sikai Cheng","Zelin Tan","Jiajun Bao","Junwei Zhou","Shuaicheng Tong","Shuwen Xu","Penghao Jiang","Terry Yue Zhuo","Tinghan Ye","Chujun Tao","Qi Qi","Miao Li","Longtai Liao","Zhikang Dong","Chang Shi","Steven Dillmann","Han-chung Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.12670","noteUrl":"/review?paper=review-neurips-2026-139353","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139355","title":"DD-Ranking: Rethinking the Evaluation of Dataset Distillation","authors":["Zekai Li","Xinhao Zhong","Samir Khaki","Zhiyuan Liang","Yuhao Zhou","Mingjia Shi","Ziqiao Wang","Xuanlei Zhao","Wangbo Zhao","Ziheng Qin","Mengxuan Wu","Pengfei Zhou","Haonan Wang","David Junhao Zhang","Jia-Wei Liu","Shaobo Wang","Dai Liu","Linfeng Zhang","Guang Li","Kun Wang","Zheng Zhu","Zhiheng Ma","Joey Tianyi Zhou","Jiancheng Lv","Yaochu Jin","Peihao Wang","Kaipeng Zhang","Yiran Huang","Zhiwei Deng","Xindi Wu","George Cazenavette","Yuzhang Shang","Justin CUI","Jindong Gu","Qian Zheng","Hao Ye","Shuo Wang","Xiaobo Wang","Yan Yan","Angela Yao","Mike Zheng Shou","Tianlong Chen","Hakan Bilen","Baharan Mirzasoleiman","Manolis Kellis","Konstantinos N Plataniotis","Zhangyang &amp;quot;Atlas&amp;quot; Wang","Bo Zhao","Yang You","Kai Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DD-Ranking: Rethinking the Evaluation of Dataset Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.13300","noteUrl":"/review?paper=review-neurips-2026-139355","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139358","title":"Handwriting decoding as a challenging motor task for EEG Foundation Models","authors":["Srinivas Ravishankar","Ishayu Ghosh","Nora Zajzon","Teng (Simon) Fei","Virginia de Sa"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Handwriting decoding as a challenging motor task for EEG Foundation Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15698","noteUrl":"/review?paper=review-neurips-2026-139358","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139360","title":"HERO: A Heterogeneity-Aware Benchmark Library for Federated Continual Learning","authors":["Thinh Nguyen","Le-Tuan Nguyen","Minh-Duong Nguyen","Nhi Trinh","Anh T Nguyet","Dung Le","Kok-Seng Wong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"HERO: A Heterogeneity-Aware Benchmark Library for Federated Continual Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.08784","noteUrl":"/review?paper=review-neurips-2026-139360","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139361","title":"$\\chi$-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?","authors":["Haolin Chen","Deon Metelski","Leon Qi","Tao Xia","Joonyul Lee","Steve Brown","Kevin Riley","Frank Wang","T. Y Liu","Hank Capps","Zeyu Tang","Xiangchen Song","Lingjing Kong","Fan Feng","Tianyi Zeng","Zhiwei Liu","Zixian Ma","Hang Jiang","Fangli Geng","Yuan Yuan","Chenyu You","Qingsong Wen","Hua Wei","Yanjie Fu","Yue Zhao","Carl Yang","Biwei Huang","Kun Zhang","Caiming Xiong","Sanmi Koyejo","Eric Xing","Philip S Yu","Weiran Yao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"$\\chi$-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.16679","noteUrl":"/review?paper=review-neurips-2026-139361","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139362","title":"Distinguishing Performance From Competence in Evaluations of Humanlike Abstract Reasoning","authors":["Claas Beger","Ryan Yi","Shuhao Fu","Kaleda K Denton","Arseny Moskvichev","Sarah Tsai","Sivasankaran Rajamanickam","Melanie Mitchell"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Distinguishing Performance From Competence in Evaluations of Humanlike Abstract Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.02125v4","noteUrl":"/review?paper=review-neurips-2026-139362","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139366","title":"Don't Pause! Every prediction matters in a streaming video","authors":["Dibyadip Chatterjee","Zhanzhong Pang","Fadime Sener","Yale Song","Angela Yao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The matching pseudocode can credit one response more than once","plainExplanation":"The algorithm has neither an argmax selection nor response consumption.","primaryUrl":"https://arxiv.org/html/2604.24317","noteUrl":"/review?paper=review-neurips-2026-139366","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139367","title":"ELMA: Benchmarking Anaphoric Compositions for Long-Term Text-to-Motion Generation","authors":["Shun Kato","Dixin Yang","Kotaro Amaya","Yuiga Wada","Kazumi Fukuda","Takashi Shibuya","Yuki Mitsufuji","Mariko Isogawa"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"ELMA: Benchmarking Anaphoric Compositions for Long-Term Text-to-Motion Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.13549","noteUrl":"/review?paper=review-neurips-2026-139367","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139368","title":"An Open-Source Training Dataset for Foundation Models for Black-box Optimization","authors":["Aaron Klein","Herilalaina Rakotoarison","Luca Thale-Bombien","David Salinas"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"An Open-Source Training Dataset for Foundation Models for Black-box Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23417","noteUrl":"/review?paper=review-neurips-2026-139368","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139369","title":"OpenMHC: Accelerating the Science of Wearable Foundation Models","authors":["Narayan Schuetz","Yuze Bai","Lianggang Pan","Edgar Eggert","Favour Nerrise","Juan Delgado-SanMartin","Max Rosenblattl","Milana Gurbanova","Mohammad Asadi","Anders Johnson","Paul Schmiedmayer","Dennis Wang","Allan Lawrie","Daniel S Kim","Xin Liu","Akshay Paruchuri","Ehsan Adeli","Euan Ashley","Kelly W Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"OpenMHC: Accelerating the Science of Wearable Foundation Models","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2607.16235","noteUrl":"/review?paper=review-neurips-2026-139369","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139370","title":"RoutingBench: Can Agentic Routing Analysis Scale to Production Datacenter Networks?","authors":["Wenlong Ding","Zhixiong Niu","Jianan Yang","Fajun Zhang","Bo Zhang","Ling Liang","Yongqiang Xiong","Tianyin Xu","Hong Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RoutingBench: Can Agentic Routing Analysis Scale to Production Datacenter Networks?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.22848","noteUrl":"/review?paper=review-neurips-2026-139370","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139372","title":"Chem-PerturBridge: a harmonized compendium of small molecule perturbation transcriptomic effects","authors":["Artur Szałata","Olga Novitskaia","Maiia Shulman","Matthew Mella","Altynbek Zhubanchaliyev","Fabian Theis"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Chem-PerturBridge: a harmonized compendium of small molecule perturbation transcriptomic effects","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31522","noteUrl":"/review?paper=review-neurips-2026-139372","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139373","title":"EgoStream: A Diagnostic Benchmark for Streaming Episodic Memory in Egocentric Vision","authors":["Rosario Forte","Giuseppe Lando","Antonino Furnari"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EgoStream: A Diagnostic Benchmark for Streaming Episodic Memory in Egocentric Vision","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.31557","noteUrl":"/review?paper=review-neurips-2026-139373","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139374","title":"NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces","authors":["Konstantinos Kontras","Trui Osselaer","Stylianos G Mouslech","Angeliki I. Karaiskou","Guido Gagliardi","Thomas Strypsteen","Mohammad Hossein Badiei","Anku Rani","Maarten Vanmarcke","Miguel Bhagubai","Chanakya Ekbote","Jaedong Hwang","Christos Chatzichristos","Paul Liang","Maarten De Vos"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A pooled binomial law does not calibrate imbalanced balanced accuracy","plainExplanation":"The significance formula applies a total-sample binomial threshold to balanced accuracy across potentially imbalanced classes. Balanced accuracy is the mean of class recalls.","primaryUrl":"https://arxiv.org/html/2605.14698","noteUrl":"/review?paper=review-neurips-2026-139374","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139376","title":"AgentKernelArena: Benchmarking Performance and Generalization of AI Coding Agents on GPU Kernels Optimization","authors":["Sharareh Younesian","Wenwen Ouyang","sina rafati","Mehdi Rezagholizadeh","Sharon Zhou","Ji Liu","Yue Liu","Yuchen Yang","Hao Li","Ziqiong Liu","Dong Li","Vikram Appia","Zhenyu Gu","Emad Barsoum"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"AgentKernelArena: Benchmarking Performance and Generalization of AI Coding Agents on GPU Kernels Optimization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.16819","noteUrl":"/review?paper=review-neurips-2026-139376","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139377","title":"Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security","authors":["Devina Jain","David Hartmann","Chuan Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.18063","noteUrl":"/review?paper=review-neurips-2026-139377","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139378","title":"FinReasoning: A Hierarchical Benchmark for Reliable Financial Research Reporting","authors":["Yiyun Zhu","Yidong Jiang","Ziwen Xu","Yinsheng Yao","Dawei Cheng","Jinru Ding","JIE XU"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The body-strength formula drops the absolute value","plainExplanation":"A down-close day has a negative value under the printed formula, although its real-body magnitude is positive.","primaryUrl":"https://arxiv.org/html/2603.19254","noteUrl":"/review?paper=review-neurips-2026-139378","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_ambiguity","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139379","title":"Towards Characterizing Scientific Image Utility and Upgradability","authors":["Wenzhe Li","Qihang Yan","Liang Chen","Junying Wang","Farong Wen","Yijin Guo","Chunyi Li","Zicheng Zhang","Guangtao Zhai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Towards Characterizing Scientific Image Utility and Upgradability","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.03401","noteUrl":"/review?paper=review-neurips-2026-139379","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139380","title":"Toward Interactive Understanding of Code APIs","authors":["Dhananjay Ashok","Jesse Thomason","Jonathan May"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Toward Interactive Understanding of Code APIs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.32081","noteUrl":"/review?paper=review-neurips-2026-139380","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139381","title":"Partition-Aware Unlearning for Removing Spurious Correlations in Large Vision-Language Models","authors":["Aditi Sarker","Nazreen Shah","Rafi Ibn Sultan","Rhongho Jang","Dongxiao Zhu","Prashant Khanduri"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Partition-Aware Unlearning for Removing Spurious Correlations in Large Vision-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.29996","noteUrl":"/review?paper=review-neurips-2026-139381","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139382","title":"MegaStyle: Constructing Diverse and Scalable Style Dataset via Consistent Text-to-Image Style Mapping","authors":["Junyao Gao","Sibo Liu","Jiaxing Li","Yanan SUN","Yuanpeng Tu","Fei Shen","Weidong Zhang","Cairong Zhao","Jun Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MegaStyle: Constructing Diverse and Scalable Style Dataset via Consistent Text-to-Image Style Mapping","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.08364","noteUrl":"/review?paper=review-neurips-2026-139382","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139385","title":"TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders","authors":["Wei Pang","Xiangru Jian","Hehan Li","Zhixuan Yu","Jingxi Xue","Jinyang Li","Zhengyuan Dong","Xinjian Zhao","Hao Xu","Chao Zhang","Reynold Cheng","M. Tamer Özsu","Tianshu Yu"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.09323","noteUrl":"/review?paper=review-neurips-2026-139385","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139387","title":"MLLMs Fail to Refuse when Using Tools Agentically","authors":["Rikiya Takehi","Ryo Hachiuma","Shaona Ghosh","Dan Zhao","Frank Wang","Yusuke Hirota"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MLLMs Fail to Refuse when Using Tools Agentically","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.03938","noteUrl":"/review?paper=review-neurips-2026-139387","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139388","title":"Offloading Score: Measuring AI Reliance through Counterfactual Workflows","authors":["Vishakh Padmakumar","Lujain Ibrahim","Zora Wang","Jennifer Wang","Q.Vera Liao","Diyi Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Offloading Score: Measuring AI Reliance through Counterfactual Workflows","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.29392","noteUrl":"/review?paper=review-neurips-2026-139388","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139390","title":"DIBench: Benchmarking Decision Integrity of GUI-based Mobile Agents Under Deceptive Injections","authors":["Li Hu","Kanghua Mo","Yingbin Jin","Qingqing Ye","Haibo Hu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DIBench: Benchmarking Decision Integrity of GUI-based Mobile Agents Under Deceptive Injections","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06898","noteUrl":"/review?paper=review-neurips-2026-139390","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139391","title":"MedHorizon: Towards Long-context Medical Video Understanding in the Wild","authors":["Bodong Du","Bowen Liu","Yang YU","Xinpeng Ding","Zhiheng Wu","Shuning Wang","Shuo Nie","Naiming Liu","Qifeng Chen","Yangqiu Song","Xiaomeng Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MedHorizon: Towards Long-context Medical Video Understanding in the Wild","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06537","noteUrl":"/review?paper=review-neurips-2026-139391","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139393","title":"Global PIQA: Evaluating Commonsense Reasoning Across 100+ Languages and Cultures","authors":["Tyler Chang","Catherine Arnett","Abdelrahman Sadallah","Abdelrahman Eldesokey","Abeer Kashar","Daud Abolade","Abosede G Olanihun","Adam Labaran M","Adhikarimayum Meerajita Sharma","Aditi Gupta","Adril P Merin","Adwoa A Bremang","Afitab Iyigun","Afonso Simplício","Ahmed Essouaied","Aicha Chorana","Akhil Eppa","Akintunde Oladipo","Akriti Kuri","Akshay Ramesh","Aleksei Dorkin","Alfred M Kondoro","Alham Aji","Ali E Çetintaş","Allan Hanbury","Alp Niksarli","Alvaro Arroyo","Amin Bajand","Amol Khanna","Ana Chkhaidze","Ana C Condez","Anamaria Roberta Hartl","Andiswa Mkhonto","Andrew Hoblitzell","Angelos Poulis","Anirban Majumder","Anjali Chaudhary","Anna Vacalopoulou","Annika Simonsen","Anton Kovalev","Anupam Nayak","Ashvanth.S","Ayodeji Lana","Ayu Purwarianti","Bashar Alhafni","C. Benedict C. Busole","Bernard Ghanem","Bharti Nathani","Biljana Stojanovska Đurić","Ogundipe B Funmilola","Bolaotan Agbonile","Bruce Torres Fischer","Burak Tutar","Burcu Çınar","Cade Kane","Can Udomcharoenchaikit","Chadi Helwe","Cecilia Liu","Chiamaka G Nwokolo","Christopher Homan","Clément Sampebgo","Cristina España-Bonet","Cynthia Amol","Daeyeop Lee","Dan Saattrup Smart","Dana Arad","Daniil Dzenhaliou","Dasol Choi","David Liu","David Semedo","David Anugraha","OLAOYE D Oluwaseun","Deborah Popoola","Deividas Mataciunas","Dennis Asamoah Owusu","Dhyuthy Krishna Kumar","Diogo Tavares","Diogo Glória-Silva","Divyanshu Goyal","DongGeon Lee","Estefany Kelly Buchanan","Ebele N Anajemba","Ngozi G Egonu","Elias Herranen","Eman Nisar","Emile Anand","Emuobonuvie M Ajiboye","Eryawan P Yulianrifat","Esther Adenuga","Ewa Rudnicka","Faith O Itiola","Fareeha Fayyaz Sheikh","Fathima Thekkekara","Fatima Haouari","Faustin NSENGIYUMVA","Fenal A Ilasariya","Filbert A Tjiaranata","Firas Laakom","Francesca Grasso","Francesco Periti","Francesco Orabona","Gbenga K Solomon","Genta Winata","Nghia G Ngo","Gloria Udhedhe-oze","Gonçalo Vinagre","Gopi N Challagolla","Gorka Urbizu-Garmendia","Gouthami Vadithya","Guijin Son","Gyan S Mohapatra","Hafeez Ullah","Hafsteinn Einarsson","Hai Hu","Hamidreza Saffari","Hamza Zaidi","Haopeng Zhang","Harethah Abu Shairah","Hele-Andra Kuulmets","Hitesh Patel","Houda Bouamor","Hwanjo Yu","Iben N Debess","İbrahim Ethem Deveci","Ikhlasul A Hanif","Ikhyun Cho","Inês Vieira","Inês Calvo","Ismail O Daud","Yusuf I Abayomi","Itay Itzhak","Ivan Zhelyazkov","Ivan Belashkin","Ivan Spada","Jacob Brinton","Jafar Isbarov","Jaka Čibej","Jan Kocon","Jan Cuhel","Jauza A Krito","Jebish Purbey","Jennifer Za","Jennifer Mickel","Jenny Kunz","Jessica Ratovondranto","Jeyarajalingam Varsha","Jihae Jeong","Jimena T Dávalos","Jinu Lee","Joao Magalhaes","John S Yi","Jongin Kim","Joseph Chataignon","Joseph Marvin Imperial","Jubeerathan Thevakumar","Judith Land","Iuliia Alekseenko","Jiang Junchen","Jungwhan Kim","Kairit Sirts","Kamesh R","Kanda Patrick Tshinu","Kätriin Kukk","Kaustubh Ponkshe","Kavsar Huseynova","Ke He","Kenneth Enevoldsen","Kent J Alvarez","Kerem Zaman","Khalil Mrini","Kian Kyars","Komal Gour","Krishnakumar Lainitha","Krister Kruusmaa","Kunal Mukherjee","Kusum Chouhan","Laura Castro","Laura M Porrino-Moscoso","Lenny S Nzambi","Leshem Choshen","Lilja Øvrelid","Lisa Alazraki","Lovina Ehimen-Ugbede","Luheerathan Thevakumar","Luxshan Thavarasa","Mamadou K KEITA","Mansi Jangid","Marco De Santis","Marcos Garcia","Marek Suppa","Mariam D'Ciofalo","Marii Ojastu","Marium Atta","Maryam Sikander","Maximos Skandalis","Mehak Soomro","Mehmet İlteriş Bozkurt","Melaku Bayu","Menan Velayuthan","Michael Leventhal","Michał Marcińczuk","Mina Almasi","Mirna Potočnjak","Mithil Bangera","Mohammadamin Shafiei","MOHIBA ANSARI","Mridul Sharma","Mrityunjaya Indoria","Mughees U Rehman","Muhammad Ravi Shulthan Habibi","Murat B Kınay","Nada Galant","Naina S Rathore","Narada Maugin","Nathalie H Norman","Nicholas Kluge Corrêa","Nikola Ljubešić","Nirmal Thomas","Nisansa de Silva","Nisheeth Joshi","Nitish Ponkshe","Nizar Habash","Nneoma C Udeze","Noel Thomas","Noémi Ligeti-Nagy","Odunayo Ogundepo","Odunayo K Buliaminu","God'spraise Okechukwu","Olanrewaju Samuel","Olga Snissarenko","Onyinye A Chiemezie","Orkun Kınay","osman tursun","Pablo Rodríguez","Pablo Gamallo","Pedro Valente","Peter Rupnik","Philip O. EKIUGBO","Prakhar Agarwal","Prokopis Prokopidis","Quadri Yahya","Rachele Mignone","Raghav Singhal","Rahul Raja","Ram Mohan Rao Kadiyala","Raphaël Merx","Rasmus Larsen","Ratnavel Rajalakshmi","Rishav Ghosh","Romina Oji","Rui P Guerra","Rushikesh Zawar","Sa'ad N Bashir","Saeed Alzaabi","Sahil Sandeep","Sai Sandeep Kantareddy","Saleha Muzammil","Salsabila Z Pranida","Sam Buchanan","Samuel Rutunda","Sander Land","Sarah Sulollari","Sardar Ali","Kengatharaiyer Sarveswaran","Saulius Tautvaisas","Sayambhu Sen","Sayantani Banerjee","Sebastien Diarra","Sewoong Lee","Shaan Shah","Shankar Venkitachalam","Sharifa Djurabaeva","Sharon Ibejih","Shivanya S Dutta","Siddhant Gupta","Silvia P Suárez","Sina Ahmadi","Sivasuthan Sukumar","Siyuan Song","Sokratis Sofianopoulos","Sona E Simon","Sonja Benčina","Sophie Gvasalia","Sphurti K More","Spyridon Konstantinos Dragazis","Stefan Milosavljević","Stephan Kaufhold","AlRashed","Surangika Ranathunga","Taiga Someya","Taja Kuzman Pungeršek","Tal Haklay","Tasi'u Jibril","Tatsuya Aoyama","Tea Abashidze","Terenz Jomar Dela Cruz","Terra Blevins","Themistoklis Nikas","Theresa D Idoko","Tilek Chubakov","Tina Munda","Tobiloba Moses Owoeye","Tommaso Gargiani","Uni Johannesen","Vallerie A Putra","Vanya BK","Varvara Arzt","Vasily Konovalov","Vasudevan Nedumpozhimana","Viktória Ondrejová","Viktoryia Horbik","Vuk Dinić","Walelign Sewunetie","Winston Wu","Xiaojing Zhao","Yacouba Diarra","Yaniv Nikankin","Yash Mathur","Yash Bagla","Yeshil Bangera","Yixi Chen","Yiyuan Li","Yolanda Xavier","Yonatan Belinkov","Zaid Alyafeai","Zhengyang Shan","Zhi Rui Tam","Zilu Tang","Zuzana Nadova","Baber Abbasi","Stella Biderman","David Stap","Duygu Ataman","Fabian D Schmidt","Hila Gonen","Jiayi Wang","David Adelani"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Global PIQA: Evaluating Commonsense Reasoning Across 100+ Languages and Cultures","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.24081","noteUrl":"/review?paper=review-neurips-2026-139393","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139395","title":"SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits","authors":["Edward C Lin","Sahil Modi","Siva Kumar Sastry Hari","Qijing Huang","Zhifan Ye","Nestor Y. Qin","Fengzhe Zhou","Yuan Zhang","Jingquan Wang","Sana Damani","Dheeraj Peri","Ouye Xie","Aditya Kane","Michael behar","Rishabh Mehta","Vartika Singh","Vikram Sharma Mailthody","Terry Chen","Zihao Ye","Hanfeng Chen","Tianqi Chen","Vinod Grover","Wei Chen","Wei Liu","Eric Chung","Luis Ceze","Roger Bringmann","Michael Lightstone","Christos Kozyrakis","Humphrey Shi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.19173","noteUrl":"/review?paper=review-neurips-2026-139395","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139396","title":"Beyond Visual Boundaries: Rethinking Scene Segmentation for Movie RAG","authors":["Dong-Hee Kim","Seonwoo Choi","Changbeen Kim","Jungmyung Wi","Juyeon Ko","Youngju Choi","Il Hyeon Mun","Hyunwoo J. Kim","Donghyun Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Beyond Visual Boundaries: Rethinking Scene Segmentation for Movie RAG","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.28699","noteUrl":"/review?paper=review-neurips-2026-139396","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139397","title":"Are Text-to-Image Models Inductivist Turkeys? A Counterfactual Benchmark for Causal Reasoning","authors":["jiayi lei","Yuandong Pu","Xingyu Han","Rongpeng Zhu","Jing Xu","Jinyao Wang","Zijian Zhou","Bin Fu","Yuewen Cao","Yihao Liu","Hongsheng Li"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Are Text-to-Image Models Inductivist Turkeys? A Counterfactual Benchmark for Causal Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.24548","noteUrl":"/review?paper=review-neurips-2026-139397","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139398","title":"ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents","authors":["Yuanzheng Zhu","Mingyu Zhao","Chinmay Savadikar","Han Li","Shuang Xie","Alberto Castelo","Tianfu Wu","Lingyun Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.16116","noteUrl":"/review?paper=review-neurips-2026-139398","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139399","title":"Auto-Annotation with Expert-Crafted Guidelines: A Study through 3D LiDAR Detection Benchmark","authors":["Yechi Ma","Wei Hua","Shu Kong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Auto-Annotation with Expert-Crafted Guidelines: A Study through 3D LiDAR Detection Benchmark","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2506.02914","noteUrl":"/review?paper=review-neurips-2026-139399","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139401","title":"Fluxtrapolation: A benchmark on extrapolating ecosystem fluxes","authors":["Anya Fries","Jacob A Nelson","Martin Jung","Markus Reichstein","Jonas Peters"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Fluxtrapolation: A benchmark on extrapolating ecosystem fluxes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19812","noteUrl":"/review?paper=review-neurips-2026-139401","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139404","title":"Evaluating Whether LLMs Can Reliably Connect the DOTs?","authors":["Eftekhar Hossain","John Salvador","Santu Karmaker"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Evaluating Whether LLMs Can Reliably Connect the DOTs?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.38406","noteUrl":"/review?paper=review-neurips-2026-139404","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139405","title":"Probing the Trajectories of Reasoning Traces in Large Language Models","authors":["Marthe Ballon","Brecht F Verbeken","Vincent Ginis","Andres Algaba"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A discrimination gap does not determine calibration","plainExplanation":"Correctness-conditioned confidence separation measures discrimination, not calibration. The stated trends can accompany an improvement to perfect calibration.","primaryUrl":"https://arxiv.org/html/2601.23163","noteUrl":"/review?paper=review-neurips-2026-139405","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139406","title":"Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning","authors":["Lukas Twist","Helen Yannakoudakis","Jie Zhang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Reasoning-Trace Collapse: Evaluating the Loss of Explicit Reasoning During Fine-Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21127","noteUrl":"/review?paper=review-neurips-2026-139406","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139407","title":"Physical AI Smart Spaces: A Large-Scale Benchmark for Multi-Camera 3D Perception in Smart Spaces","authors":["Yuxing Wang","Yizhou Wang","Anqi Li","Shuo Wang","Sameer Satish Pusegaonkar","Haoquan Liang","Jiajun Li","Shenxin Jiang","Jianhe Yuan","Shangru Li","Tongwei Dai","Zihao Chen","David Anastasiu","Zheng (Thomas) Tang","Sujit Biswas","Xunlei Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Physical AI Smart Spaces: A Large-Scale Benchmark for Multi-Camera 3D Perception in Smart Spaces","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02580","noteUrl":"/review?paper=review-neurips-2026-139407","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139408","title":"EgoBabyVLM: Benchmarking cross-modal learning from naturalistic egocentric video data","authors":["Dongyan Lin","Phillip Rust","Angel Villar-Corrales","Alvin Tan","Mahi Luthra","Charles-Éric Saint-James","Robin Algayres","Rashel Moritz","Vanessa Stark","Surya Parimi","Jiayi Shen","Youssef Benchekroun","Yosuke Higuchi","Angelo Ortiz Tandazo","Martin Gleize","Nicolas Hamilakis","manel khentout","Sho Tsuji","Balázs Kégl","Juan Pino","Michael C Frank","Emmanuel Dupoux"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EgoBabyVLM: Benchmarking cross-modal learning from naturalistic egocentric video data","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.19130","noteUrl":"/review?paper=review-neurips-2026-139408","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139409","title":"VAANI: Capturing the language landscape for an inclusive digital India","authors":["Sujith Pulikodan","Abhayjeet Singh","Agneedh Basu","Nihar Desai","Pavan K J","Pranav D Bhat","Raghu Dharmaraju","Ritika Gupta","Sathvik Udupa","Saurabh Kumar","Sumit Sharma","Visruth Sanka","Dinesh Tewari","Harsh Dhand","Amrita Kamat","Sukhwinder Singh","Shikhar Vashishth","Partha Talukdar","Raj Acharya","Prasanta Ghosh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"VAANI: Capturing the language landscape for an inclusive digital India","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.28714","noteUrl":"/review?paper=review-neurips-2026-139409","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139411","title":"AgroOmni: A Large-Scale Multi-view Agricultural Dataset for Cross-Scale Multimodal Reasoning","authors":["Jiarui Zhang","Junqi Hu","Zurong Mai","Yang Liu","Yuhang Chen","Loushuohong","Henglian Huang","Hong Cheng","Lingyuan Zhao","Huang Jianxi","Yutong Lu","Haohuan Fu","Juepeng Zheng"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"AgroOmni: A Large-Scale Multi-view Agricultural Dataset for Cross-Scale Multimodal Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.14342","noteUrl":"/review?paper=review-neurips-2026-139411","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139412","title":"Same Words, Different Judgments: How Preferences Vary Across Modalities","authors":["Aaron Broukhim","Nadir Weibel","Eshin Jolly"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Same Words, Different Judgments: How Preferences Vary Across Modalities","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.22710","noteUrl":"/review?paper=review-neurips-2026-139412","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139415","title":"General Agent Evaluation","authors":["Elron Bandel","Asaf Yehudai","Lilach Edelstein","Yehoshua Sagron","Yotam Perlitz","Elad Venezian","Natalia Razinkov","Natan Ergas","Shlomit S Ifergan","Segev Shlomov","Michal Jacovi","Leshem Choshen","Liat Ein-Dor","Yoav Katz","Michal Shmueli-Scheuer"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"General Agent Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.22953","noteUrl":"/review?paper=review-neurips-2026-139415","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139416","title":"CS-Dialogue: A 104-Hour Dataset of Spontaneous Mandarin-English Code-Switching Dialogues for Speech Recognition","authors":["Jiaming Zhou","Yujie Guo","Shiwan Zhao","Haoqin Sun","Hui Wang","Jiabei He","Aobo Kong","Shiyao Wang","Yong Qin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CS-Dialogue: A 104-Hour Dataset of Spontaneous Mandarin-English Code-Switching Dialogues for Speech Recognition","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2502.18913","noteUrl":"/review?paper=review-neurips-2026-139416","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139418","title":"Measuring Cross-Modal Synergy: A Benchmark for VLM Explainability","authors":["Joël Roman Ky","Salah GHAMIZI","Maxime Cordy"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Measuring Cross-Modal Synergy: A Benchmark for VLM Explainability","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22168","noteUrl":"/review?paper=review-neurips-2026-139418","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139419","title":"TRIAGE: Evaluating Prospective Metacognitive Control in LLMs under Resource Constraints","authors":["Zabir Al Nazi","Shubhashis Roy Dipta"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Integer budgets prevent the claimed divergent endpoint","plainExplanation":"The integer budget remains one unit below full cost on its final open interval, so the baseline need not approach the oracle and the score need not diverge. The asserted endpoint limit is false; the scoring procedure can still be defined with its explicit clipping convention.","primaryUrl":"https://arxiv.org/html/2605.13414","noteUrl":"/review?paper=review-neurips-2026-139419","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139420","title":"RESCAST-100K: A Comprehensive Dataset for Cross-Domain Residential Load and Indoor Temperature Forecasting","authors":["Jainam Dhruva","Yousaf Raza","A.B. Siddique","Simone Silvestri"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RESCAST-100K: A Comprehensive Dataset for Cross-Domain Residential Load and Indoor Temperature Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02852","noteUrl":"/review?paper=review-neurips-2026-139420","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139421","title":"BELLS-O: Evaluating the Operational Trade-offs of LLM Supervision Systems","authors":["Leonhard Waibl","Felix Michalak","Hadrien Mariaccia"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BELLS-O: Evaluating the Operational Trade-offs of LLM Supervision Systems","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.20668","noteUrl":"/review?paper=review-neurips-2026-139421","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139423","title":"Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs","authors":["Dylan Feng","Pragya Srivastava","Anca Dragan","Cassidy Laidlaw"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.21602","noteUrl":"/review?paper=review-neurips-2026-139423","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139425","title":"Beyond Truthfulness: Evaluating Honesty in Large Language Models","authors":["Richard Ren","Arunim Agarwal","Mantas Mazeika","Cristina Menghini","Brad Kenstler","Robert Vacareanu","Mick Yang","Isabelle Barrass","Alice Gatti","Xuwang Yin","Eduardo Trevino","Matias Geralnik","Dean Lee","Summer Yue","Dan Hendrycks"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Beyond Truthfulness: Evaluating Honesty in Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2503.03750v3","noteUrl":"/review?paper=review-neurips-2026-139425","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139426","title":"Beyond Drug Discovery: The Nanotechnology Molecular Optimization (NMO) Benchmark","authors":["Matthias Blaschke","Daniel Kienzle","Zsuzsanna Koczor-Benda","Julian Lorenz","Rainer Lienhart","Fabian Pauly"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Beyond Drug Discovery: The Nanotechnology Molecular Optimization (NMO) Benchmark","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30170","noteUrl":"/review?paper=review-neurips-2026-139426","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139427","title":"COCOTree: A Dataset and Benchmark for Open Tree-Structured Visual Decomposition","authors":["Junhyub Lee","Seunghun Chae","Hyosu Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"COCOTree: A Dataset and Benchmark for Open Tree-Structured Visual Decomposition","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22068","noteUrl":"/review?paper=review-neurips-2026-139427","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139428","title":"InvestigationWorlds: An Agentic Environment for Legal Investigation","authors":["Albert Sun","Andrew J Benard","Sil Hamilton","Anna Teresita Marcelo","Yong J Kim","Carl-Leander Henneking","Rundong Hu","Yuhong Wang","David Mimno","Bishan Yang","Igor Labutov"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"InvestigationWorlds: An Agentic Environment for Legal Investigation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04129","noteUrl":"/review?paper=review-neurips-2026-139428","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139429","title":"Beyond Success Rates: Trainability and Extractability in Offline GCRL","authors":["Jan Malte Töpperwien","Aditya Mohan","Marius Lindauer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond Success Rates: Trainability and Extractability in Offline GCRL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2602.05459","noteUrl":"/review?paper=review-neurips-2026-139429","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139430","title":"STRABLE: Benchmarking Tabular Machine Learning with Strings","authors":["Gioia Blayer","Myung Jun Kim","Félix Lefebvre","Lennart Purucker","Alan Arazi","Eilam Shapira","Roi Reichart","Frank Hutter","Marine Le Morvan","David Holzmüller","Gael Varoquaux"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"STRABLE: Benchmarking Tabular Machine Learning with Strings","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12292","noteUrl":"/review?paper=review-neurips-2026-139430","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139434","title":"Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks","authors":["Benjamin Warner","Ratna S Grandhi","Max Kieffer","Aymane Ouraq","Saurav Panigrahi","Geetu Ambwani","Kunal Bagga","Nikhil Khandekar","Arya Hariharan","Nishant Mishra","Manish Saravanan","Shamus S Yang","Ahmed Essouaied","Adepoju J Moyondafoluwa","Robert Scholz","Bofeng Huang","Molly Beavers","Srishti Gureja","Anish Mahishi","Sameed Khan","Maxime Griot","Hunar Batra","Jean-Benoit Delbrouck","Siddhant Bharadwaj","Ronald Clark","Ashish Vashist","Anas Zafar","Leema K Murali","Harsh Deshpande","Ameen Patel","William Brown","Johannes Hagemann","Connor Lane","Paul Scotti","Tanishq M Abraham"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The win-rate sum includes a self-comparison","plainExplanation":"The displayed win-rate average includes a self-tie while dividing by the number of other models. It therefore exceeds the stated other-model win rate by \\(1/[2(M-1)]\\) for \\(M>1\\) models. This is an identity error in the stated win probability. With a common model set and common dataset weights, the added offset is the same for every model, so it does not change their ordering.","primaryUrl":"https://arxiv.org/html/2605.01417","noteUrl":"/review?paper=review-neurips-2026-139434","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-139435","title":"Trimming the Long-Tail of Visual World Modeling Evaluation","authors":["Bingxuan Li","Yining Hong","Cheng Qian","Hyeonjeong Ha","Jiateng Liu","Zhenhailong Wang","Yue Guo","Yunzhu Li","Heng Ji"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Trimming the Long-Tail of Visual World Modeling Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.24256","noteUrl":"/review?paper=review-neurips-2026-139435","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139436","title":"ReefNet: A Large-Scale Dataset and Benchmark for Fine-Grained Coral Reef Recognition","authors":["Abdulwahab Felemban","Yahia Battach","Faizan F Khan","Yuqian Fu","Xuhui Liu","Yesmeen Khattab","Yousef A Radwan","Xiang Li","Natalie Dunn","Susanne Bähr","Tullia I Terraneo","Fabio Marchese","Sara Beery","Burton H Jones","Francesca Benzoni","Mohamed Elhoseiny"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ReefNet: A Large-Scale Dataset and Benchmark for Fine-Grained Coral Reef Recognition","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.16822","noteUrl":"/review?paper=review-neurips-2026-139436","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139437","title":"CrafterDojo: A Suite of Foundation Models for Building Open-Ended Embodied Agents in Crafter","authors":["Junyeong Park","Hyeonseo Cho","Sungjin Ahn"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The diagonal-only loss does not enforce contrastive alignment","plainExplanation":"The printed loss equalizes diagonal similarities without comparing mismatched pairs, allowing collapsed and aligned embeddings to share a minimum.","primaryUrl":"https://arxiv.org/html/2508.13530","noteUrl":"/review?paper=review-neurips-2026-139437","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139438","title":"Evaluating Epistemic Uncertainty: Beyond OOD Detection and Active Learning","authors":["Jakub Paplhám","Vojtech Franc","Willem Waegeman","Eyke Hüllermeier"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Evaluating Epistemic Uncertainty: Beyond OOD Detection and Active Learning","plainExplanation":"Slack risk and regret budgets can make accepting every input optimal, which needs an infinite threshold for an unbounded score. A different optimal cutoff after normalization is already allowed by the appendix.","primaryUrl":"https://arxiv.org/html/2607.14817","noteUrl":"/review?paper=review-neurips-2026-139438","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139439","title":"Not Another Text Benchmark: Putting the “Visual\" Back in Visual Question Answering for Large Video Models","authors":["Rwiddhi Chakraborty","Yinong O Wang","Cheng Zhang","Fan Bai","Zhuoran Yu","Michael Kampffmeyer","Yong Jae Lee","Fernando D De la Torre","Robert Jenssen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Not Another Text Benchmark: Putting the “Visual\" Back in Visual Question Answering for Large Video Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.17112","noteUrl":"/review?paper=review-neurips-2026-139439","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139443","title":"Argus: A Cross-Regime Benchmark for the Transferability of Uncertainty Quantification in Computer-Use Agents","authors":["DIVAKE KUMAR","Sina Tayebati","Devashri Naik","Amanda Rios","Nilesh Ahuja","Omesh Tickoo","Ranganath Krishnan","Amit Trivedi"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Argus: A Cross-Regime Benchmark for the Transferability of Uncertainty Quantification in Computer-Use Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.25760v2","noteUrl":"/review?paper=review-neurips-2026-139443","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139445","title":"The FACTS Leaderboard: A Comprehensive Benchmark for Large Language Model Factuality","authors":["Aileen Cheng","Alon Jacovi","Amir Globerson","Ben Golan","Zhaobin Kuang","Chris Alberti","Connie Tao","Eyal Ben-David","Gaurav Singh Tomar","Lukas Haas","Yonatan Bitton","Adam Bloniarz","Aijun Bai","Andrew Wang","Anfal Siddiqui","Aravindan Raghuveer","Arturo Bajuelos Castillo","Aviel Atias","Chang Liu","Corey Fry","Daniel Balle","Deepanway Ghosal","Doron Kukliansky","Dror Marcus","Elena Gribovskaya","Eran Ofek","Honglei Zhuang","Itay Laish","Jan Ackermann","Lily Wang","Megan Risdal","Megan Barnes","Michael Fink","Mohamed Amin","Moran R Ambar","Natan Potikha","Nikita Gupta","Nitzan Katz","Noam Velan","Ofir Roval","Ori Ram","Polina Zablotskaia","Prathamesh Bang","Priyanka Agrawal","Rakesh Ghiya","Sanjay Ganapathy","Simon Baumgartner","Sofia Erell","Sushant Prakash","Thibault Sellam","Vikram R Sudarshan","Xuanhui Wang","Yaroslav Akulov","Yulong Yang","ZHEN YANG","Zhixin Lai","Zhongru Wu","Avinatan Hassidim","Fernando Pereira","Slav Petrov","Srinivasan Venkatachary","Tulsee Doshi","Yossi Matias","Sasha Goldshtein","Dipanjan Das"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The FACTS Leaderboard: A Comprehensive Benchmark for Large Language Model Factuality","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.10791","noteUrl":"/review?paper=review-neurips-2026-139445","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139446","title":"TorchUMM: A Unified Multimodal Model Codebase for Evaluation, Analysis, and Post-training","authors":["Yinyi Luo","Wenwen Wang","Haoyue Bai","Hongyu Zhu","Hao Chen","Pan He","Marios Savvides","Sharon Li","Jindong Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TorchUMM: A Unified Multimodal Model Codebase for Evaluation, Analysis, and Post-training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.10784","noteUrl":"/review?paper=review-neurips-2026-139446","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139450","title":"AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators","authors":["Aritra Mazumder","Shubhashis Roy Dipta","Nusrat Jahan Lia","Tanzila Khan","Kainat R Hossain","Nehaa Shri","Humayra Tasnim","Gour G Shawon","Shubhrangshu Debsarkar","Sumaiya A Rani","Al Jami Islam Anik","Debjoty Mitra","Al Nafeu Khan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Predecessor layering cannot assign finite depth to a directed cycle","plainExplanation":"Every edge increases the proposed depth by at least one. A directed cycle makes a finite assignment impossible.","primaryUrl":"https://arxiv.org/html/2605.08647","noteUrl":"/review?paper=review-neurips-2026-139450","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139451","title":"RA-ClipScore: Making Generative Model Evaluation More Interpretable","authors":["Yifan Lu","Taras Kucherenko","Hedvig Kjellstrom","Judith Bütepage"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"An attention row requires contraction before scalar aggregation","plainExplanation":"The product \\(A[j]\\hat r_j\\) retains the patch dimension of the attention row. The displayed score has an uncontracted dimension.","primaryUrl":"https://arxiv.org/html/2608.12088","noteUrl":"/review?paper=review-neurips-2026-139451","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139453","title":"3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code","authors":["Yipeng Gao","Lei Shu","Genzhi Ye","Xi Xiong","Ameesh Makadia","Meiqi Guo","Laurent Itti","Jindong Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.01057","noteUrl":"/review?paper=review-neurips-2026-139453","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139456","title":"Mirror, Mirror on the Wall: Can VLM Agents Tell Who They Are at All?","authors":["Filippo Ziliotto","Ciro Beneduce","Luciano Serafini","Bruno Lepri","Massimiliano Luca","Tommaso Campari"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mirror, Mirror on the Wall: Can VLM Agents Tell Who They Are at All?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08816","noteUrl":"/review?paper=review-neurips-2026-139456","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139457","title":"Principia: Relational Physics Tests for Video Models","authors":["Varun V Thozhiyoor","Shivam Tripathi","Venkatesh Babu Radhakrishnan","Anand Bhattad"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The first turning point is not the equilibrium extension","plainExplanation":"The appendix measures a release turning point, whose distance includes an additive natural-length term, rather than equilibrium extension proportional to mass. The stated mass proportionality does not describe the appendix's first-turning-point measurement.","primaryUrl":"https://arxiv.org/html/2609.04200","noteUrl":"/review?paper=review-neurips-2026-139457","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139458","title":"SimplexUQ: An Evaluation Framework and Benchmark for Conformal Uncertainty on Simplex-Valued Predictions","authors":["Liang You","Hengyu Shi","Dongwen Ou"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"SimplexUQ: An Evaluation Framework and Benchmark for Conformal Uncertainty on Simplex-Valued Predictions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.00523","noteUrl":"/review?paper=review-neurips-2026-139458","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139459","title":"LibEvoBench: Probing Temporal Knowledge Stratification in Code Generation Models","authors":["Daniele Cipollone","Sergey Titov","Maliheh Izadi","Egor Bogomolov","Arie v Deursen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"LibEvoBench: Probing Temporal Knowledge Stratification in Code Generation Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.25402","noteUrl":"/review?paper=review-neurips-2026-139459","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139460","title":"SayNext-Bench: Why Do LLMs Struggle with Next-Utterance Anticipation?","authors":["Yueyi Yang","Haotian Liu","Fang Kang","Mengqi Zhang","Zheng Lian","Hao Tang","Haoyu Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SayNext-Bench: Why Do LLMs Struggle with Next-Utterance Anticipation?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.00327","noteUrl":"/review?paper=review-neurips-2026-139460","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139461","title":"How to Interpret Agent Behavior","authors":["Sophia Gao","Kaiser Sun","Jen-Tse Huang","Katherine Van Koevering","Sijie Ji","Heyuan Huang","Weiyan Shi","Zhuoran Lu","Ziang Xiao","Daniel Khashabi","Mark Dredze"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"How to Interpret Agent Behavior","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13625","noteUrl":"/review?paper=review-neurips-2026-139461","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139462","title":"Beyond Chamfer Distance: Granular Order-aware Evaluation Metric For Online Mapping","authors":["Chouaib Bencheikh Lehocine","Adam Lilja","Junsheng Fu","Lars Hammarstrand"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The empty-set normalization needs a convention","plainExplanation":"For two empty sets, all sums in Equation (9) vanish, giving \\(d=R_X=R_Y=0\\).","primaryUrl":"https://arxiv.org/html/2605.22578","noteUrl":"/review?paper=review-neurips-2026-139462","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139464","title":"More Is Not More: What Matters for Diversity in LLM Opinions？","authors":["Qiyang Yao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Paired sign counts are not signed-rank biserial effect size","plainExplanation":"Paired sign counts ignore magnitudes, while signed-rank biserial weights by absolute-difference ranks. They are different effect sizes.","primaryUrl":"https://arxiv.org/html/2607.20429","noteUrl":"/review?paper=review-neurips-2026-139464","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139476","title":"PrivacySIM: Evaluating LLM Simulation of User Privacy Behavior","authors":["James Flemings","Murali Annavaram"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PrivacySIM: Evaluating LLM Simulation of User Privacy Behavior","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12147","noteUrl":"/review?paper=review-neurips-2026-139476","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139478","title":"Can Agents Price a Reaction? Evaluating LLMs on Chemical Cost Reasoning","authors":["Yuyang Wu","Yue Huang","Shuaike Shen","Xujian Wang","Shuhao Zhang","Qiyao Xue","Weichen Liu","Runtian Gao","Jian Ma","Xiangliang Zhang","Olexandr Isayev"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Can Agents Price a Reaction? Evaluating LLMs on Chemical Cost Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07251","noteUrl":"/review?paper=review-neurips-2026-139478","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139479","title":"Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth","authors":["Viktor Kocur","Sithu Aung","Gabrielle Flood","Yaqing Ding","Lukas Bujnak","Torsten Sattler","Zuzana Kukelova"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19797","noteUrl":"/review?paper=review-neurips-2026-139479","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139480","title":"Block-R1: Rethinking the Role of Block Size in Multi-domain Reinforcement Learning for Diffusion Large Language Models","authors":["Yan Jiang","Ruihong Qiu","Zi Huang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Block-R1: Rethinking the Role of Block Size in Multi-domain Reinforcement Learning for Diffusion Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11726","noteUrl":"/review?paper=review-neurips-2026-139480","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139481","title":"GeoSym127K: Scalable Symbolically-verifiable Synthesis for Multimodal Geometric Reasoning","authors":["Jinhao Jing","Zheng Ma","Jinwei Liang","Qiannian Zhao","Shuang Chen","Jing Yang","Prayag Tiwari","Jingjing Bai","Benyou Wang","Por L Yee","Zhan Su","Lewei Lu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"GeoSym127K: Scalable Symbolically-verifiable Synthesis for Multimodal Geometric Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.16371","noteUrl":"/review?paper=review-neurips-2026-139481","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139483","title":"E-GEO: A Testbed for Generative Engine Optimization in E-Commerce","authors":["Puneet S Bagga","Vivek Farias","Tamar Korkotashvili","Tianyi Peng","Yuhang Wu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"E-GEO: A Testbed for Generative Engine Optimization in E-Commerce","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2511.20867","noteUrl":"/review?paper=review-neurips-2026-139483","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139484","title":"The Smart Buildings Control Suite: A Diverse Open Source Benchmark to Evaluate and Scale HVAC Control Policies for Sustainability","authors":["Judah Goldfeder","Victoria Dean","Zixin Jiang","Xuezheng Wang","Bing Dong","Hod Lipson","John Sipple"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Cell-specific face conductivities fail to conserve interface energy","plainExplanation":"The heat stencil uses each cell’s own conductivity for the same shared interface, without imposing a reciprocal face conductance. Take two closed cells of unit capacity and time step, conductivities \\(1,2\\), initial temperatures \\((0,1)\\), and no sources.","primaryUrl":"https://arxiv.org/html/2410.03756","noteUrl":"/review?paper=review-neurips-2026-139484","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139486","title":"JuICE: A Benchmark for Evaluating LLM-Judge in Identifying Cultural Errors","authors":["Jiho Jin","Junho Myung","Juhyun Oh","Junyeong Park","Rifki A Putri","Sunipa Dev","Vinodkumar Prabhakaran","Alice Oh"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"JuICE: A Benchmark for Evaluating LLM-Judge in Identifying Cultural Errors","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.26955","noteUrl":"/review?paper=review-neurips-2026-139486","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139488","title":"Thinking in Pictures: A Systematic Benchmark for Reasoning-driven Image Generation","authors":["Yutong Liu","Nan Huang","Xu Cao","James Rehg"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Thinking in Pictures: A Systematic Benchmark for Reasoning-driven Image Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.02864","noteUrl":"/review?paper=review-neurips-2026-139488","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139489","title":"Castle-in-the-Air: Probing the Foundational Visual Deficits of MLLMs via Bottom-Up Cognitive Factors","authors":["Jen-Tse Huang","Dasen Dai","Jen-Yuan Huang","Youliang Yuan","Xiaoyuan Liu","Wenxuan Wang","Wenxiang Jiao","Pinjia He","Zhaopeng Tu","Haodong Duan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Castle-in-the-Air: Probing the Foundational Visual Deficits of MLLMs via Bottom-Up Cognitive Factors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2502.16435v4","noteUrl":"/review?paper=review-neurips-2026-139489","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139490","title":"Beyond IID: How General Are Tabular Foundation Models, Really?","authors":["Lennart Purucker","Andrej Tschalzev","Nick Erickson","David Holzmüller","Alan Arazi","Alexander Pfefferle","Mustafa Tajjar","Gioia Blayer","Gael Varoquaux","Frank Hutter"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The zero split-budget endpoint contradicts Lite","plainExplanation":"Failing to reject a difference null can include uniformly worse methods; it does not certify equivalence or noninferiority. The statistical inference is unsupported; the underlying observed scores and comparisons are not refuted.","primaryUrl":"https://arxiv.org/html/2606.30410","noteUrl":"/review?paper=review-neurips-2026-139490","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139491","title":"Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials","authors":["Amir Masoud Nourollah","Irtaza Khalid","Stefano Leoni","Steven Schockaert"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08988","noteUrl":"/review?paper=review-neurips-2026-139491","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139494","title":"GenScale: A Benchmark for Relative Object Scale in Image Generation and Editing","authors":["Lingxiao Li","Max Whitton","Ledell Wu","Boqing Gong"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"GenScale: A Benchmark for Relative Object Scale in Image Generation and Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.00525","noteUrl":"/review?paper=review-neurips-2026-139494","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139495","title":"Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets?","authors":["Trisha Mittal","Akshay Mehra","Joshua Kimball"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.18209","noteUrl":"/review?paper=review-neurips-2026-139495","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139496","title":"DRAGON : A Benchmark for Evidence-Grounded Visual Reasoning over Diagrams","authors":["Anirudh Iyengar Kaniyar Narayana Iyengar","Tampu Ravi Kumar","Gaurav Najpande","Manan Suri","Dinesh Manocha","Puneet Mathur","Vivek Gupta"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"DRAGON : A Benchmark for Evidence-Grounded Visual Reasoning over Diagrams","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.25231","noteUrl":"/review?paper=review-neurips-2026-139496","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139497","title":"L2-Bench: An Evaluation Benchmark for Measuring LLM Capabilities in Second Language Education","authors":["James Edgell","Wm. Matthew Kennedy","Ben Knight","Danielle Carvalho","Martin Ku","Isaac J Pattis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The SEM formula counts within-task response noise twice","plainExplanation":"The displayed SEM adds average within-task variance divided by run count to the variance of the observed task-level means. For \\(X_i\\), the mean of \\(k\\) runs, total variance already gives \\(\\operatorname{Var}(X_i)=\\operatorname{Var}(\\mu_i)+\\mathbb E\\sigma_i^2/k\\).","primaryUrl":"https://arxiv.org/html/2607.08842","noteUrl":"/review?paper=review-neurips-2026-139497","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139498","title":"ASPI: Seeking Ambiguity Clarification Amplifies Prompt Injection Vulnerability in LLM Agents","authors":["Udari Madhushani Sehwag","Zhengyang Shan","Heming Liu","Dileepa Lakshan","Joseph Brandifino","Max Fenkell"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ASPI: Seeking Ambiguity Clarification Amplifies Prompt Injection Vulnerability in LLM Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.17324","noteUrl":"/review?paper=review-neurips-2026-139498","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139500","title":"Offline Preference-Based Trajectory Evaluation","authors":["Fernando Diaz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Offline Preference-Based Trajectory Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.17541","noteUrl":"/review?paper=review-neurips-2026-139500","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139501","title":"SCDBench: A Benchmark for LLM-Based Smart Contract Decompilers","authors":["Kaihua Qin","Dawn Song","Arthur Gervais"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SCDBench: A Benchmark for LLM-Based Smart Contract Decompilers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29059","noteUrl":"/review?paper=review-neurips-2026-139501","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139502","title":"MPDocBench-Parse: Benchmarking Practical Multi-page Document Parsing","authors":["Bangbang Zhou","Hangdi Xing","Yifan Chen","Jianjun Xu","Qi Zheng","Feiyu Gao","Zhibo Yang","Shuai Bai","Ming Yan","Jieping Ye","Hongtao Xie"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MPDocBench-Parse: Benchmarking Practical Multi-page Document Parsing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22100","noteUrl":"/review?paper=review-neurips-2026-139502","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139503","title":"MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays","authors":["Emre Hayir","Lorin Crawford","Alex X Lu"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"A rejection fraction is not false discovery rate","plainExplanation":"The reported quantity is \\(R/N\\), whereas FDR is \\(\\mathbb E[V/\\max(R,1)]\\).","primaryUrl":"https://arxiv.org/pdf/2605.15383","noteUrl":"/review?paper=review-neurips-2026-139503","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139504","title":"Realism VS Accuracy: Event Sequence Forecasting from a Generative Modeling Perspective","authors":["Dmitry Osin","Egor Surkov","Petr Mokrov","Igor Udovichenko","Ivan Karpukhin","Evgeny Burnaev","Egor Shvetsov","Aleksandr Korotin"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Realism VS Accuracy: Event Sequence Forecasting from a Generative Modeling Perspective","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.12586","noteUrl":"/review?paper=review-neurips-2026-139504","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139505","title":"RISE-Video: Can Video Generators Decode Implicit World Rules?","authors":["Mingxin Liu","Shuran Ma","Shibei Meng","Xiangyu Zhao","Zicheng Zhang","Shaofeng Zhang","Zhihang Zhong","Peixian Chen","Haoyu Cao","Xing Sun","Haodong Duan","Xue Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RISE-Video: Can Video Generators Decode Implicit World Rules?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.05986","noteUrl":"/review?paper=review-neurips-2026-139505","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139506","title":"MedFlowBench: Auditing Medical Agents in Full-Study Workflows","authors":["Weixiang Shen","Chengzhi Shen","Che Liu","Junde Wu","Jiayuan Zhu","Xiao Han","Zongyue Li","Jingpei Wu","Min Xu","Daguang Xu","Yueming Jin","Benedikt Wiestler","Daniel Rueckert","Jiazhen Pan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MedFlowBench: Auditing Medical Agents in Full-Study Workflows","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.24649","noteUrl":"/review?paper=review-neurips-2026-139506","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139507","title":"gfnx: Fast and Scalable Library for Generative Flow Networks in JAX","authors":["Daniil Tiapkin","Artem Agarkov","Nikita Morozov","Ian Maksimov","Askar Tsyganov","Timofei Gritsaev","Sergey Samsonov"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The Metropolis-Hastings proposal ratio is reversed","plainExplanation":"The acceptance rule uses the proposal ratio in the wrong direction and can violate detailed balance even for a uniform one-spin target.","primaryUrl":"https://arxiv.org/html/2511.16592","noteUrl":"/review?paper=review-neurips-2026-139507","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139510","title":"When Stored Evidence Stops Being Usable: Scale-Conditioned Evaluation of Agent Memory","authors":["Jiaqi Shao","Yiyi Lu","Yunzhen Zhang","Bing Luo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"When Stored Evidence Stops Being Usable: Scale-Conditioned Evaluation of Agent Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07313","noteUrl":"/review?paper=review-neurips-2026-139510","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139512","title":"GeoFidelity-Bench: Evaluating Block-Conditioned Geographic Fidelity of Street-View Generation","authors":["Kaizhen Tan"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"GeoFidelity-Bench: Evaluating Block-Conditioned Geographic Fidelity of Street-View Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.23669","noteUrl":"/review?paper=review-neurips-2026-139512","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139513","title":"Agent$^2$ RL-Bench: Can LLM Agents Engineer Agentic RL Post-Training?","authors":["Wanyi Chen","Xiao Yang","Xu Yang","Tianming Sha","Qizheng Li","Zhuo Wang","Bowen Xian","Fang Kong","Weiqing Liu","Jiang Bian"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Agent$^2$ RL-Bench: Can LLM Agents Engineer Agentic RL Post-Training?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.10547","noteUrl":"/review?paper=review-neurips-2026-139513","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139514","title":"Text-to-CAD Evaluation with CADTests","authors":["Dimitrios Mallis","Marco Wang","Ahmet Karadeniz","Elisa Ricci","Anis Kacem","Djamila Aouada"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Text-to-CAD Evaluation with CADTests","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07807","noteUrl":"/review?paper=review-neurips-2026-139514","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139515","title":"CORAL: A Benchmark for Structure-aware and Brain-wide Neuron Reconstruction in Light Microscopy","authors":["Zekang Yang","Jiamin Li","Zhenghua Li","Jiaqi Fan","Zengcai V Guo","Xiaolin Hu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"CORAL: A Benchmark for Structure-aware and Brain-wide Neuron Reconstruction in Light Microscopy","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.30768","noteUrl":"/review?paper=review-neurips-2026-139515","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139516","title":"AssayBench: An Assay-Level Virtual Cell Benchmark for LLMs and Agents","authors":["Edward De Brouwer","Carl Edwards","Alexander P Wu","Jenna L Collier","Graham Heimberg","Xiner Li","Meena Subramaniam","Ehsan Hajiramezanali","David Richmond","Jan-Christian Huetter","Sara Mostafavi","Gabriele Scalia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Clipping destroys the centered random baseline","plainExplanation":"The appendix clips a centered score at zero, removing its negative range and changing its random-ranking expectation. The two score definitions differ; the zero-expectation claim holds before clipping.","primaryUrl":"https://arxiv.org/html/2605.10876","noteUrl":"/review?paper=review-neurips-2026-139516","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139517","title":"Agents' Last Exam","authors":["Yiyou Sun","Xinyang Han","Weichen Zhang","Yuanbo Pang","Tianyu Wang","Yuhan Cao","Yixiao Huang","Christopher Duroiu","Haoyun Zhang","Jeffrey Lin","Weishu Zhang","Tianren Zeng","Ying Yan","Bo Liu","Hanson Wen","Mingyang Xu","Xiaoyuan Liu","Zimeng Chen","Weiyan Shi","Amanda Dsouza","Vincent Chen","Yushan Li","Wenxi Deng","Huiqi Wang","Justin Xu","Tao Sun","Zhun Wang","Chris Liu","Yafei Cheng","Rongwang Hu","Aras Bacho","Shengcao Cao","Zengyi Qin","Yixiong Chen","Hengduan Fan","Hao Liu","Lin Zeng","Shashank M Bharadwaj","Litian Gong","Maojia Song","Ruheng Wang","Zongzheng Zhang","Jianhong Tu","Zhonghua Wang","Zheng Zhang","Zijiao Chen","Yanqiong Jiang","Zhendong Li","Bohan Lyu","Ma Chang","Peiran Xu","Benran Zhang","Shangding Gu","Haoyue Hua","Haoyang Li","Wanzhe Liao","Chengzhi Liu","Junbo Peng","Zechen Xu","Bo Chen","Jiayi Cheng","Yi Jiang","Maggie Kuang","Yuan Li","Youbang Pan","Ziyan Rao","Alexander Schubert","Ivan Shen","Vincent Siu","Xiatao Sun","Xiaopan Zhang","Yuchen Zhu","Ishaan S Chandok","Lei Ding","Jingxuan Fan","Andrew Glover","Jiaming Hu","Yiran HU","Wenbo Huang","Haoran Jin","Lukas Kim","Ming Liu","Alireza Rafiei","Kunyang Sun","Ting Sun","Yingze Wang","Yixin Wang","Hanwen Xing","Sihan Xu","Yuzheng Xu","Zhongxing Xu","Zhiling Yan","Boqin Yuan","Ruiqi Zhang","Yifan Zhang","Zibo Zhao","Haoyue Bai","Carlo Bosio","Joe Cavanagh","Tianxing Chen","Yipu Chen","CHENYU ZHU","Chen Dai","Stefano De Castro","Yunfu Deng","Kaustubh Dhole","Jiayuan Ding","Chenchen Du","Zhehang Du","Hao Fan","Run-Ze Fan","Hengyu Fu","Shi Gu","Yifan Gu","Baihe Huang","Baixiang Huang","Ran Jin","Xin Lan","Joseph Lee","Deren Lei","Chenyu Li","Daofeng Li","Jingyan Li","Yi Li","Yuangang Li","Zhixu Li","Yinsheng Li","Wenyu Liang","Longtai Liao","Kevin Qinghong Lin","Zeyi (Andy) Liu","Jiaming Liu","Kaiyuan Liu","Xuan Liu","Pan Lu","Yicheng Lyu","Qiuyang Mang","Kyle Montgomery","Ruoxi Ning","Jorin Overwiening","Xu Pan","Core Francisco Park","Justin Purnomo","Scott Rankin","Bixuan Ren","HaoYang Shang","Hejia Shen","JIAWEI SHEN","Shi Qiu","Tianneng Shi","Jonah So","Vladislav Susoy","Haocheng Wang","Jialu Wang","Wei Wang","Xinyu J Wang","Zehao Wang","Ruoxi Wu","Dehao Wu","Fangyu Wu","Mengyuan Wu","Yu Wu","Yuchen Wu","Yuhao Wu","Qingpo Wuwu","Weihang Xiao","Yongyi Xiong","Fan Xu","Ruiling Xu","Mingxuan Yan","Benjamin Yang","Jirong Yang","Xiaoli Yang","Yushi Yang","Haoran Ye","Xiaohu Yu","Zhengming Yu","Chenlong Zhang","Chi Zhang","Hanning Zhang","Hanwen Zhang","Junge Zhang","Kunpeng Zhang","Song Zhang","Wenjin Zhang","Wenshuo Zhang","Qijian Zhao","Yimin Zhao","Yuhaohua Zheng","Liwei Zhou","Tianyue Zhou","Sichen Zhu","Yan Zhu","Yishu Zhu","Jerry Zuo","Chonghao Cai","Helena Casademunt","Cheng Cheng","Nawen Deng","Rao Fu","Yifan Han","He Ren","Zhenyu He","Qiao Jin","Langlang Li","Yuetai Li","Lu Lu","Luqing Zhou","Subhabrata Mukherjee","Yunqi Ouyang","Yin Ren","Dawei Shi","Haoran Wu","Zhiyue Wu","Hannah Yao","Zhuoran Yi","Jenny Yu","Rhea Zhan","Hang Zhou","Shiqiao Zhu","Junfan Zhu","Zhen Dong","Ali Emami","Yanjun Gao","Haimin Hu","Erika J Schneider","Zhenglu Li","Jiachen Li","Lihui Liu","Murphy Niu","Yi Shao","Wenqi Shi","Lichao Sun","Jianxin Sun","Chenguang Wang","Xuhai &amp;quot;Orson&amp;quot; Xu","Huaxiu Yao","Aylin Caliskan","Jing Huang","Yang Liu","Mark W Mueller","Russell Poldrack","Sanjiv R Das","Zhiyong Lu","Radha Poovendran","Somayeh Sojoudi","Costas J Spanos","Molei Tao","Mikko Tolonen","Ting Wang","Alan Yuille","Patrick Bryant","Carl Boettiger","Arvind Rao","Tapio Schneider","Georgios N Yannakakis","Laure Zanna","Ida Sim","Tarek Zohdi","Jack Gallant","Teresa Head-Gordon","Dawn Song"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Agents' Last Exam","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.05405","noteUrl":"/review?paper=review-neurips-2026-139517","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139519","title":"MVVBench: Benchmarking 4D Reasoning in Vision-Language Models","authors":["Hyungjin Chung","Byeongjun Park","Joonseok Lee","Hojun Kim","Jaeho Choi","Byung-Hoon Kim"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"MVVBench: Benchmarking 4D Reasoning in Vision-Language Models","plainExplanation":"","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2609.30952","noteUrl":"/review?paper=review-neurips-2026-139519","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139522","title":"SC$^3$: A Multi-Solvent Solubility Challenge and Benchmark","authors":["Vansh Ramani","Tarak Karmakar","Sayan Ranu","Dhairya Kuchhal","Har Ashish Arora","Sergei V Tatarin","Lev Krasnov"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Equal solvent weighting does not cancel residual location bias","plainExplanation":"PS-RMSE gives each solvent equal weight but retains the squared conditional mean prediction error. The mathematical correction concerns this location-cancellation interpretation; the metric remains a valid equal-weighted prediction-error measure.","primaryUrl":"https://arxiv.org/html/2606.07656","noteUrl":"/review?paper=review-neurips-2026-139522","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139524","title":"TALES: Text-Adventure Learning Environment Suite","authors":["Christopher Z. Cui","Xingdi Yuan","Ziang Xiao","Prithviraj Ammanabrolu","Marc-Alexandre Côté"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TALES: Text-Adventure Learning Environment Suite","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2504.14128","noteUrl":"/review?paper=review-neurips-2026-139524","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139525","title":"Models That Know How Evaluations Are Designed Score Safer","authors":["Katharina Deckenbach","Haritz Puerto","Jonas Geiping","Sahar Abdelnabi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Models That Know How Evaluations Are Designed Score Safer","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.28591","noteUrl":"/review?paper=review-neurips-2026-139525","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139526","title":"CausalCompass: Evaluating the Robustness of  Time-Series Causal Discovery in Misspecified Scenarios","authors":["Huiyang Yi","Xiaojian Shen","Yonggang Wu","Duxin Chen","He Wang","Wenwu Yu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"CausalCompass: Evaluating the Robustness of  Time-Series Causal Discovery in Misspecified Scenarios","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.07915","noteUrl":"/review?paper=review-neurips-2026-139526","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139527","title":"EvoCodeBench: Evaluating Coding Agents in Multi-Turn Iterative Interactions","authors":["Haiyang Shen","Xuanzhong Chen","Wendong XU","Yun Ma","Liang Chen","Kuan Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EvoCodeBench: Evaluating Coding Agents in Multi-Turn Iterative Interactions","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.24110","noteUrl":"/review?paper=review-neurips-2026-139527","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139530","title":"EMERGE: A Benchmark for Updating Knowledge Graphs with Emerging Textual Knowledge","authors":["Klim Zaporojets","Daniel Daza","Edoardo Barba","Ira Assent","Roberto Navigli","Paul Groth"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"EMERGE: A Benchmark for Updating Knowledge Graphs with Emerging Textual Knowledge","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2507.03617","noteUrl":"/review?paper=review-neurips-2026-139530","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139532","title":"NORA: Evaluating Grounded Reasonableness in Visual First-person Normative Action Reasoning","authors":["Sichao Li","Sai Ma","Zhuang Li","Daniel Kilov","Secil Yanik Guyot","Seth Lazar"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"NORA: Evaluating Grounded Reasonableness in Visual First-person Normative Action Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.04806","noteUrl":"/review?paper=review-neurips-2026-139532","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139533","title":"GlucoFM-Bench: Benchmarking Time-Series Foundation Models for Blood Glucose Forecasting","authors":["Baiying Lu","Zhaohui Liang","Ryan Pontius","Shengpu Tang","Temiloluwa Prioleau"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"GlucoFM-Bench: Benchmarking Time-Series Foundation Models for Blood Glucose Forecasting","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.06881","noteUrl":"/review?paper=review-neurips-2026-139533","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139534","title":"MathlibPR: Pull Request Merge-Readiness Benchmark for Formal Mathematical Libraries","authors":["Zixuan Xie","Xinyu Liu","Shangtong Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MathlibPR: Pull Request Merge-Readiness Benchmark for Formal Mathematical Libraries","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07147","noteUrl":"/review?paper=review-neurips-2026-139534","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139535","title":"InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection","authors":["Łukasz Marcjan","Dorota Wilk-Kołodziejczyk","Roberto Corizzo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.15032","noteUrl":"/review?paper=review-neurips-2026-139535","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139537","title":"FineVision: Open Data Is All You Need","authors":["Luis Wiedmann","Orr Zohar","Amir Mahla","Xiaohan Wang","Rui Li","Thibaud Frere","Leandro Von Werra","Aritra Roy Gosthipaty","Andrés Marafioti"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FineVision: Open Data Is All You Need","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.17269","noteUrl":"/review?paper=review-neurips-2026-139537","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139538","title":"ThousandWorlds: A benchmark for climate emulation of potentially habitable exoplanets","authors":["Edward Stevenson","Mei T Mak","Eric Wolf","Denis E Sergeev","Tobi Hammond","Nathan Mayne","Miles Cranmer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ThousandWorlds: A benchmark for climate emulation of potentially habitable exoplanets","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.18338","noteUrl":"/review?paper=review-neurips-2026-139538","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139539","title":"STRIDE: Automated Evaluation of Text-to-Trajectory Alignment across Diverse Contexts","authors":["Wanchun Ni","Tao Qi","Leonel Aguilar","Jiugeng Sun","Marlene Wagner","Verena Zimmermann","Mennatallah El-Assady"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"STRIDE: Automated Evaluation of Text-to-Trajectory Alignment across Diverse Contexts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34799","noteUrl":"/review?paper=review-neurips-2026-139539","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139543","title":"A Matched-Budget Audit Framework for Recaptioned Image-Text Supervision Distributions","authors":["Giyeong Oh","Junghun Park","Yuhan Bae","Youngjae Yu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"A Matched-Budget Audit Framework for Recaptioned Image-Text Supervision Distributions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.00952","noteUrl":"/review?paper=review-neurips-2026-139543","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139544","title":"K9-Bench: Evaluating Multimodal LLMs on Canine-Centric Videos","authors":["Khush Attarde","Yusuf Ali","Megha Thukral","Divye Bhutani","Thomas Ploetz","Zsolt Kira"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"K9-Bench: Evaluating Multimodal LLMs on Canine-Centric Videos","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.02680","noteUrl":"/review?paper=review-neurips-2026-139544","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139546","title":"LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training","authors":["Andreas Hochlehnert","Marianna Nezhurina","Mehdi Cherti","Andrej Radonjic","Thaddäus Wiedemer","Christoph Schuhmann","Romain Beaumont","Wieland Brendel","Bernhard Schölkopf","A. Sophia Koepke","Jenia Jitsev","Matthias Bethge"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training","plainExplanation":"","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2608.24845","noteUrl":"/review?paper=review-neurips-2026-139546","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139547","title":"fev-bench: A Realistic Benchmark for Time Series Forecasting","authors":["Oleksandr Shchur","Abdul Fatir Ansari","Ali Caner Turkmen","Lorenzo Stella","Nick Erickson","Pablo A Guerron","Michael Bohlke-Schneider","Yuyang (Bernie) Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"fev-bench: A Realistic Benchmark for Time Series Forecasting","plainExplanation":"Clipping relative errors can reverse the ranking of two forecasting models when the reference baseline changes. Geometric-mean baseline invariance holds when clipping is inactive, but does not justify the unrestricted claim for the implemented score.","primaryUrl":"https://arxiv.org/html/2509.26468","noteUrl":"/review?paper=review-neurips-2026-139547","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139548","title":"MedEvoEval: Evaluating Continual Evolution of Doctor Agents through Simulated Clinical Episodes","authors":["Hui Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MedEvoEval: Evaluating Continual Evolution of Doctor Agents through Simulated Clinical Episodes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.28900","noteUrl":"/review?paper=review-neurips-2026-139548","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139550","title":"DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs","authors":["Jiazhen Pan","Weixiang Shen","Jun Li","Julian Canisius","Felix Bitzer","Paula Roßmüller","Jiancheng Yang","Virginie Kreutzinger","Daniel Rueckert","Benedikt Wiestler"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.23629","noteUrl":"/review?paper=review-neurips-2026-139550","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139551","title":"EHRNote-ChatQA: A Benchmark for Evidence-Grounded Multi-Turn Clinical Question Answering over Longitudinal Discharge Summaries","authors":["Jiyoun Kim","Muhan Yeo","Eunhye Jang","Jeewon Yang","Hangyul Yoon","Su Ji Lee","Hee J Han","Hee-Jae Jung","Doyun Kwon","Jun y Lee","Jaehun Lee","Jung-Oh Lee","Sunjun Kweon","Jong-Hak Moon","Daseul Kim","Minjae Cho","Edward Choi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EHRNote-ChatQA: A Benchmark for Evidence-Grounded Multi-Turn Clinical Question Answering over Longitudinal Discharge Summaries","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.15735","noteUrl":"/review?paper=review-neurips-2026-139551","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139552","title":"Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models","authors":["Haoxiang Wang","Da Yu","Huishuai Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Repeated-trial Fisher information must retain the trial count","plainExplanation":"The calibration appendix computes information and posterior uncertainty for binomial cells with \\(N\\) repeated trials. For \\(K\\sim\\operatorname{Binomial}(N,p)\\), \\(p=\\sigma(\\theta-\\beta)\\), the score is \\(K-Np\\) and negative second derivative is \\(Np(1-p)\\).","primaryUrl":"https://arxiv.org/html/2605.06213","noteUrl":"/review?paper=review-neurips-2026-139552","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139557","title":"MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents","authors":["Jifeng Gao","Kang Xia","Yi Zhang","Xiaobin Hong","Mingkai Lin","Xingshen Wei","Wenzhong Li","Sanglu Lu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.14651","noteUrl":"/review?paper=review-neurips-2026-139557","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139558","title":"AgentAbstain: Do LLM Agents Know When Not to Act?","authors":["Xun Liu","Yi Evie Zhang","Vira Kasprova","Parisa Rabbani","Pardissadat Zahraei","Tianyu Zhang","Ali Ebrahimpour-Boroojeny","Varun Chandrasekaran"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AgentAbstain: Do LLM Agents Know When Not to Act?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.10059","noteUrl":"/review?paper=review-neurips-2026-139558","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139560","title":"mRNABench: A curated benchmark for mature mRNA property and function prediction","authors":["Ruian (Ian) Shi","Taykhoom Dalal","Philip Fradkin","Divya Koyyalagunta","Simran Chhabria","Andrew Jung","Cyrus L Tam","Defne Ceyhan","Jessica Lin","Kaitlin U Laverty","Ilyes Baali","Bo Wang","Quaid Morris"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"mRNABench: A curated benchmark for mature mRNA property and function prediction","plainExplanation":"","primaryUrl":"https://www.biorxiv.org/content/biorxiv/early/2025/07/08/2025.07.05.662870.full.pdf","noteUrl":"/review?paper=review-neurips-2026-139560","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139561","title":"MedMisBench: Measuring Epistemic Resilience of LLMs Under Misleading Medical Context","authors":["Hongjian Zhou","Xinyu Zou","Jinge Wu","Sean Wu","Junchi Yu","Bradley M Segal","Tobias E Niebuhr","Sara Amro","Michael Petrus","Sheikh Momin","Alexandra M Pinto","Rachel Niesen","Laura S Wegner","Dhruv Darji","Jung M Koo","Joshua Fieggen","Kapil Narain","Mingde Zeng","Lei Clifton","Linda Shapiro","Fenglin Liu","David Clifton"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MedMisBench: Measuring Epistemic Resilience of LLMs Under Misleading Medical Context","plainExplanation":"","primaryUrl":"https://www.biorxiv.org/content/10.64898/2026.05.25.727671v3.full.pdf","noteUrl":"/review?paper=review-neurips-2026-139561","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139565","title":"ReasoningShield: Safety Moderation over Reasoning Traces of Large Reasoning Models","authors":["Changyi Li","Jiayi Wang","Xudong Pan","Geng Hong","Min Yang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"ReasoningShield: Safety Moderation over Reasoning Traces of Large Reasoning Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.17244","noteUrl":"/review?paper=review-neurips-2026-139565","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139567","title":"OpenView: Empowering MLLMs with Out-of-view VQA","authors":["Qixiang Chen","Cheng Zhang","Chi-Wing Fu","Jingwen Ye","Jianfei Cai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"OpenView: Empowering MLLMs with Out-of-view VQA","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.18563","noteUrl":"/review?paper=review-neurips-2026-139567","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139568","title":"RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations","authors":["Hanyu Li","Yichi Zhang","Speed Zhu","Hang Su","Jun Zhu","Yinpeng Dong"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26177","noteUrl":"/review?paper=review-neurips-2026-139568","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139569","title":"Doomed to Re-Annotate, Forever: The ImageNet Story","authors":["Illia Volkov","Nikita Kisel","Tetiana Mishkina","Klara Janouskova","Jiri Matas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Doomed to Re-Annotate, Forever: The ImageNet Story","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.13783","noteUrl":"/review?paper=review-neurips-2026-139569","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139571","title":"$\\pi$-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows","authors":["Haoran Zhang","Luxin Xu","Zhilin Wang","Runquan Gui","Shunkai Zhang","Haodi Lei","Tong Zhu","Xiaoye Qu","Yang Yang","Yu Cheng","Yafu Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"$\\pi$-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14678","noteUrl":"/review?paper=review-neurips-2026-139571","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139572","title":"LiveOption: Evaluating LLM Agents in Structured Option Trading with Nonlinear Payoffs","authors":["Haochen Luo","YIFAN LI","An B Minh","Xiaolong Luo","Zhengzhao Lai","Yuan Zhang","Chen Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LiveOption: Evaluating LLM Agents in Structured Option Trading with Nonlinear Payoffs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.33470","noteUrl":"/review?paper=review-neurips-2026-139572","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139573","title":"Can 4D Foundation Models Remember?","authors":["Guangzhao He","Hadar Averbuch-Elor","Wei-Chiu Ma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Can 4D Foundation Models Remember?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.20819","noteUrl":"/review?paper=review-neurips-2026-139573","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139576","title":"VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding","authors":["He Haichen","Jiayi Zhou","Sifeng SHANG","Yihan Hu","Yuanhan Zhang","Kaiyang Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"VideoOdyssey: A Benchmark for Ultra-Long-Context and Omni-Modal Video Understanding","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22907","noteUrl":"/review?paper=review-neurips-2026-139576","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139577","title":"EnterpriseBench: Evaluating LLM Agents on End-to-End Spreadsheet Tasks in Finance","authors":["Thomson Yen","Julian Poeltl","Harshith S Gear","Yilin Meng","Joshua Fan","Adam Shen","Yili Liu","Ali Bauyrzhan","Siri Du","Haoyang Liu","C. Guetta","Hongseok Namkoong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EnterpriseBench: Evaluating LLM Agents on End-to-End Spreadsheet Tasks in Finance","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22664","noteUrl":"/review?paper=review-neurips-2026-139577","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139579","title":"HierSVA: A Synthesis Pipeline, Dataset, and Benchmark for LLM-Driven Hierarchical Hardware Formal Verification","authors":["Maohua Nie","Jiang Zhu","Jingqun Zhang","Zhichen Zeng","Jiayi Wang","Sibo Zhang","Jialin Wang","C-J R Shi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"HierSVA: A Synthesis Pipeline, Dataset, and Benchmark for LLM-Driven Hierarchical Hardware Formal Verification","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.13706","noteUrl":"/review?paper=review-neurips-2026-139579","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139580","title":"AndroidReality: How Far Are Mobile Agents from the Real World?","authors":["Xiaoou Liu","Longchao Da","Hanyang Chen","Yuan Ling","Hua Wei"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AndroidReality: How Far Are Mobile Agents from the Real World?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.07775","noteUrl":"/review?paper=review-neurips-2026-139580","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139581","title":"Metropolis-Scale Road Network Datasets for Fine-Grained Urban Traffic Modeling","authors":["Fedor Velikonivtsev","Oleg Platonov","Ekaterina Alimaskina","Gleb Bazhenov","Liudmila Prokhorenkova"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Metropolis-Scale Road Network Datasets for Fine-Grained Urban Traffic Modeling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.02278","noteUrl":"/review?paper=review-neurips-2026-139581","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139583","title":"Towards Visual Query Segmentation in the Wild","authors":["Bing Fan","Minghao Li","Hanzhi Zhang","Shaohua Dong","NAGA PRUDHVI MAREEDU","Xiaoqiong Liu","Weishi Shi","Yunhe Feng","Yan Huang","Heng Fan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Towards Visual Query Segmentation in the Wild","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.08898","noteUrl":"/review?paper=review-neurips-2026-139583","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139585","title":"CEO-Bench: Can Agents Play the Long Game?","authors":["Haozhe Chen","Karthik Narasimhan","Zhuang Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The surplus argmax must be restricted to acceptable plans","plainExplanation":"The algorithm defines an acceptable set by budget and quality constraints but maximizes surplus over all plans. With budget \\(100\\), prices \\(50,150\\), required qualities \\(5,10\\) and perceived qualities \\(10,20\\), only the first plan is acceptable.","primaryUrl":"https://arxiv.org/html/2606.18543","noteUrl":"/review?paper=review-neurips-2026-139585","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139588","title":"Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents","authors":["Bowen Ye","Rang Li","Qibin Yang","Yuanxin Liu","Linli Yao","Hanglong Lv","Zhihui Xie","Chenxin An","Lei Li","Lingpeng Kong","Qi Liu","Zhifang Sui","Tong Yang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.06132","noteUrl":"/review?paper=review-neurips-2026-139588","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139589","title":"Benchmarking Open-Ended Multi-Agent Coordination in Language Agents","authors":["Kaleab Tessera","Andras Szecsenyi","Cameron Barker","Alexander Rutherford","Davide Paglieri","Aidan Scannell","Henry Gouk","Elliot Crowley","Tim Rocktäschel","Amos Storkey"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Benchmarking Open-Ended Multi-Agent Coordination in Language Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.08340","noteUrl":"/review?paper=review-neurips-2026-139589","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139590","title":"Semantic Search over 9 Million Mathematical Theorems","authors":["Luke Alexander","Eric Leonen","Sophie Szeto","Artemii Remizov","Ignacio Tejeda","Jarod Alper","Giovanni Inchiostro","Vasily Ilin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Semantic Search over 9 Million Mathematical Theorems","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.05216","noteUrl":"/review?paper=review-neurips-2026-139590","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139591","title":"GISA: A Benchmark for General Information-Seeking Assistant","authors":["Yutao Zhu","Xingshuo Zhang","Maosen Zhang","Jiajie Jin","Liancheng Zhang","Xiaoshuai Song","Kangzhi Zhao","Wencong Zeng","Ruiming Tang","Han Li","Ji-Rong Wen","Zhicheng Dou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Lower adjacent-query Jaccard means more diverse reformulation","plainExplanation":"Search Diversity is explicitly average adjacent-query Jaccard similarity, with lower values meaning greater diversity, but the interpretation assigns the opposite direction. Identical term sets give similarity one and disjoint sets zero.","primaryUrl":"https://arxiv.org/html/2602.08543","noteUrl":"/review?paper=review-neurips-2026-139591","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139595","title":"ECG-Reasoning-Benchmark: A Benchmark for Evaluating Clinical Reasoning Capabilities in ECG Interpretation","authors":["Jungwoo Oh","Hyunseung Chung","Junhee Lee","Min-Gyu Kim","Hangyul Yoon","Ki Seong Lee","Youngchae Lee","Muhan Yeo","Edward Choi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ECG-Reasoning-Benchmark: A Benchmark for Evaluating Clinical Reasoning Capabilities in ECG Interpretation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.14326","noteUrl":"/review?paper=review-neurips-2026-139595","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139596","title":"SR-Prominence: A Crowdsourced Protocol and Dataset Suite for Perceptually-Weighted Super-Resolution Artifact Evaluation","authors":["Ivan Molodetskikh","Kirill Malyshev","Mark Mirgaleev","Nikita Zagainov","Evgeney Bogatyrev","Dmitriy Vatolin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SR-Prominence: A Crowdsourced Protocol and Dataset Suite for Perceptually-Weighted Super-Resolution Artifact Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14847","noteUrl":"/review?paper=review-neurips-2026-139596","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139597","title":"BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors","authors":["Jonas Ricker","Asja Fischer","Erwin Quiring"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31153","noteUrl":"/review?paper=review-neurips-2026-139597","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139598","title":"B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data","authors":["Magdalena Proszewska","Tomasz Danel","Antoni Antoszek","Dawid Damian Rymarczyk"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2505.22252","noteUrl":"/review?paper=review-neurips-2026-139598","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139599","title":"TerraVis: Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows","authors":["Shuai Fu","Jing Gu","Jian Zhou","Zicheng Duan","Gengze Zhou","Qi Wu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"TerraVis: Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02959","noteUrl":"/review?paper=review-neurips-2026-139599","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139602","title":"MultiTalk: Scaling Full-Duplex Speech Models to Long, Multi-Party, Bilingual Conversation","authors":["Ke Wang","Houxing Ren","Zimu Lu","Yunqiao Yang","ZHUOFAN ZONG","Mingjie Zhan","Hongsheng Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MultiTalk: Scaling Full-Duplex Speech Models to Long, Multi-Party, Bilingual Conversation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36903","noteUrl":"/review?paper=review-neurips-2026-139602","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139607","title":"GUIGuard-Bench:  Toward a General Evaluation for Privacy-Preserving GUI Agents","authors":["Yanxi Wang","Zhiling Zhang","Wenbo Zhou","Weiming Zhang","Jie Zhang","Qiannan Zhu","Yu Shi","Shuxin Zheng","Jiyan He"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"GUIGuard-Bench:  Toward a General Evaluation for Privacy-Preserving GUI Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.18842","noteUrl":"/review?paper=review-neurips-2026-139607","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139609","title":"TELEVAL: A Benchmark Designed for Spoken Language Models in Chinese Interactive Scenarios","authors":["Zehan Li","Hongjie Chen","Qing Wang","Yuxin Zhang","Jing Zhou","Hang Lv","Mengjie Du","Yaodong Song","Jie Lian","Jian Kang","Jie Li","Yongxiang Li"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"TELEVAL: A Benchmark Designed for Spoken Language Models in Chinese Interactive Scenarios","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2507.18061v4","noteUrl":"/review?paper=review-neurips-2026-139609","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139610","title":"SenseBench: A Benchmark for Remote Sensing Low-Level Visual Perception and Description in Large Vision-Language Models","authors":["Chen Zhong","Xiao An","Jiaxing Sun","Zihan Gui","Guangyi Yang","Wei He"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SenseBench: A Benchmark for Remote Sensing Low-Level Visual Perception and Description in Large Vision-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10576","noteUrl":"/review?paper=review-neurips-2026-139610","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139611","title":"StakeBench: Evaluating Language Understanding Grounded in Market Commitment","authors":["Yunhua Pei","Jingyu Hu","Yiwei Shi","Hongnan Ma","Weiru Liu","John Cartlidge"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"StakeBench: Evaluating Language Understanding Grounded in Market Commitment","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.26074","noteUrl":"/review?paper=review-neurips-2026-139611","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139612","title":"EchoXFlow: A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function","authors":["Elias Stenhede","Joanna Sulkowska","Eivind B Orstad","Henrik Schirmer","Arian Ranjbar"],"year":2026,"category":"Computer vision & medical imaging","status":"finding","issueTitle":"EchoXFlow: A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function","plainExplanation":"No issue was found in the checked claims. The equality in v1 was false for the architecture it described. A convolution with a shorter temporal kernel has fewer learned weights when its channels and groups are fixed. The current manuscript states different counts, so this exact historical issue is addressed.","primaryUrl":"https://arxiv.org/html/2605.05447","noteUrl":"/review?paper=review-neurips-2026-139612","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139613","title":"StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs","authors":["Pierre Le Jeune","Etienne Duchesne","Weixuan Xiao","Stefano Palminteri","Bazire Houssin","Benoît Malézieux","Matteo Dora"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10442","noteUrl":"/review?paper=review-neurips-2026-139613","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139616","title":"LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale","authors":["Francesco Mantegna","Dulhan Jayalath","Gereon Elvers","Tasha Kim","Benjamin Ballyk","SungJun Cho","Teyun Kwon","Luisa Kurth","Miran Özdogan","Gilad Landau","Pratik Somaiya","Natalie Voets","Mark Woolrich","Oiwi Parker Jones"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.25204","noteUrl":"/review?paper=review-neurips-2026-139616","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139617","title":"PolyTopoBench: A Benchmark for Complex Vector Polygon Generation from Remote Sensing Imagery","authors":["Zeping Liu","Ni Lao","Weiwei Sun","Gil Wolff","Yiqun Xie","Liang Zhao","Junfeng Jiao","Gengchen Mai"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PolyTopoBench: A Benchmark for Complex Vector Polygon Generation from Remote Sensing Imagery","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.32856","noteUrl":"/review?paper=review-neurips-2026-139617","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139619","title":"Mini Amusement Parks (MAPs): A Testbed for Modelling Business Decisions","authors":["Stéphane Aroca-Ouellette","Ian Berlot-Attwell","Panagiotis Lymperopoulos","Abhiramon Rajasekharan","Tongqi Zhu","Herin Kang","Kaheer Suleman","Sam Pasupalak"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mini Amusement Parks (MAPs): A Testbed for Modelling Business Decisions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.15830","noteUrl":"/review?paper=review-neurips-2026-139619","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139620","title":"Beyond Bag-of-Words: Diagnosing Compositional Binding Failures in Vision-Language Models","authors":["Cristian Sbrolli","Toshihiko Yamasaki","Matteo Matteucci"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Beyond Bag-of-Words: Diagnosing Compositional Binding Failures in Vision-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.02043","noteUrl":"/review?paper=review-neurips-2026-139620","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139621","title":"BenchRep-T: A Systematic Evaluation of T-Cell Repertoire-Based Disease Diagnostics","authors":["Chiho Im","Liel Cohen-Lavi","Alejandro Buendia","Anshul Kundaje","Scott D Boyd"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BenchRep-T: A Systematic Evaluation of T-Cell Repertoire-Based Disease Diagnostics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://www.biorxiv.org/content/10.64898/2026.06.09.727013v1.full.pdf","noteUrl":"/review?paper=review-neurips-2026-139621","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139622","title":"Measuring AI Agents' Progress on Multi-Step Cyber Attack Scenarios","authors":["Linus Folkerts","Will Payne","Simon Inman","Philippos M Giavridis","Joe Skinner","Sam Deverett","James Aung","Ekin Zorer","Michael Schmatz","Mahmoud Ghanem","John Wilkinson","Alan Steer","The Vy Hong","Harry Coppock","Jessica Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Measuring AI Agents' Progress on Multi-Step Cyber Attack Scenarios","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.11214","noteUrl":"/review?paper=review-neurips-2026-139622","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139623","title":"TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation","authors":["Hanyu Guo","JieDong Yang","Chao Chen","Longfei Xu","Kaikui Liu","Xiangxiang Chu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22355","noteUrl":"/review?paper=review-neurips-2026-139623","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139624","title":"Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors","authors":["Jonas Wiedermann-Möller","Leonard Dung","Maksym Andriushchenko"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.06490v1","noteUrl":"/review?paper=review-neurips-2026-139624","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139626","title":"MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI","authors":["Bohan Lyu","Yucheng Yang","Siqiao Huang","Jiaru Zhang","Qixin Xu","Xinghan Li","Xinyang Han","Huaqing Zhang","Yicheng Zhang","Runhan Huang","Kaicheng Yang","Zitao Chen","Wentao Guo","Junlin Yang","Xinyue Ai","Wenhao Chai","Yadi Cao","Ziran Yang","Kun Wang","Dapeng Jiang","Huan-ang Gao","Shange Tang","Chengshuai Shi","Simon Du","Max Simchowitz","Jiantao Jiao","Dawn Song","Chi Jin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.08678v3","noteUrl":"/review?paper=review-neurips-2026-139626","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139628","title":"OBLIQ-Bench: Exposing Overlooked Bottlenecks in Modern Retrievers with Latent and Implicit Queries","authors":["Diane Tchuindjo","Devavrat Shah","Omar Khattab"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OBLIQ-Bench: Exposing Overlooked Bottlenecks in Modern Retrievers with Latent and Implicit Queries","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.06235v2","noteUrl":"/review?paper=review-neurips-2026-139628","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139629","title":"ScAn-Bench: Evaluating Scaling Analysis Methodology","authors":["Artin Sermaxhaj","Nastaran Alipour","Donat Sinani","Johannes Hog","Neeratyoy Mallik","Steven Adriaensen","Jenia Jitsev","Danny Stoll"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ScAn-Bench: Evaluating Scaling Analysis Methodology","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35707","noteUrl":"/review?paper=review-neurips-2026-139629","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139631","title":"RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades","authors":["Xinbo Xu","Ruihan Yang","Haiyang Shen","Wendong XU","Bofei Gao","Ruoyu Wu","Kean Shi","Weichu Xie","Xuanzhong Chen","Ming Wu","Jason Zeng","Michael Heinrich","Liang Chen","Kuan Li","Baobao Chang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.15846v2","noteUrl":"/review?paper=review-neurips-2026-139631","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139634","title":"Evaluating Test-Time Scaling of General LLM Agents","authors":["Xiaochuan Li","Tianshi Ming","Pranav Setlur","Abhijay S Paladugu","Andy Tang","Hao Kang","Shuai Shao","Rong Jin","Chenyan Xiong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Finite-population correction does not reduce fresh inference noise","plainExplanation":"The confidence calculation applies a without-replacement finite-population correction to task differences that also contain newly sampled stochastic inference outcomes. Write \\(d_j=\\mu_j+\\epsilon_j\\) with inference variance \\(v_j\\).","primaryUrl":"https://arxiv.org/html/2602.18998","noteUrl":"/review?paper=review-neurips-2026-139634","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139635","title":"Measuring Collapse and Correction in Homogeneous-Panel LLM Debate","authors":["Xin Li","Mengbing LIU","Chau Yuen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Measuring Collapse and Correction in Homogeneous-Panel LLM Debate","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.35279","noteUrl":"/review?paper=review-neurips-2026-139635","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139636","title":"IDEAFix: Evaluation Framework for Creative Defixation Prompting in LLMs","authors":["Florian Carichon","Soumya Sharma","Meaghan J. Girard","Romain Rampa","Golnoosh Farnadi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"IDEAFix: Evaluation Framework for Creative Defixation Prompting in LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.00875","noteUrl":"/review?paper=review-neurips-2026-139636","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139637","title":"Efficient Benchmarking Is Just Feature Selection and Multiple Regression","authors":["Sam Bowyer","Acyr Locatelli","Kris Cao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Efficient Benchmarking Is Just Feature Selection and Multiple Regression","plainExplanation":"The stated objective optimizes the intercept jointly with the coefficients. Eliminating that intercept centers the design and target, so the uncentered expression need not minimize the displayed objective. The relevant objective and formula are unchanged from v2.","primaryUrl":"https://arxiv.org/html/2605.25773","noteUrl":"/review?paper=review-neurips-2026-139637","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139640","title":"FinMTM: A Multi-Turn Multimodal Benchmark for Financial Reasoning and Agent Evaluation","authors":["Chenxi Zhang","Ziliang Gan","Liyun Zhu","Youwei Pang","Qing Zhang","Rongjunchen Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"FinMTM: A Multi-Turn Multimodal Benchmark for Financial Reasoning and Agent Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.03130","noteUrl":"/review?paper=review-neurips-2026-139640","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139641","title":"Auditing Sabotage Bench: A Benchmark for Detecting and Fixing Research Sabotage in ML Codebases","authors":["Eric Gan","Aryan Bhatt","Buck Shlegeris","Julian Stastny","Vivek Hebbar"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Auditing Sabotage Bench: A Benchmark for Detecting and Fixing Research Sabotage in ML Codebases","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.16286","noteUrl":"/review?paper=review-neurips-2026-139641","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139645","title":"Tracing the Cascade: A Topology-Aware Evaluation Framework for Scientific Agent Hallucinations","authors":["Xinshun Feng","Ziqi Miao","Lijun Li","Jing Shao"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Tracing the Cascade: A Topology-Aware Evaluation Framework for Scientific Agent Hallucinations","plainExplanation":"Replacing a refuted hub claim with a refuted off-graph claim can lower weighted hallucination severity while leaving raw severity unchanged. The scoring formulas support a narrower append-only guarantee when the original judged claims remain fixed.","primaryUrl":"https://arxiv.org/html/2608.00711","noteUrl":"/review?paper=review-neurips-2026-139645","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139647","title":"CryptanalysisBench: Can LLMs do cryptanalysis?","authors":["Lukas Fluri","Avital Shafran","Nicholas Carlini","Matthew Jagielski","Milad Nasr","Orr Dunkelman","Eyal Ronen","Florian Tramer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CryptanalysisBench: Can LLMs do cryptanalysis?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.18538","noteUrl":"/review?paper=review-neurips-2026-139647","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139648","title":"OpenClawBench: Benchmarking Process-side Anomalies in Real-world Agent Execution Trajectories","authors":["Yibing Liu","Yangze Liu","Xiao-Long Yin","Bin Wang","Chong Zhang","Hao Yin","Zhongyi Han"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"OpenClawBench: Benchmarking Process-side Anomalies in Real-world Agent Execution Trajectories","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.29253","noteUrl":"/review?paper=review-neurips-2026-139648","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139649","title":"WAXAL: A Large-Scale Multilingual African Language Speech Corpus","authors":["MohamedElfatih MohamedKhair","Emmanuel Asiedu Brempong","Subhashini Venugopalan","Abdoulaye Diack","Perry Nelson","Mireku","Tavonga Siyavora","Pooja Rao","Uche Okonkwo","Angela Nakalembe","Abhishek Bapna","Aisha Walcott-Bryant"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"WAXAL: A Large-Scale Multilingual African Language Speech Corpus","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.02734","noteUrl":"/review?paper=review-neurips-2026-139649","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139650","title":"MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers","authors":["Chaithanya Bandi","Razvan Dumitru","Ben Hertzberg","Divyansh Agarwal","Geobio Boo","Tejas Polakam","Jeff Da","HiJae Kim","Vipul Gupta","Manasi Sharma","Andrew Park","Martin Villanueva Dimakis","Ernesto G Montoya","Rafael Cruz","MohammadHossein Rezaei","Chetan Rane","Daniel Y Zhang","Brad Kenstler","Bing Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.00933","noteUrl":"/review?paper=review-neurips-2026-139650","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139651","title":"s2n-bignum-bench: A practical benchmark for evaluating low-level code reasoning of LLMs","authors":["Balaji Rao","Soonho Kong","Juneyoung Lee","Carlo Lipizzi"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"s2n-bignum-bench: A practical benchmark for evaluating low-level code reasoning of LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.14628","noteUrl":"/review?paper=review-neurips-2026-139651","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139652","title":"ContextShift: A Controlled Benchmark for Context Dependence in Object Detection","authors":["Dan Zlotnikov","Alex Lazarovich","Ohad Ben-Shahar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"NPMI needs a joint probability","plainExplanation":"The prose defines a conditional probability where NPMI requires a joint probability, breaking its independence and range properties. The printed probability definition is incompatible with the NPMI formula.","primaryUrl":"https://arxiv.org/html/2606.09495","noteUrl":"/review?paper=review-neurips-2026-139652","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139653","title":"Stress-Testing Neural Network Verifiers with Provably Robust Instances","authors":["David Troxell","Yulia Alexandr","Sofia Hunt","Stephanie Lei","Guido Montufar"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Stress-Testing Neural Network Verifiers with Provably Robust Instances","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17153","noteUrl":"/review?paper=review-neurips-2026-139653","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139660","title":"CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI","authors":["Austin T D Andrews","Liam Wilkinson","Jamie Heagerty","Harry Coppock","Jakob Foerster","Rui Costa"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.02459","noteUrl":"/review?paper=review-neurips-2026-139660","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139663","title":"SciResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery","authors":["Lei Xiong","Kun Luo","Ziyi Xia","Wenbo Zhang","Jingying Shao","Jianlyu Chen","Hongjin Qian","Zhicheng Dou","Zheng Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SciResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.25256","noteUrl":"/review?paper=review-neurips-2026-139663","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139664","title":"An Assessment of Human vs. Model Uncertainty in Soft-Label Learning and Calibration","authors":["Maja Pavlovic","Silviu Paun","Massimo Poesio"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"An Assessment of Human vs. Model Uncertainty in Soft-Label Learning and Calibration","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.18648","noteUrl":"/review?paper=review-neurips-2026-139664","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139665","title":"LoopNav: Benchmarking Spatial Consistency in World Models","authors":["Kewei Lian","Shaofei Cai","Yitao Liang","Anji Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LoopNav: Benchmarking Spatial Consistency in World Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2505.22976","noteUrl":"/review?paper=review-neurips-2026-139665","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139667","title":"RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models","authors":["Ruinan Jin","Xinting Liao","Hanlin Yu","Deval Pandya","Xiaoxiao Li"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.00443","noteUrl":"/review?paper=review-neurips-2026-139667","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139669","title":"K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs","authors":["Hao Liang","Qihan Lin","Mingrui Chen","Hengyi Feng","ZhaoYang Han","Xiaochen Ma","Zhen H Wong","Meiyi Qiang","Linzhuang Sun","Wentao Zhang"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09635","noteUrl":"/review?paper=review-neurips-2026-139669","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139670","title":"Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses","authors":["Jonathan Stray","David Yang","Steven Luo","Miu N Takagi","Serina Chang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28911","noteUrl":"/review?paper=review-neurips-2026-139670","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139671","title":"Feedback Forensics: A Toolkit to Measure AI Personality","authors":["Arduin Findeis","Timo Kaufmann","Eyke Hüllermeier","Robert Mullins"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Feedback Forensics: A Toolkit to Measure AI Personality","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.26305","noteUrl":"/review?paper=review-neurips-2026-139671","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139672","title":"AgenticVBench: Can AI Agents Complete Real-World Video Production Tasks?","authors":["Zongheng Cao","Yi Zheng","Xinyu Hu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Chance adjustment permits negative assembly scores","plainExplanation":"A valid all-distractor answer has \\(r=0\\), giving adjusted score \\(-1/(k-1)=-1/2\\).","primaryUrl":"https://arxiv.org/html/2605.27705","noteUrl":"/review?paper=review-neurips-2026-139672","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139673","title":"KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards","authors":["Pengfei Li","Naufal Suryanto","Sicheng Zhang","Muhammad Muzammal Naseer"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02206","noteUrl":"/review?paper=review-neurips-2026-139673","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139674","title":"ObjView-Bench: Rethinking Difficulty and Deployment for Object-Centric View Planning","authors":["Sicong Pan","Hao Hu","Xuying Huang","Benno Wingender","Maren Bennewitz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ObjView-Bench: Rethinking Difficulty and Deployment for Object-Centric View Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10707","noteUrl":"/review?paper=review-neurips-2026-139674","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139675","title":"Eyes on VLM: Benchmarking Gaze Following and Social Gaze Prediction in Vision Language Models","authors":["Hengfei Wang","Anshul Gupta","Pierre Vuillecard","Jean-marc Odobez"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Eyes on VLM: Benchmarking Gaze Following and Social Gaze Prediction in Vision Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.19859","noteUrl":"/review?paper=review-neurips-2026-139675","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139676","title":"PICID: A Modular Evaluation Infrastructure for Reproducible PHM Across Tasks and Domains","authors":["Lev Telyatnikov","Raffael Theiler","Leandro Von Krannichfeldt","Olga Fink"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The at-most-one-hot encoding permits a cross-dataset collision","plainExplanation":"The allowed zero vector has \\(\\tilde c=0\\). This contradicts the stated positive concept-code range and overlaps the adjacent dataset’s class block.","primaryUrl":"https://arxiv.org/html/2605.28345","noteUrl":"/review?paper=review-neurips-2026-139676","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139677","title":"Evaluating Multimodal Narrative Understanding of Popular Hollywood Films","authors":["David Bamman","Kent K Chang","Allison Cooper","Juishan Hsu","Reina Kushihashi","Madison Mar","Arnav Podichetty","Rachael G Samberg","Ipek N Sancak","Yuhan Shao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Evaluating Multimodal Narrative Understanding of Popular Hollywood Films","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.21430","noteUrl":"/review?paper=review-neurips-2026-139677","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139678","title":"AccelEval: A Whole-Program Benchmark for LLM-Generated CPU-to-GPU Code Acceleration","authors":["Chendong Song","Tianci Bu","Meixuan Wang","Zijie Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AccelEval: A Whole-Program Benchmark for LLM-Generated CPU-to-GPU Code Acceleration","plainExplanation":"","primaryUrl":"https://papers.ssrn.com/sol3/Delivery.cfm/6835659.pdf?abstractid=6835659&mirid=1","noteUrl":"/review?paper=review-neurips-2026-139678","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139679","title":"CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations","authors":["Zixian Su","Hongkai Zhang","Fan Gao","Encheng Su","Taiping Qu","Jingwei Guo","Zhang Nan","Hui Wang","Zhen Zhou","Kairui Bo","Yan Chen","Yue Ren","Shuai Li","Lei Xu","Henggui Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.00123","noteUrl":"/review?paper=review-neurips-2026-139679","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139680","title":"stable-worldmodel: A Platform for Reproducible World Modeling Research and Evaluation","authors":["Lucas Maes","Quentin Le Lidec","Luiz Facury","Nassim Massaudi","Ayush Chaurasia","Francesco Capuano","Richard Gao","Taj Gillin","Dan Haramati","Damien Scieur","Yann LeCun","Randall Balestriero"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The policy-extraction weight uses an unbound next-state latent","plainExplanation":"The displayed extraction block does not bind the next observation or latent needed to evaluate its weight. Separately, the chosen value difference should not be silently identified with the training residual.","primaryUrl":"https://arxiv.org/html/2605.21800","noteUrl":"/review?paper=review-neurips-2026-139680","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["algorithm_specification","definition_ambiguity","notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-139682","title":"RepoZero: Can LLMs Generate a Code Repository from Scratch?","authors":["Zhaoxi Zhang","Yiming Xu","Jiahui Liang","Weikang Li","Xiaoshuai Chen","Liwei Qian","Xin Pei","Jizhou Huang","Rui Sun","Yunfang Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RepoZero: Can LLMs Generate a Code Repository from Scratch?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07122","noteUrl":"/review?paper=review-neurips-2026-139682","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139683","title":"JudgmentBench: Comparing Rubric and Preference Evaluation for Quality Assessment","authors":["Russell Yang","Ruishi Chen","Pierce Kelaita","Riya Ranjan","Sibo Ma","Charles Dickens","Matthew Guillod","Megan Ma","Julian Nyarko"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"JudgmentBench: Comparing Rubric and Preference Evaluation for Quality Assessment","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.25240","noteUrl":"/review?paper=review-neurips-2026-139683","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139685","title":"MemArena: An Ego-Centric Benchmark for On-Device Agentic Personal Memory Assistants at Scale","authors":["Jiadong Zhang","Xiaosong Ma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MemArena: An Ego-Centric Benchmark for On-Device Agentic Personal Memory Assistants at Scale","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.02613","noteUrl":"/review?paper=review-neurips-2026-139685","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139686","title":"Liars' Bench: Evaluating Lie Detectors for Language Models","authors":["Kieron Kretschmar","Walter Laurito","Sharan Maiya","Samuel Marks"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Liars' Bench: Evaluating Lie Detectors for Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2511.16035v2","noteUrl":"/review?paper=review-neurips-2026-139686","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139687","title":"WorldCoder-Bench: Benchmarking Physically Grounded 3D World Synthesis","authors":["Shuo Lu","Yinuo Xu","Kecheng Yu","Siru Jiang","Yongcan Yu","Yubin Wang","Haitao Yang","Yuxiang Zhang","Bin Wang","Ran He","Jian Liang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"WorldCoder-Bench: Benchmarking Physically Grounded 3D World Synthesis","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.01869","noteUrl":"/review?paper=review-neurips-2026-139687","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139688","title":"WildBox: A Dataset and Benchmark for Aerial Monocular 3D Detection of African Savanna Wildlife","authors":["Vandita Shukla","Kilian Meier","Lucie Laporte-Devylder","Camille R Saint-Jean","Jenna Kline","Blair R Costelloe","Devis Tuia","Fabio Remondino","Benjamin Risse"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"WildBox: A Dataset and Benchmark for Aerial Monocular 3D Detection of African Savanna Wildlife","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.21309","noteUrl":"/review?paper=review-neurips-2026-139688","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139690","title":"SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of Foundation Model-based Embodied Agents","authors":["Simon Zhan","Philip Wang","Yao Liu","Yiyan Peng","Yiqi Lyu","Zinan Wang","Qineng Wang","Zhian Ruan","Xiangyu Shi","Xinyu Cao","Frank Yang","Zhenyang Ni","Kangrui Wang","Ruohan Zhang","Huajie Shao","Manling Li","Qi Zhu"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of Foundation Model-based Embodied Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.12985","noteUrl":"/review?paper=review-neurips-2026-139690","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139691","title":"Explanation Multiplicity in SHAP: Characterization and Assessment","authors":["Hyunseung Hwang","Seungeun Lee","Lucas Rosenblatt","Steven Whang","Julia Stoyanovich"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Explanation Multiplicity in SHAP: Characterization and Assessment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.12654","noteUrl":"/review?paper=review-neurips-2026-139691","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139692","title":"Driving on Memory","authors":["Christian Löwens","Thorben Funke","Alexandru P P Condurache"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Driving on Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2608.31029","noteUrl":"/review?paper=review-neurips-2026-139692","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139693","title":"OpenMedReason: Scientific Reasoning Supervision for Medical Vision–Language Models","authors":["Negin Baghbanzadeh","Pritam Sarkar","Michael Colacci","Abeer Badawi","Adibvafa Fallahpour","Arash Afkanpour","Leonid Sigal","Ali Etemad","Elham Dolatabadi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"OpenMedReason: Scientific Reasoning Supervision for Medical Vision–Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.12169","noteUrl":"/review?paper=review-neurips-2026-139693","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139694","title":"Gender Artifacts from Art History to Text-to-Image Generation","authors":["Piera Riccio","Miriam Doh","Benedikt Höltgen","Noa Garcia","Nanne van Noord"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Gender Artifacts from Art History to Text-to-Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05829","noteUrl":"/review?paper=review-neurips-2026-139694","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139695","title":"SREGym: A Live Benchmark for AI SRE Agents with High-Fidelity Failure Scenarios","authors":["Jackson Clark","Yiming Su","Saad Mohammad Rafid Pial","Yifang Tian","Lily Gniedziejko","Hans Arno Jacobsen","Yinfang Chen","Tianyin Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SREGym: A Live Benchmark for AI SRE Agents with High-Fidelity Failure Scenarios","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07161","noteUrl":"/review?paper=review-neurips-2026-139695","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139696","title":"Mosaic: A Benchmark Suite for Differentiable Physics Solvers","authors":["Andrin Rehmann","Heiko Zimmermann","Dion Häfner"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mosaic: A Benchmark Suite for Differentiable Physics Solvers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.27895","noteUrl":"/review?paper=review-neurips-2026-139696","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139697","title":"MM-OptBench: A Solver-Grounded Benchmark for Multimodal Optimization Modeling","authors":["Zhong Li","Qi Huang","Yuxuan Zhu","Mohammad Mohammadi Amiri","Niki van Stein","Thomas Bäck","Matthijs van Leeuwen","Zaiwen Wen","Lincen Yang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"MM-OptBench: A Solver-Grounded Benchmark for Multimodal Optimization Modeling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12154","noteUrl":"/review?paper=review-neurips-2026-139697","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139698","title":"Where Root Cause Analysis Fails: A Retrieval-Reranking Decomposition","authors":["Hada Melino Muhammad","Luan Pham","Laure Barrière","Sachin Shetty","Leonardo Pulga","Flora Salim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Deleting one edge need not remove the causal node","plainExplanation":"With recovered edges \\(m\\to b\\), \\(b\\to c\\), \\(a\\to c\\), and \\(c\\to a\\), the missing \\(m\\to a\\) does not remove \\(m\\) as a root or its path to \\(c\\).","primaryUrl":"https://arxiv.org/html/2609.36686","noteUrl":"/review?paper=review-neurips-2026-139698","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139699","title":"VeriContest: A Competitive-Programming Benchmark for Verifiable Code Generation","authors":["Zichen Xie","Mrigank Pawagi","Yuxin Liu","Aaditi Rai","Lize Shao","John Berberian","Sicong Che","Wenxi Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"VeriContest: A Competitive-Programming Benchmark for Verifiable Code Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08553","noteUrl":"/review?paper=review-neurips-2026-139699","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139700","title":"MCPHunt: An Evaluation Framework for Cross-Boundary Data Propagation in Multi-Server MCP Agents","authors":["Haonan Li","Tianjun Sun","Yongqing Wang","Qisheng Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MCPHunt: An Evaluation Framework for Cross-Boundary Data Propagation in Multi-Server MCP Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.27819","noteUrl":"/review?paper=review-neurips-2026-139700","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139701","title":"Models Recall What They Violate: Constraint Adherence in Multi-Turn LLM Ideation","authors":["Garvin Kruthof"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Models Recall What They Violate: Constraint Adherence in Multi-Turn LLM Ideation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.28031v2","noteUrl":"/review?paper=review-neurips-2026-139701","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139702","title":"FindStatBench: Evaluating Large Language Models on Combinatorial Code Synthesis","authors":["Soham Dan"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"FindStatBench: Evaluating Large Language Models on Combinatorial Code Synthesis","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2607.18260","noteUrl":"/review?paper=review-neurips-2026-139702","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139703","title":"ITPEval: Benchmarking Formal Translation Across Interactive Theorem Provers","authors":["Jiayi Wu","Robert Joseph George","Animashree Anandkumar"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ITPEval: Benchmarking Formal Translation Across Interactive Theorem Provers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.19407","noteUrl":"/review?paper=review-neurips-2026-139703","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139704","title":"SciReason: A Controllable Benchmark for Scientific Reasoning in LLMs","authors":["Pierre Beckmann","Marco Valentino","Andre Freitas"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"SciReason: A Controllable Benchmark for Scientific Reasoning in LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.13020","noteUrl":"/review?paper=review-neurips-2026-139704","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139707","title":"SETA: Scaling Environments for Terminal Agents","authors":["Qijia Shen","Zhiqi Huang","Vamsidhar Kamanuru","Aznaur Aliev","Jay Rainton","Ahmed Awelkair","Boyuan Ma","Qizheng Zhang","Jiwei Fu","Yuzhen Mao","Wendong Fan","Ping Nie","Philip Torr","Bernard Ghanem","Changran Hu","Jonathan Li","Urmish Thakker","Guohao Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SETA: Scaling Environments for Terminal Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.10891","noteUrl":"/review?paper=review-neurips-2026-139707","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139708","title":"SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark","authors":["Taeyoung Yu","Seonbin Jo","Jiwon Kim","Junyoung Byun"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.13730","noteUrl":"/review?paper=review-neurips-2026-139708","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139709","title":"CalArena: A Large Scale Post-Hoc Calibration Benchmark","authors":["Eugène Berta","David Holzmüller","Francis Bach","Michael Jordan"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"CalArena: A Large Scale Post-Hoc Calibration Benchmark","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30188","noteUrl":"/review?paper=review-neurips-2026-139709","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139711","title":"DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving","authors":["Anh Hao Vo","Khoa Vo","Phu Loc Nguyen","Sieu Tran","Duc Nguyen","Ngo X Cuong","Gladys Gawugah","Sreevenkata A Godavarthi","Chase Rainwater","Nghi Bui","Anh Nguyen","Duy M. H. Nguyen","Ngan Le"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.23176","noteUrl":"/review?paper=review-neurips-2026-139711","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139715","title":"A Multimodal Benchmark for Evaluating Cause-of-Death Inference Using Child Health and Mortality Data","authors":["Junhe Yang","Soumyakanti Pan","Hyun Seung Lim","YUE CHU","Yuting Guo","Nishtha Agarwal","Varun Babbar","Gaurav Rajesh Parikh","Yiqun Chen","Chris A Rees","Ziyaad Dangor","Sanjay G Lala","Zehang Li","Samuel J Clark","Zhenke Wu","Abhirup Datta","Li Liu","Cynthia Rudin","Samuel Scarpino","Benjamin M Gyori","Tyler H. McCormick"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"A Multimodal Benchmark for Evaluating Cause-of-Death Inference Using Child Health and Mortality Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://www.medrxiv.org/content/10.64898/2026.07.13.26357980v1.full.pdf","noteUrl":"/review?paper=review-neurips-2026-139715","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139717","title":"What Gets Measured Gets Managed: Sign-aware Recommendation Needs Sign-aware Evaluation","authors":["Minchan Kim","Jungmin Hwang","Hyunwoo Park"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The orthogonal gradient requires weight variation on the user embedding’s support","plainExplanation":"Coordinates where $u_j=0$ vanish from the elementwise product. Variation of $w$ only on those coordinates cannot create an orthogonal component. Equation (15) is valid for $u\\ne0$; the erroneous step is the subsequent assertion that collinearity requires all entries of $w$ to be equal. The correction concerns Proposition 3, the inference after Eq. (15), and uses of its unconditional orthogonal-gradient claim in Section 6.3 and Appendix F. The projection identity is valid for $u\\ne0$; the missing hypothesis is variation of $w$ on $\\operatorname{supp}(u)$. The result concerns the item-embedding derivative for a fixed user embedding. It does not establish that a full training update preserves relevance, and the paper already acknowledges that the parallel component changes the relevance score.","primaryUrl":"https://arxiv.org/html/2609.33346","noteUrl":"/review?paper=review-neurips-2026-139717","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139720","title":"HERO: Improving the Reliability and Sensitivity of Generative Model Evaluation Using Historical Data","authors":["Zhenyu Zhao","Xinrui Ruan","Waverly Wei","Yueshan Zhang","Zeyu Zheng","Sui Huang","Jingshen Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The prevalence rate requires informative current labelers","plainExplanation":"Calibration can identify a labeler's sensitivity and specificity without making its current labels informative about prevalence. If $\\alpha_j=\\beta_j=1/2$ for every current labeler, the class-conditional label laws coincide. The current likelihood is constant in $\\mu$, so the inverse-information step in the rate proof is unavailable. The contradiction concerns the missing current-information condition in Theorem 1. Theorems 2–3 already write or use $I_0>0$; their hypotheses should state this positivity condition explicitly.","primaryUrl":"https://arxiv.org/html/2606.29784","noteUrl":"/review?paper=review-neurips-2026-139720","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139721","title":"ColorConceptBench: A Benchmark for Probabilistic Color-Concept Understanding in Text-to-Image Models","authors":["Chenxi Ruan","Yihan Hou","Yu Xiao","Guosheng Hu","Wei Zeng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The printed correlation formula normalizes each coordinate separately","plainExplanation":"Each nonzero summand becomes a sign rather than a magnitude-weighted covariance contribution. The resulting sum is not a Pearson coefficient and can be undefined on uniform distributions.","primaryUrl":"https://arxiv.org/html/2601.16836","noteUrl":"/review?paper=review-neurips-2026-139721","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139722","title":"Where Do We (Not) Need Temporal Context in Low-Resource Video Task Adaptation?","authors":["Luc Sträter","Hazel Doughty"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Where Do We (Not) Need Temporal Context in Low-Resource Video Task Adaptation?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03837","noteUrl":"/review?paper=review-neurips-2026-139722","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139724","title":"Evaluating AI-based Scientific Knowledge Synthesis with Epidemiological Systematic Reviews","authors":["Shreyansh Padarha","Ryan Othniel Kearns","Tristan M Naidoo","Lingyi Yang","Łukasz Borchmann","Piotr Blaszczyk","Christian Morgenstern","Ruth McCabe","Sangeeta Bhatia","Philip Torr","Jakob Foerster","Scott Hale","Thomas Rawson","Anne Cori","Elizaveta Semenova","Adam Mahdi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Evaluating AI-based Scientific Knowledge Synthesis with Epidemiological Systematic Reviews","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.22327","noteUrl":"/review?paper=review-neurips-2026-139724","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139725","title":"Bench-MFG: A Benchmark Suite for Learning in Stationary Mean Field Games","authors":["Lorenzo Magnino","Jiacheng Shen","Matthieu Geist","Olivier Pietquin","Mathieu Lauriere"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The fictitious-play weights do not preserve probability mass","plainExplanation":"The weights sum to \\(1+1/(k(k+1))\\), not one. The printed averaging recurrence fails normalization; complementary weights in the main damped-fixed-point formulation already avoid this problem.","primaryUrl":"https://arxiv.org/html/2602.12517","noteUrl":"/review?paper=review-neurips-2026-139725","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139726","title":"MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios","authors":["Zhang Li","Lin Zhibo","Qiang Liu","Ziyang Zhang","Shuo Zhang","Zidun Guo","Jiajun Song","Jiarui Zhang","Xiang Bai","Yuliang Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.28130","noteUrl":"/review?paper=review-neurips-2026-139726","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139727","title":"SPATIALEPIBENCH: Benchmarking Spatial Information and Epidemic Priors in Forecasting","authors":["Ruiqi Lyu","Alistair Turcan","Bryan Wilder"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Nonnegative rates do not ensure that the propagation inverse exists","plainExplanation":"The allowed parameter domain includes singular matrices. Even when the inverse exists, the stated conditions do not ensure a nonnegative propagation output.","primaryUrl":"https://arxiv.org/html/2605.06530","noteUrl":"/review?paper=review-neurips-2026-139727","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139729","title":"RamanBench: A Large-Scale Benchmark for Machine Learning on Raman Spectroscopy","authors":["Mario Koddenbrock","Christoph Lange","Robin Legner","Martin Jaeger","Martin Kögler","Mariano N Bournazou","Peter Neubauer","Felix Biessmann","Erik Rodner"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RamanBench: A Large-Scale Benchmark for Machine Learning on Raman Spectroscopy","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.02003","noteUrl":"/review?paper=review-neurips-2026-139729","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139730","title":"DriveHierarchy: A Benchmark for Diagnosing VLM Driving Capabilities from Open-Loop Understanding to Closed-Loop Execution","authors":["Chengkai Xu","Jiaqi Liu","Yicheng Guo","Peng Hang","Jian Sun"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"DriveHierarchy: A Benchmark for Diagnosing VLM Driving Capabilities from Open-Loop Understanding to Closed-Loop Execution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31814","noteUrl":"/review?paper=review-neurips-2026-139730","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139735","title":"NEST: Nascent Encoded Steganographic Thoughts","authors":["Artem Karpov"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"NEST: Nascent Encoded Steganographic Thoughts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.14095","noteUrl":"/review?paper=review-neurips-2026-139735","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139736","title":"Is Your LLM-as-a-Recommender Agent Trustable? LLMs' Recommendation is Easily Hacked by Biases (Preferences)","authors":["Zichen TANG","Zirui Zhang","Qian Wang","Zhenheng Tang","Xiaowen Chu","Bo Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Is Your LLM-as-a-Recommender Agent Trustable? LLMs' Recommendation is Easily Hacked by Biases (Preferences)","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.17417","noteUrl":"/review?paper=review-neurips-2026-139736","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139738","title":"PolarScale: A Physics-Grounded Benchmark for Radiometrically Consistent RGB-to-Stokes Estimation","authors":["Beibei Lin","Tingting Chen","Xin Zhang","Wenhao Zhao","DONGJUN LI","Zifeng Yuan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AoLP needs a quadrant-aware arctangent","plainExplanation":"Equation (3) sets \\(\\mathrm{AoLP}=\\tfrac12\\arctan(S_2/S_1)\\), while the supervised angle spans the stated half-turn orientation domain. For physical linear polarization, \\(S_1=p\\cos(2\\psi)\\) and \\(S_2=p\\sin(2\\psi)\\).","primaryUrl":"https://arxiv.org/html/2610.08346","noteUrl":"/review?paper=review-neurips-2026-139738","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139739","title":"Self Driving Datasets: From 20 Million Papers to Nuanced Biomedical Knowledge at Scale","authors":["Haydn Jones","Yimeng Zeng","Alden Rose","Yifei Li","Yining Huang","Kaiwen Wu","Jiaming Liang","Maggie Huan","Yoseph Barash","Cesar de la Fuente-Nunez","Osbert Bastani","Zachary Ives","Mark Yatskar","Jacob Gardner"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Self Driving Datasets: From 20 Million Papers to Nuanced Biomedical Knowledge at Scale","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07022","noteUrl":"/review?paper=review-neurips-2026-139739","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139740","title":"LUMOS: Tracing Parametric Knowledge from Training Data to Behavioral Outputs in LLMs","authors":["Seoyeon Ye","Gayoung Kim","Jiyoung Hong","Soo Kyung Kim","Hyunsoo Cho"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Probe confidence is not probe accuracy","plainExplanation":"The test statistic omits incorrect answers, so it cannot identify test classification accuracy or discrimination. Development-set layer selection does not turn this different test estimand into an accuracy measure.","primaryUrl":"https://arxiv.org/html/2610.02902","noteUrl":"/review?paper=review-neurips-2026-139740","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139742","title":"VIPER: An Expert-Curated Benchmark for Vision-Language Models in Veterinary Pathology","authors":["Luca Weishaupt","Simone de Brot","Javier Asin","Llorenç Grau-Roma","Nic G Reitsam","Andrew H Song","Dongmin Bang","Stefan T Kaluziak","Long P Le","Jakob N Kather","Faisal Mahmood","Guillaume Jaume"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"VIPER: An Expert-Curated Benchmark for Vision-Language Models in Veterinary Pathology","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.26382","noteUrl":"/review?paper=review-neurips-2026-139742","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139743","title":"SLS-Bench: A Benchmark for Incident Log Summarization with Synthetic Observability Data","authors":["Valerii Iakovlev","Markus Spanring","Martin Flechl"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SLS-Bench: A Benchmark for Incident Log Summarization with Synthetic Observability Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16412","noteUrl":"/review?paper=review-neurips-2026-139743","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139744","title":"MONET: A Massive, Open, Non-redundant and Enriched Text-to-image dataset","authors":["Benjamin Aubin","Gonzalo Iñaki Quintana","Onur Tasar","Sanjeev Sreetharan","Urszula Czerwinska","Damien Henry","Clément Chadebec"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MONET: A Massive, Open, Non-redundant and Enriched Text-to-image dataset","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.21272","noteUrl":"/review?paper=review-neurips-2026-139744","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139746","title":"PotARCin: Multi-Dimensional Evaluation of Skill Acquisition in Abstract Reasoning Tasks","authors":["Claas Beger","Ryan Yi","Melanie Mitchell"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"PotARCin: Multi-Dimensional Evaluation of Skill Acquisition in Abstract Reasoning Tasks","plainExplanation":"No issue was found in the checked claims. The lower-bound sentence summarizes the observed comparison on the tested models and datasets. Pairwise positive correlations alone would not imply the corresponding distribution-free probability inequality, but the source does not state that theorem.","primaryUrl":"https://arxiv.org/html/2609.27288","noteUrl":"/review?paper=review-neurips-2026-139746","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139747","title":"Are We Making Progress in Multimodal Domain Generalization? A Comprehensive Benchmark Study","authors":["Hao Dong","Hongzhao Li","Shupan Li","Muhammad Haris Khan","Eleni Chatzi","Olga Fink"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Are We Making Progress in Multimodal Domain Generalization? A Comprehensive Benchmark Study","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06643","noteUrl":"/review?paper=review-neurips-2026-139747","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139748","title":"FlyAOC: Evaluating Agentic Ontology Curation of Drosophila Scientific Knowledge Bases","authors":["Xingjian Zhang","Sophia Moylan","Ziyang Xiong","Qiaozhu Mei","Yichen Luo","Jiaqi Ma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FlyAOC: Evaluating Agentic Ontology Curation of Drosophila Scientific Knowledge Bases","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09163","noteUrl":"/review?paper=review-neurips-2026-139748","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139749","title":"MOVEBENCH: A Benchmark for Global-Scale Wildlife Movement Forecasting","authors":["Justin Kay","Shir Bar","Ellen O Aikens","Martin Becker","Francesca Cagnacci","Juliet Cohen","Scott W Forrest","Jessica Kendall-Bar","Madeleine Lucas","Macon Overcast","Meredith S Palmer","Will Rogers","Nicholas J Russo","Christian Rutz","Larissa T Beumer","Michael B Brown","Ying-Chi Chan","Sarah C Davidson","Diego E Soto","Anne G Hertel","Roland Kays","Benjamin Koger","Guram Mikaberidze","Thomas Mueller","Ruth Oliver","Thorsten Papenbrock","Robert Patchett","Jared A Stabach","Dane Taylor","Scott W Yanco","Sara Beery"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MOVEBENCH: A Benchmark for Global-Scale Wildlife Movement Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.15780","noteUrl":"/review?paper=review-neurips-2026-139749","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139753","title":"Aligning Language Model Benchmarks with Pairwise Preferences","authors":["Marco Gutierrez","Xinyi Leng","Hannah Chen","Jonathan Richard Schwarz","Ahmed Alaa","Tom Hartvigsen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Signed-weight and pair-count normalization","plainExplanation":"The signed sum is not guaranteed positive. Dividing by a negative sum reverses every nonzero score difference; a zero sum makes the formula undefined.","primaryUrl":"https://arxiv.org/html/2602.02898","noteUrl":"/review?paper=review-neurips-2026-139753","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139754","title":"PHOEBI: An Open-World Benchmark for Bacterial Identification in Phase-Contrast Microscopy","authors":["Aaditya Baranwal","Md Jahid Hasan","Shruti Vyas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PHOEBI: An Open-World Benchmark for Bacterial Identification in Phase-Contrast Microscopy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.22890","noteUrl":"/review?paper=review-neurips-2026-139754","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139755","title":"MotionHalluc: Diagnosing Kinematic Hallucinations in Fine-Grained Motion Reasoning","authors":["Weile Guo","Shenghong He","Danying Mo","Chengdong Xu","Xuexun Liu","Chao Yu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"MotionHalluc: Diagnosing Kinematic Hallucinations in Fine-Grained Motion Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.23061","noteUrl":"/review?paper=review-neurips-2026-139755","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139756","title":"PRISM-Bench: A Benchmark of Puzzle-Based Visual Tasks with CoT Error Detection","authors":["Cheng Wan","Yusu Qian","Qingyu Zhao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PRISM-Bench: A Benchmark of Puzzle-Based Visual Tasks with CoT Error Detection","plainExplanation":"","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2510.23594","noteUrl":"/review?paper=review-neurips-2026-139756","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139760","title":"Expected Harm: Rethinking Safety Evaluation of (Mis)Aligned LLMs","authors":["Yen-Shan Chen","Zhi Rui Tam","Cheng-Kuang Wu","Yun-Nung (Vivian) Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Joint completion and probe-identification limits","plainExplanation":"Making each factor larger than direct-query success does not make their product larger. The number of required subtasks matters even if their outcomes are independent.","primaryUrl":"https://arxiv.org/html/2602.01600","noteUrl":"/review?paper=review-neurips-2026-139760","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139761","title":"SMMBench: A Benchmark for Source-Distributed Multimodal Agent Memory","authors":["Huacan Chai","YUKAI WANG","Yingxuan Yang","Dan Peng","Yuanyi Song","Zhihui Fu","Jun Wang","Weiwen Liu","Jianghao Lin","Weinan Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SMMBench: A Benchmark for Source-Distributed Multimodal Agent Memory","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15710","noteUrl":"/review?paper=review-neurips-2026-139761","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139762","title":"SPDEBench: An Extensive Benchmark for Learning Stochastic PDEs","authors":["Yuantu Zhu","Zheyan Li","Dai Shi","Luke Thompson","Oliver Nash","Jose M Rangel","Siran Li","Bingguang Chen","Rongchan Zhu","Qi Meng","Hao Ni"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Finite expected signatures need not separate path laws","plainExplanation":"The finite-order signature discrepancy need not be a metric on path laws: distinct distributions can have identical expected finite features and hence zero discrepancy. The separation axiom fails for unrestricted path laws at fixed truncation order. The finite feature comparison remains usable, and this argument does not challenge infinite expected-signature uniqueness under its hypotheses.","primaryUrl":"https://arxiv.org/html/2505.18511","noteUrl":"/review?paper=review-neurips-2026-139762","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-139765","title":"VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models","authors":["Alex S Huang","Jiahui Zhang","Shiqing Tang","Yu Xiang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.20774","noteUrl":"/review?paper=review-neurips-2026-139765","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139768","title":"DeEscalWild: A Real-World Benchmark for Automated De-Escalation Training with SLMs","authors":["Md Hasebul Hasan","Krity H Charu","Eshwara Prasad Sridhar","Shuchisnigdha Deb","Mohammad A Islam"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DeEscalWild: A Real-World Benchmark for Automated De-Escalation Training with SLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.13075","noteUrl":"/review?paper=review-neurips-2026-139768","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139770","title":"ContractBench: Can LLM Agents Preserve Observation Contracts?","authors":["Jicheng Wang","Yifeng He","Zili Wang","Hanwen Xing","Arkaprava De","Hao Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Four contract outcomes require an admissible time outside the window","plainExplanation":"The proposition claims all four integrity/temporal-validity combinations for every admissible contract with a nonempty validity window and nontrivial artifact predicate. The definitions allow admissible submission times \\(T\\) to equal the validity window \\(W=[0,1)\\).","primaryUrl":"https://arxiv.org/html/2605.17281","noteUrl":"/review?paper=review-neurips-2026-139770","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139771","title":"LithoBench: Benchmarking Large Multimodal Models for Remote-Sensing Lithology Interpretation","authors":["jun wang","Fengpeng Li","Tianjin Huang","Hang Dong","Wei Han"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LithoBench: Benchmarking Large Multimodal Models for Remote-Sensing Lithology Interpretation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07640","noteUrl":"/review?paper=review-neurips-2026-139771","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139772","title":"OmniTraffic: A Controllable Generation Pipeline and Benchmark for Spatio-Temporal Traffic Reasoning","authors":["Maonan Wang","Zhengyan Huang","Kemou Jiang","Yuhang Fu","Jiayue Zhu","Yuxin Cai","Xingchen Zou","Qiaosheng Zhang","Yi Yu","Ding Wang","XI CHEN","Ben Chen","Yuxuan Liang","Zhiyong Cui","Man On Pun","Yirong Chen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"OmniTraffic: A Controllable Generation Pipeline and Benchmark for Spatio-Temporal Traffic Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.15749","noteUrl":"/review?paper=review-neurips-2026-139772","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139774","title":"Can LLM Agents Respond to Disasters? Benchmarking Heterogeneous Geospatial Reasoning in Emergency Operations","authors":["Junjue Wang","Weihao Xuan","Heli Qi","Pengyu Dai","Kunyi Liu","Hongruixuan Chen","Zhuo Zheng","Junshi Xia","Stefano Ermon","Naoto YOKOYA"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Can LLM Agents Respond to Disasters? Benchmarking Heterogeneous Geospatial Reasoning in Emergency Operations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11633","noteUrl":"/review?paper=review-neurips-2026-139774","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139776","title":"XL-SafetyBench: A Country-Grounded Cross-Cultural Benchmark for LLM Safety and Cultural Sensitivity","authors":["Dasol Choi","Eugenia Kim","Jaewon Noh","Sang Seo","Eunmi Kim","Myunggyo Oh","Yunjin Park","Brigitta J Kartono","Josef Pichlmeier","Helena Berndt","Sai Krishna Mendu","Glenn J Tungka","Özlem Gökçe","Suresh Gehlot","Katherine Pratt","Amanda Minnich","Haon Park"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"XL-SafetyBench: A Country-Grounded Cross-Cultural Benchmark for LLM Safety and Cultural Sensitivity","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.05662","noteUrl":"/review?paper=review-neurips-2026-139776","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139777","title":"StreamGaze: Gaze-Guided Temporal Reasoning and Proactive Understanding in Streaming Videos","authors":["Daeun Lee","Subhojyoti Mukherjee","Branislav Kveton","Ryan Rossi","Viet Lai","Seunghyun Yoon","Trung Bui","Franck Dernoncourt","Mohit Bansal"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Subtracting only the current fixation does not select never-fixated objects","plainExplanation":"The never-fixated candidate set subtracts the current fixation set \\(O_i\\) from the scene-wide union \\(A\\). An object fixated earlier but absent from \\(O_i\\) remains in \\(A\\setminus O_i\\).","primaryUrl":"https://arxiv.org/html/2512.01707","noteUrl":"/review?paper=review-neurips-2026-139777","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139778","title":"Benchmark for Assessing Olfactory Perception of Large Language Models","authors":["Eftychia Makri","Nikolaos Nakis","Laura Sisson","Geetanjali Minsky","Leandros Tassiulas","Vahid Satarifard","Nicholas A Christakis"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Benchmark for Assessing Olfactory Perception of Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.00002","noteUrl":"/review?paper=review-neurips-2026-139778","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139780","title":"SpatialBench: Is Your Spatial Foundation Model an All-Round Player","authors":["Haosong Peng","Hao Li","jiaqi chen","Yuhao Pan","Runmao Yao","Yalun Dai","Fushuo Huo","Fangzhou Hong","Zhaoxi Chen","Haozhao Wang","Dingwen Zhang","Ziwei Liu","Wenchao Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"World-to-camera relative poses require the opposite product order","plainExplanation":"The pose convention is \\(x_i=G_i x_{\\rm world}\\), but the relative translation formula uses \\(G_i^{-1}G_j\\). Under this convention \\(x_j=G_jG_i^{-1}x_i\\), with translation \\(t_j-R_jR_i^\\top t_i\\).","primaryUrl":"https://arxiv.org/html/2605.27367","noteUrl":"/review?paper=review-neurips-2026-139780","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139785","title":"EmoPhone: A Multi-Wave Dataset for In-the-Wild Mobile and Wearable Affect Sensing","authors":["Panyu Zhang","minseo Park","Soowon Kang","Ismatzoda Tomiris","Azizbek Mustafakulov","Otabek Najimov","Woohyeok Choi","JUMABEK Alikhanov","Surjya Ghosh","Uichin Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EmoPhone: A Multi-Wave Dataset for In-the-Wild Mobile and Wearable Affect Sensing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.16581","noteUrl":"/review?paper=review-neurips-2026-139785","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139786","title":"Latent Video Prediction for World Modeling: An Evaluation Uncovering Intriguing Favorable Evidence","authors":["Ali J Alrasheed","Aryan Y Parast","Basim Azam","James Bailey","NAVEED AKHTAR"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Latent Video Prediction for World Modeling: An Evaluation Uncovering Intriguing Favorable Evidence","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15618","noteUrl":"/review?paper=review-neurips-2026-139786","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139788","title":"PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments","authors":["Ruoqi Liu","Imran Mohiuddin","Austin J Schoeffler","Kavita Renduchintala","Ashwin Nayak","Prasantha L Vemu","Shivam Vedak","Kameron C Black","John L Havlik","Isaac Ogunmola","Stephen P.","Roopa Dhatt","Jonathan H Chen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.02240","noteUrl":"/review?paper=review-neurips-2026-139788","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139789","title":"EsoLang-Bench: Evaluating Large Language Models via Esoteric Programming Languages","authors":["Aman Sharma","Paras Chopra"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"EsoLang-Bench: Evaluating Large Language Models via Esoteric Programming Languages","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.09678","noteUrl":"/review?paper=review-neurips-2026-139789","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139793","title":"Benchmarking Optimizers for Large Language Model Pretraining","authors":["Andrei Semenov","Matteo Pagliardini","Martin Jaggi"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The right rotation needs the right covariance eigenbasis","plainExplanation":"An eigenbasis of \\(GG^\\top\\) has shape \\(m\\times m\\), while the right basis must have shape \\(n\\times n\\). The initialization pseudocode is incorrect for matrix gradients; actual optimizer runs are not assessed.","primaryUrl":"https://arxiv.org/html/2509.01440","noteUrl":"/review?paper=review-neurips-2026-139793","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139794","title":"NARRA-Gym for Evaluating Interactive Narrative Agents","authors":["Yue Huang","Yuchen Ma","Jiayi Ye","Wenjie Wang","Zipeng Ling","Xingjian Hu","Yuexing Hao","Zichen Chen","Zhangchen Xu","Yunhong He","Zhengqing Yuan","Yujun Zhou","Kehan Guo","Chaoran Chen","Toby Li","Stefan Feuerriegel","Xiangliang Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"NARRA-Gym for Evaluating Interactive Narrative Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08503","noteUrl":"/review?paper=review-neurips-2026-139794","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139795","title":"PhysEval.Weather: An Evaluation Framework for Physical Consistency in ML Weather Models","authors":["Emma Kasteleyn","Timo Maier","Axel Lauer","Veronika Eyring","Pierre Gentine","Ana Lucic"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PhysEval.Weather: An Evaluation Framework for Physical Consistency in ML Weather Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.10642","noteUrl":"/review?paper=review-neurips-2026-139795","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139796","title":"Tactile MNIST: Benchmarking Active Tactile Perception","authors":["Tim Schneider","Guillaume Duret","Cristiana de Farias","Roberto Calandra","Liming Chen","Jan Peters"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Tactile MNIST: Benchmarking Active Tactile Perception","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2506.06361","noteUrl":"/review?paper=review-neurips-2026-139796","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139798","title":"When Are Multimodal Predictions Biologically Supported? A Diagnostic Evaluation Framework","authors":["Dylan Steiner","Gustavo Arango","Gerald J Sun","Etai Jacob"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Raw cross-covariance SVD is not CCA without whitening","plainExplanation":"Raw cross-covariance SVD favors covariance magnitude, while CCA optimizes variance-normalized correlation. They need not retain the same directions.","primaryUrl":"https://arxiv.org/html/2605.31504","noteUrl":"/review?paper=review-neurips-2026-139798","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139799","title":"Where Tabular Foundation Models Falter on Genetic Data: Datasets That Expose and Provide a Path to Address the Gap","authors":["ANIRBAN DAS","Yan Cui"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The noise formula does not generally achieve cluster heritability","plainExplanation":"Within-cluster heritability depends on the within-cluster genetic variance \\(V_k\\), not the population-wide variance \\(V\\). The printed calibration reaches \\(h_k^2\\) only when those variances agree.","primaryUrl":"https://www.biorxiv.org/content/biorxiv/early/2026/09/28/2026.09.25.754520.full.pdf","noteUrl":"/review?paper=review-neurips-2026-139799","collection":"neurips-2026","severity":"repairable","issueTypes":["calibration_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139802","title":"ReaLM: A Unified Red-Teaming Benchmark for Physical-World VLMs","authors":["Yifei Zhao","Qian Lou","Mengxin Zheng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ReaLM: A Unified Red-Teaming Benchmark for Physical-World VLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.23892","noteUrl":"/review?paper=review-neurips-2026-139802","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139804","title":"GUITAR: Structured Failure Diagnosis of GUI Agents via State Transitions","authors":["Shaoqing Zhang","Kehai Chen","Xuefeng Bai","Zhuosheng Zhang","Pengfei Zhang","Yang Xiang","Min zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GUITAR: Structured Failure Diagnosis of GUI Agents via State Transitions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34113","noteUrl":"/review?paper=review-neurips-2026-139804","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139805","title":"EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective","authors":["Yuyao Wang","Zhongjian Zhang","Mo Chi","Kaichi Yu","Yuhan Li","Miao Peng","Bing Tong","Chen Zhang","Yan Zhou","Jia Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18421","noteUrl":"/review?paper=review-neurips-2026-139805","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139806","title":"Dream-Cubed: Controllable Generative Modeling in Minecraft by Training on Billions of Cubes","authors":["Tim Merino","Sam Earle","Ryunosuke Iwai","Julian Togelius","Edoardo Cetin"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Dream-Cubed: Controllable Generative Modeling in Minecraft by Training on Billions of Cubes","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.22847","noteUrl":"/review?paper=review-neurips-2026-139806","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139807","title":"URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment","authors":["Bogdan Zagribelnyy","Ivan Ilin","Nikita Bondarev","Anton Morgunov","Arkadii Lin","Maksim Kuznetsov","Rim Shayakhmetov","Vlad Aladinskiy","Alex Aliper","Alex Zhavoronkov"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.04688","noteUrl":"/review?paper=review-neurips-2026-139807","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139808","title":"FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs","authors":["David Restrepo","Chenwei Wu","Luis Nakayama","Miguel Martins","Stergios Christodoulidis","Maria Vakalopoulou","Enzo Ferrante"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33158","noteUrl":"/review?paper=review-neurips-2026-139808","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139809","title":"VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents","authors":["Zirui Wang","Junyi Zhang","Jiaxin Ge","Long Lian","Max Fu","Lisa Dunlap","Ken Goldberg","XuDong Wang","Ion Stoica","David Chan","Sewon Min","Joseph Gonzalez"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.16973","noteUrl":"/review?paper=review-neurips-2026-139809","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139814","title":"StemBind: When MLLMs Get Lost Between Rules and Instances in Abstract Visual Reasoning","authors":["Xixiang He","Baiqi Wu","Xingming Li","Ao Cheng","Qiyao Sun","Xuanyu Ji","Qingyong Hu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"StemBind: When MLLMs Get Lost Between Rules and Instances in Abstract Visual Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.00148","noteUrl":"/review?paper=review-neurips-2026-139814","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139815","title":"Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale","authors":["Amit Roth","Ankur Samanta","Matan Halevy","Yoav Levine","Yonathan Efroni"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20744","noteUrl":"/review?paper=review-neurips-2026-139815","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139816","title":"What DNA Foundation Models Learn Beyond Sequence Composition","authors":["Vincenzo Y. Civale","Andrew Bagdanov","Alberto Magi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"What DNA Foundation Models Learn Beyond Sequence Composition","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.15711","noteUrl":"/review?paper=review-neurips-2026-139816","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139818","title":"MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI","authors":["Paula Arguello","Berk Tinaz","Mohammad Shahab Sepehri","Maryam Soltanolkotabi","Mahdi Soltanolkotabi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.11762","noteUrl":"/review?paper=review-neurips-2026-139818","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139819","title":"The Alignment Curse: Modality Alignment Supercharges Audio Attacks via Text Transfer","authors":["Yupeng Chen","Junchi Yu","Aoxi Liu","Baoyuan Wu","Philip Torr","Adel Bibi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Projected KL cannot certify an upper bound on full-representation KL","plainExplanation":"The transfer theorem needs full-representation KL, while the empirical procedure estimates KL after a fifteen-dimensional PCA projection. Data processing gives \\(D_{\\rm KL}(\\Pi P\\|\\Pi Q)\\leq D_{\\rm KL}(P\\|Q)\\), the wrong direction for the needed upper bound.","primaryUrl":"https://arxiv.org/html/2602.02557","noteUrl":"/review?paper=review-neurips-2026-139819","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139821","title":"VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing","authors":["Zhiming Luo","Hebaixu Wang","Haonan Guo","Jing Zhang","Bo Du"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.07045","noteUrl":"/review?paper=review-neurips-2026-139821","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139822","title":"\\$OneMillion-Bench: How Far are Language Agents from Human Experts?","authors":["Yang Liu","Jiaqi Li","Jun Bai","Qianyu Yang","Zixia Jia","Tiliang Duan","Chun Zhang","Jiayun Dong","Lingyue Yin","Jianpeng Jiao","Yanglihong Xiao","Zaiyuan Wang","Tao Peng","Xiaobo Hu","Kaiyuan Chen","Ge Zhang","Gang Yao","Hao Chen","Zilong Zheng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"\\$OneMillion-Bench: How Far are Language Agents from Human Experts?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.07980","noteUrl":"/review?paper=review-neurips-2026-139822","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139823","title":"MMTA: Benchmarking Multimodal Temporal Analysis with Time Series, Text, and Vision","authors":["Ziyang Zhang","Shenyi Li","Yilin wang","Ziyun Cui","Bowen Zhou","Wen Wu","Chao Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MMTA: Benchmarking Multimodal Temporal Analysis with Time Series, Text, and Vision","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.17856","noteUrl":"/review?paper=review-neurips-2026-139823","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139825","title":"The Text Uncanny Valley: Non-Monotonic Performance Degradation in LLM Information Retrieval","authors":["Zekai Tong","Ruiyao Xu","Aryan Shrivastava","Chenhao Tan","Ari Holtzman"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Text Uncanny Valley: Non-Monotonic Performance Degradation in LLM Information Retrieval","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07186","noteUrl":"/review?paper=review-neurips-2026-139825","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139826","title":"The Illusion of Multi-Agent Advantage","authors":["Prathyusha Jwalapuram","Hehai Lin","Chuyuan Li","Fangkai Jiao","Sudong Wang","Yifei Ming","Zixuan Ke","Chengwei Qin","Giuseppe Carenini","Shafiq Joty"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Illusion of Multi-Agent Advantage","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.13003","noteUrl":"/review?paper=review-neurips-2026-139826","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139827","title":"ROK-FORTRESS: Measuring the Effect of Geopolitical Transcreation for National Security and Public Safety","authors":["Michael S Lee","Yash Maurya","Drew Rein","Bert Herring","Jonathan Nguyen","Kyungho Song","Udari Madhushani Sehwag","Jiyeon Cho","Kaustubh Deshpande","Yeongkyun Jang","Jiyeon Joo","Minn S Choi","Evi Fuelle","Christina Q Knight","Joseph Brandifino","Max Fenkell"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ROK-FORTRESS: Measuring the Effect of Geopolitical Transcreation for National Security and Public Safety","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14152","noteUrl":"/review?paper=review-neurips-2026-139827","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139828","title":"HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot Manipulation","authors":["Jaehwi Song","Suchae Jeong","Byeongguk Jeon","Sungdong Kim","Hyungmok Son","Kimin Lee"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot Manipulation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.31682","noteUrl":"/review?paper=review-neurips-2026-139828","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139830","title":"RedVLA: Physical Red Teaming for Vision-Language-Action Models","authors":["Yuhao Zhang","Borong Zhang","Jiaming Fan","Jiachen Shen","Yishuai Cai","Yaodong Yang","Jiaming Ji"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Existence of a successful path does not ensure victim-policy reward","plainExplanation":"The safety formulation constrains reward under a fixed victim policy, but the construction claims to enforce it by preserving some feasible benign solution. Existence of a safe successful path gives a statement about \\(\\max_\\pi J_R\\), not about \\(J_R(\\pi_\\theta)\\).","primaryUrl":"https://arxiv.org/html/2604.22591","noteUrl":"/review?paper=review-neurips-2026-139830","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139831","title":"Offline Reinforcement Learning for Plasma Control in Nuclear Fusion: Codebase and Benchmark","authors":["YANG FU","Haomin Bao","Rohit Sonker","Xiaoyan Hu","Aravind Venugopal","Jeff Schneider","Jiayu Chen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Offline Reinforcement Learning for Plasma Control in Nuclear Fusion: Codebase and Benchmark","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.07550","noteUrl":"/review?paper=review-neurips-2026-139831","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139833","title":"VistaQA: Benchmarking Joint Visual Question Answering and Pixel-Level Evidence","authors":["Mozhgan Nasr Azadani","Yimu Wang","Yongpeng Zhu","Lihong Chen","Milan Ganai","Sean Sedwards","Marco Pavone","Krzysztof Czarnecki"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"VistaQA: Benchmarking Joint Visual Question Answering and Pixel-Level Evidence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20676","noteUrl":"/review?paper=review-neurips-2026-139833","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139834","title":"ProcObject-10K: Benchmarking Object-Centric Procedural Understanding in Instructional Videos","authors":["Wenliang Guo","Yu Kong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The auxiliary objective and interval metric need local corrections","plainExplanation":"Under loss minimization with positive auxiliary weights, these signs encourage disagreement with the pseudo-labels. The printed terms become arbitrarily negative as a positive label receives probability approaching zero.","primaryUrl":"https://arxiv.org/abs/2512.03479v2","noteUrl":"/review?paper=review-neurips-2026-139834","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139837","title":"Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action","authors":["Ben Slater","Lucy G Cheke","John Burden","Winnie Street"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The false-belief odds contrast depends on task size","plainExplanation":"The interaction with continuous task size remains in the false-versus-true log-odds contrast, so a single size-independent magnitude is incorrect.","primaryUrl":"https://arxiv.org/html/2606.31916","noteUrl":"/review?paper=review-neurips-2026-139837","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139838","title":"MT-JailBench: A Modular Benchmark for Understanding Multi-Turn Jailbreak Attacks","authors":["Xinkai Zhang","Zhipeng Wei","Huanli Gong","Jing Ting Zheng","Yuchen Zhang","Yue Dong","N. Benjamin Erichson"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MT-JailBench: A Modular Benchmark for Understanding Multi-Turn Jailbreak Attacks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11002","noteUrl":"/review?paper=review-neurips-2026-139838","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139839","title":"Benchmarking Attention for Tabular Foundation Models","authors":["Maximilian Schambach","Clemens Biehl","Sam Thelin"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Benchmarking Attention for Tabular Foundation Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31306","noteUrl":"/review?paper=review-neurips-2026-139839","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139841","title":"StaminaBench: Stress-Testing Coding Agents over 100 Interaction Turns","authors":["Uladzislau Sobal","Shuo Yang","Yuting Zhang","Wei Xia","Stefano Soatto"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"StaminaBench: Stress-Testing Coding Agents over 100 Interaction Turns","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.19613","noteUrl":"/review?paper=review-neurips-2026-139841","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139844","title":"Faithfulness Metrics Don't Measure Faithfulness: A Meta-Evaluation with Ground Truth","authors":["Yoav Gur-Arieh","Ana Marasovic","Mor Geva"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A correct transition count does not identify every intermediate computation","plainExplanation":"An even run from \\(x\\) has \\(\\nu_2(x)\\) transitions and can be evaluated with one trailing-zero count and a shift, adding \\(\\nu_2(x)\\) to the total without materializing each state.","primaryUrl":"https://arxiv.org/html/2605.25052","noteUrl":"/review?paper=review-neurips-2026-139844","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139850","title":"RoboWits: Unexpected Challenges for Robotic Creative Problem Solving","authors":["Chunru Lin","Hongxin Zhang","Fenghao Yu","Zhehuan Chen","Tom Griffiths","Yejin Choi","David Held","Chuang Gan"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"RoboWits: Unexpected Challenges for Robotic Creative Problem Solving","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.30326","noteUrl":"/review?paper=review-neurips-2026-139850","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-139853","title":"DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules","authors":["Devin Y De Silva","Dhaval Patel","Christodoulos Constantinides","Shuxin Lin","Nianjun Zhou","Dharmashankar Subramanian","Paul J Adams","Sal Rosato","Nicolas Constantinides","Deborah L McGuinness","Jayant Kalagnanam"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules","plainExplanation":"The sole retained mathematical row concerns the normalization of option overlap. In v2 it appears in Appendix A.4 and Figure 8, with the same pair-count mismatch. Existing dataset and evaluation reporting exclusions are preserved.","primaryUrl":"https://arxiv.org/html/2605.08614","noteUrl":"/review?paper=review-neurips-2026-139853","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-139854","title":"When Guessing is Rewarded: Rethinking Language Model Evaluation with Distributional Uncertainty Scoring","authors":["Tom Burns"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Abstention changes the maximum DCS and removes the claimed negative ceiling","plainExplanation":"The proof maximizes \\(DCS=(p_c-P_W)(1-p_{\\rm idk})\\) at fixed \\(p_c\\) by setting abstention to zero. Since \\(1-p_{\\rm idk}=p_c+P_W\\), \\(DCS=p_c^2-P_W^2\\).","primaryUrl":"https://arxiv.org/html/2510.04302","noteUrl":"/review?paper=review-neurips-2026-139854","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-139994","title":"Explainability matters: The effect of liability rules on the healthcare sector","authors":["Jiawen Wei","Elena Verona","Andrea Bertolini","Gianmarco Mengaldo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Explainability matters: The effect of liability rules on the healthcare sector","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.17334","noteUrl":"/review?paper=review-neurips-2026-139994","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-139995","title":"Comparing Explanations is not Enough,Explain the Change: New Standards are Needed to Explain Behavioral Shifts in Large Language Models","authors":["Martino Ciaperoni","Marzio Di Vece","Roberto Pellungrini","Luca Pappalardo","Fosca Giannotti","Francesco Giannini"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Comparing Explanations is not Enough,Explain the Change: New Standards are Needed to Explain Behavioral Shifts in Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.02304","noteUrl":"/review?paper=review-neurips-2026-139995","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140000","title":"Risks Create a Jagged Frontier of LLM Productivity Gains Across Computer Occupations","authors":["Deepika Chawla","Gagandeep Singh","Elham k buxton","Meicen Sun","Lav Varshney","Jeremy Riel","Craig De Voto"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Risks Create a Jagged Frontier of LLM Productivity Gains Across Computer Occupations","plainExplanation":"","primaryUrl":"https://papers.ssrn.com/sol3/Delivery.cfm/6381580.pdf?abstractid=6381580&mirid=1","noteUrl":"/review?paper=review-neurips-2026-140000","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140001","title":"We Need to Rethink Benchmarking in Anomaly Detection","authors":["Philipp Röchner","Simon Klüttermann","Kevin Kammler","Franz Rothlauf","Emmanuel Müller","Daniel Schlör"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"We Need to Rethink Benchmarking in Anomaly Detection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2507.15584","noteUrl":"/review?paper=review-neurips-2026-140001","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140005","title":"Evaluation of Visual Processing Should Be Human-Centered, Not Metric-Centered","authors":["Fanghua Yu","Jinfan Hu","Zhiyuan You","Xiang Yin","Hongyu An","Xinqi Lin","Hongyang Li","Chao Dong","Jinjin Gu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Evaluation of Visual Processing Should Be Human-Centered, Not Metric-Centered","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.00643","noteUrl":"/review?paper=review-neurips-2026-140005","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140006","title":"Watermarking Should Be Treated as a Monitoring Primitive","authors":["Toluwani Aremu","Jie Zhang","Nils Lukas"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Watermarking Should Be Treated as a Monitoring Primitive","plainExplanation":"No issue was found in the checked claims. The current formula handles fewer than three classes and resolves the boundary error.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2605.13095","noteUrl":"/review?paper=review-neurips-2026-140006","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140011","title":"Mechanistic Interpretability Needs Philosophy","authors":["Iwan Williams","Ninell Oldenburg","Ruchira Dhar","Joshua Hatherley","Constanza Fierro","Nina Rajcic","Sandrine R Schiller","Filippos Stamatiou","Anders Søgaard"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mechanistic Interpretability Needs Philosophy","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2506.18852","noteUrl":"/review?paper=review-neurips-2026-140011","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140020","title":"Position: Life-Logging Video Streams Make the Privacy–Utility Trade-off Inevitable","authors":["Tianyuan Zou","Liang Yue","Yang Liu","Ya-Qin Zhang","Sijie Cheng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Position: Life-Logging Video Streams Make the Privacy–Utility Trade-off Inevitable","plainExplanation":"A minimal sufficient representation can discard all task-irrelevant information and still allow highly accurate reconstruction. Reconstruction resistance needs a specified distortion target and an information bound tied to that target; the paper already leaves formal privacy guarantees open.","primaryUrl":"https://arxiv.org/html/2605.10404","noteUrl":"/review?paper=review-neurips-2026-140020","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-140023","title":"Epistemic Infrastructures of Science in AI Era Should Rebalance Costs of Generation and Verification","authors":["Jiaqi Ma"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Epistemic Infrastructures of Science in AI Era Should Rebalance Costs of Generation and Verification","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10425v1","noteUrl":"/review?paper=review-neurips-2026-140023","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140030","title":"Large Language Model Failures from Hallucination to Homogenization Are Different Facets of Miscalibration","authors":["Tiancheng Hu","Caiqi Zhang","Dirk Hovy","Nigel Collier"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Large Language Model Failures from Hallucination to Homogenization Are Different Facets of Miscalibration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://zenodo.org/api/records/18499744/files/Calibration_Position_Paper_preprint.pdf/content","noteUrl":"/review?paper=review-neurips-2026-140030","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140033","title":"Agentic AI Scientists Are Not Built For Autonomous Scientific Discovery","authors":["Harshit Bisht","Vinay Kumar","Kevin Maik Jablonka","Mausam","N M Anoop Krishnan"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Agentic AI Scientists Are Not Built For Autonomous Scientific Discovery","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08956","noteUrl":"/review?paper=review-neurips-2026-140033","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140034","title":"Concept-Based Mechanistic Interpretability Needs a Concrete Evaluation Paradigm","authors":["Yiming Tang","Qinglin Qi","LIN Zheng","Dianbo Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Concept-Based Mechanistic Interpretability Needs a Concrete Evaluation Paradigm","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://zenodo.org/records/20444131/files/Concept_Based_Mechanitic_Interpretability_Needs_a_Concrete_Evaluation_Paradigm.pdf?download=1","noteUrl":"/review?paper=review-neurips-2026-140034","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140037","title":"A New Perspective on XAI: Scientific Theory Building for Auditable Artefacts","authors":["Sebastian R. Müller","Vanessa Toborek","Tamas Horvath","Brendan B Jackson","Christian Bauckhage"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A New Perspective on XAI: Scientific Theory Building for Auditable Artefacts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2602.02215v1","noteUrl":"/review?paper=review-neurips-2026-140037","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140044","title":"Watermarking Without Standards Is Not AI Governance","authors":["Alexander Nemecek","Yuzhou Jiang","Erman Ayday"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Watermarking Without Standards Is Not AI Governance","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2505.23814","noteUrl":"/review?paper=review-neurips-2026-140044","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140047","title":"Standardization of Post-Publication Code Verification is Possible with the Support of the Community","authors":["Susana López-Moreno","Eric R Dolores-Cuenca","Sangil Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Standardization of Post-Publication Code Verification is Possible with the Support of the Community","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.07189","noteUrl":"/review?paper=review-neurips-2026-140047","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140048","title":"Position: Let’s Strengthen Verifiability if We Can’t Enforce Reproducibility","authors":["Samet Hicsonmez","Nermin Samet","Renaud Marlet"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Position: Let’s Strengthen Verifiability if We Can’t Enforce Reproducibility","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.35854","noteUrl":"/review?paper=review-neurips-2026-140048","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140051","title":"Embracing Evolution: A Call for Body-Control Co-Design in Embodied Humanoid Robot","authors":["Guiliang Liu","Bo Yue","Yi J Kim","Kui Jia"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Embracing Evolution: A Call for Body-Control Co-Design in Embodied Humanoid Robot","plainExplanation":"No issue was found in the checked claims. The earlier objection relied on compact morphology alone. The revision now assumes that the inner policy maximum is well-defined and the resulting value function is continuous.","primaryUrl":"https://arxiv.org/html/2510.03081","noteUrl":"/review?paper=review-neurips-2026-140051","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140053","title":"Beyond One-Size-Fits-All: Diagnosis-Driven Online Reinforcement Learning with Offline Priors","authors":["Guozheng Ma","Lu Li","Zilin Wang","Pierre-Luc Bacon","Dacheng Tao"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Beyond One-Size-Fits-All: Diagnosis-Driven Online Reinforcement Learning with Offline Priors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.25527","noteUrl":"/review?paper=review-neurips-2026-140053","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140060","title":"Computer Science Conferences Should Require Nonrepudiable Experimental Results","authors":["Mamadou K KEITA","Christopher Homan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The execution-binding reduction needs an authenticated signing domain","plainExplanation":"Signature unforgeability protects messages not legitimately signed by the service. It does not prove that an authorized signature certifies the execution origin of its input. The bound therefore needs a signing-interface premise in addition to key separation.","primaryUrl":"https://arxiv.org/html/2605.08586","noteUrl":"/review?paper=review-neurips-2026-140060","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-140061","title":"No One Knows the State-of-the-Art in Geospatial Foundation Models","authors":["Isaac Corley","Caleb Robinson","Nils Lehmann","Gabriel Tseng","Anthony Fuller","Hamed Alemohammad","Evan Shelhamer","Jennifer Marcus","Hannah Kerner"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"No One Knows the State-of-the-Art in Geospatial Foundation Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12678","noteUrl":"/review?paper=review-neurips-2026-140061","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140063","title":"Causality can systematically address the monsters under the bench(marks)","authors":["Felix Leeb","Zhijing Jin","Bernhard Schölkopf"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Zero average label effect is not a fifty-percent flip rate","plainExplanation":"For binary labels, \\(\\mathbb E[Y_{\\mathrm{edit}}-Y_{\\mathrm{base}}]=\\Pr(0\\to1)-\\Pr(1\\to0)\\), while flip probability is the sum of those probabilities.","primaryUrl":"https://arxiv.org/html/2502.05085","noteUrl":"/review?paper=review-neurips-2026-140063","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-140070","title":"When Helpfulness Becomes Sycophancy: Sycophancy is a Boundary Failure Between Social Alignment and Epistemic Integrity in Large Language Models","authors":["Jiechen Li","Catherine A Barry","Rishika Randev","Janet Chen","Ella Jorgensen","Brinnae Bent"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"When Helpfulness Becomes Sycophancy: Sycophancy is a Boundary Failure Between Social Alignment and Epistemic Integrity in Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.05403","noteUrl":"/review?paper=review-neurips-2026-140070","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140074","title":"AI Safety Evaluations Need More Human-AI Experiments","authors":["Michelle Vaccaro","Jaeyoon Song","Abdullah Almaatouq","Michiel Bakker"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AI Safety Evaluations Need More Human-AI Experiments","plainExplanation":"Human–AI uplift ratios are invariant to a change of units, but can change or reverse order when the score origin changes. Cross-scale comparisons therefore require a justified measurement scale and calibration, as distinct from a pragmatic ratio under a fixed coding.","primaryUrl":"https://arxiv.org/html/2603.26676","noteUrl":"/review?paper=review-neurips-2026-140074","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-140075","title":"Auditing is not Evaluating: LLM Audit Requires Dynamic, Contextual, Budget-aware and Reliable Evidence","authors":["Cléa Chataigner","Pablo Piantanida","Golnoosh Farnadi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Auditing is not Evaluating: LLM Audit Requires Dynamic, Contextual, Budget-aware and Reliable Evidence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://hal.science/hal-05498201v1/file/Position_Auditing_Is_Not_Evaluating.pdf","noteUrl":"/review?paper=review-neurips-2026-140075","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140077","title":"Allocate Marginal Reviews to Borderline Papers Using LLM Comparative Ranking","authors":["Elliot Epstein","Rajat Vadiraj Dwaraknath","John Winnicki","Thanawat Sornwanee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The confidence interval treats dependent removals as independent trials","plainExplanation":"The confidence interval treats dependent removals as independent trials.","primaryUrl":"https://arxiv.org/html/2602.06078","noteUrl":"/review?paper=review-neurips-2026-140077","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-140079","title":"AI Evaluation Should Require Standardized Item-Level Data Releases","authors":["Han Jiang","Susu Zhang","Dongyao Zhu","Yuzhuo Bai","Sang Truong","Xiaoyuan Yi","Sanmi Koyejo","Xing Xie","Ziang Xiao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AI Evaluation Should Require Standardized Item-Level Data Releases","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.03244","noteUrl":"/review?paper=review-neurips-2026-140079","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140086","title":"Position: Adopt Constraints Over Fixed Penalties in Deep Learning","authors":["Juan Ramirez","Seyed Meraj Hashemizadehaghda","Simon Lacoste-Julien"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Position: Adopt Constraints Over Fixed Penalties in Deep Learning","plainExplanation":"No issue was found in the checked claims. The alleged target-shift defect is not present in the current source. The retained Lagrangian subtracts the target explicitly.","primaryUrl":"https://arxiv.org/html/2505.20628","noteUrl":"/review?paper=review-neurips-2026-140086","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140088","title":"Autonomous Driving Research Requires a Community-Driven Data Paradigm","authors":["Jinsu Yoo","Zanming Huang","Katie Luo","Zheda Mai","Qiyuan Wu","Bharath Hariharan","Mark Campbell","Wei-Lun (Harry) Chao"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Autonomous Driving Research Requires a Community-Driven Data Paradigm","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.08825","noteUrl":"/review?paper=review-neurips-2026-140088","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-140090","title":"Beyond Training Time, Test-Time Coordination is Essential for Cooperative MARL","authors":["Dongsu Lee","Sooraj Sathish","Joonkyung Kim","Woojun Kim","Amy Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond Training Time, Test-Time Coordination is Essential for Cooperative MARL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.17409","noteUrl":"/review?paper=review-neurips-2026-140090","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140092","title":"From Static Policies to Adaptive Priors in Offline Reinforcement Learning","authors":["Tianwei Ni","Vineet Jain","Akash Karthikeyan","Pierre-Luc Bacon"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"From Static Policies to Adaptive Priors in Offline Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35880","noteUrl":"/review?paper=review-neurips-2026-140092","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140094","title":"Trustworthy AI Must Account for Interactions","authors":["Jesse Cresswell"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The conformal examples need inclusive ties and a distinction between minimum and equal coverage","plainExplanation":"Exchangeability permits tied scores. The strict comparison excludes true labels tied with the threshold, so the usual order-statistic coverage argument does not apply.","primaryUrl":"https://arxiv.org/html/2504.07170","noteUrl":"/review?paper=review-neurips-2026-140094","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-140100","title":"AI Agents Push Humans Out of the Loop","authors":["Margaret Mitchell","Avijit Ghosh","Samir Passi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AI Agents Push Humans Out of the Loop","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.23642","noteUrl":"/review?paper=review-neurips-2026-140100","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140101","title":"The Aleatoric-Epistemic Dichotomy of Uncertainty is Meaningful and Indispensable for Machine Learning","authors":["Yusuf Sale","Nikita Kotelevskii","Maxim Panov","Eyke Hüllermeier"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"The Aleatoric-Epistemic Dichotomy of Uncertainty is Meaningful and Indispensable for Machine Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://papers.ssrn.com/sol3/Delivery.cfm/6783338.pdf?abstractid=6783338&mirid=1","noteUrl":"/review?paper=review-neurips-2026-140101","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140105","title":"Position: Semantic Uncertainty Measures Disagreement, Not Reliability","authors":["Joseph Hoche","Maxime Corlay","David Brellmann","Andrei Bursuc","Pavel Izmailov","Angela Yao","Gianni Franchi"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Position: Semantic Uncertainty Measures Disagreement, Not Reliability","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://hal.science/hal-05633506v1/file/_PositionPaper_UQ_Visual_Foundation_Model__Copy_for_Neurips___arxiv_.pdf","noteUrl":"/review?paper=review-neurips-2026-140105","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140106","title":"Agent Security is a Systems Problem","authors":["Mihai Christodorescu","Earlence Fernandes","Ashish Hooda","Somesh Jha","Johann Rehberger","Kamalika Chaudhuri","Xiaohan Fu","Khawaja Shams","Guy Amir","Jihye Choi","Sarthak Choudhary","Nils Palumbo","Andrey Labunets","Nishit V Pandya"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Agent Security is a Systems Problem","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18991","noteUrl":"/review?paper=review-neurips-2026-140106","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140107","title":"Graph Learning Should Move Beyond Restrictive Views of Spectral and Message-Passing GNNs","authors":["Antonis Vasileiou","Juan Cervino","Pascal Frossard","Charilaos Kanatsoulis","Christopher Morris","Michael T Schaub","Pierre Vandergheynst","Zhiyang Wang","Guy Wolf","Ron Levie"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The 1-WL lower bound requires a suitable graph shift","plainExplanation":"Theorem 3 compares functional-calculus spectral GNNs with 1-WL. Its statement quantifies over every aggregation graph shift operator, while Appendix E.5 establishes the matching lower bound for the adjacency operator. Definition 2 only requires that multiplication by the graph shift can be implemented by a finite sum-aggregation MPNN. This allows shifts that erase graph information. Theorem 3 therefore cannot infer the full distinguishing power of 1-WL for every such shift.","primaryUrl":"https://arxiv.org/html/2602.10031","noteUrl":"/review?paper=review-neurips-2026-140107","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-140108","title":"Wasserstein Gradient Flows and Forward-Only Diffusion Are Not Enough for Multimodal Sampling","authors":["Daniel McBride","Pratik Khandagale","Cristina Garcia-Cardona","Yen Ting Lin"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Wasserstein Gradient Flows and Forward-Only Diffusion Are Not Enough for Multimodal Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02081","noteUrl":"/review?paper=review-neurips-2026-140108","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140111","title":"Position: The Road to Generalizable Neuro-Symbolic Learning Should be Paved with Foundation Models","authors":["Adam Stein","Aaditya Naik","Neelay Velingker","Mayur Naik","Eric Wong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Position: The Road to Generalizable Neuro-Symbolic Learning Should be Paved with Foundation Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.24874","noteUrl":"/review?paper=review-neurips-2026-140111","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140112","title":"Looking Under the Streetlight: Evaluation in Generative Molecular Dynamics","authors":["Simon Olsson","Frank Noe","Grant Rotskoff","Kresten Lindorff-Larsen"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Looking Under the Streetlight: Evaluation in Generative Molecular Dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.12518","noteUrl":"/review?paper=review-neurips-2026-140112","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140117","title":"Alignment Needs 'Cognitive Control': On The Role of Regularization in LLM Alignment","authors":["Nona Ghazizadeh","Elnaz Rahmati","Morteza Dehghani","Payam Piray"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Alignment Needs 'Cognitive Control': On The Role of Regularization in LLM Alignment","plainExplanation":"Increasing the KL penalty moves the optimal policy toward its reference distribution, but need not increase policy entropy when that reference is nonuniform. The closed-form optimum and the entropy monotonicity of its normalized exponential component remain valid.","primaryUrl":"https://drive.google.com/uc?export=download&id=1Lk26QX3XxQiBSATEHzOjhgMovHSd-DFp","noteUrl":"/review?paper=review-neurips-2026-140117","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-140119","title":"Understanding Generalization Requires Universal Induction","authors":["Aram Ebtekar","Marcus Hutter","Danica J. Sutherland"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Understanding Generalization Requires Universal Induction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34458","noteUrl":"/review?paper=review-neurips-2026-140119","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-140125","title":"Understanding diffusion models requires rethinking (again) generalization","authors":["Pierre Marion","Yu-Han Wu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Understanding diffusion models requires rethinking (again) generalization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06077","noteUrl":"/review?paper=review-neurips-2026-140125","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148139","title":"DyCoRM: Dynamic Criterion-Aware Reward Modeling for Text-to-Image Generation","authors":["Jiaying Qian","Ziheng Jia","Qian Zhang","Zicheng Zhang","Jiayi Guo","Junqi Zhang","Guangtao Zhai","Xiongkuo Min"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"DyCoRM: Dynamic Criterion-Aware Reward Modeling for Text-to-Image Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.25876","noteUrl":"/review?paper=review-neurips-2026-148139","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148140","title":"Improving Audit Realism with Inference-Time Compute and Deployment Scaffolds","authors":["Axel Ahlqvist","Richard Guan","Juan-Pablo Rivera","Adeline Kassler","Alexandra Souly","Kai Fronsdal","Robert Kirk","John Hughes"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Improving Audit Realism with Inference-Time Compute and Deployment Scaffolds","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.02302","noteUrl":"/review?paper=review-neurips-2026-148140","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148141","title":"RADAR: Relative Angular Divergence Across Representations","authors":["Xavier Cadet","Mateusz Nowak","Peter Chin"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The regularization reverses the central-sample weighting","plainExplanation":"The regularization reverses the central-sample weighting. Total-variation contraction does not control every representation metric.","primaryUrl":"https://arxiv.org/html/2605.23028","noteUrl":"/review?paper=review-neurips-2026-148141","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148142","title":"Flow Matching Policy Optimization with Mirror Descent and Entropy Constraints","authors":["Ting Gao","Stavros Orfanoudakis","Nan Lin","Winnie Daamen","Serge Hoogendoorn","Elvin Isufi"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The entropy estimate needs the learned flow’s time marginals","plainExplanation":"Equal endpoints do not imply equal intermediate laws. Hutchinson’s estimator is conditionally unbiased for the trace at its evaluation point; it does not correct a mismatch in the distribution of that point.","primaryUrl":"https://arxiv.org/html/2603.17685","noteUrl":"/review?paper=review-neurips-2026-148142","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148146","title":"Learning Visual Feature-Based World Models via Residual Latent Action","authors":["Xinyu Zhang","Zhengtong Xu","Yutian Tao","Yeping Wang","Yu She","Abdeslam Boularias"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Visual Feature-Based World Models via Residual Latent Action","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07079","noteUrl":"/review?paper=review-neurips-2026-148146","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148152","title":"Graph Sparse Sampling: Breaking the Curse of the Horizon in Continuous MDP Planning","authors":["Idan Lev-Yehudi","Vadim Indelman"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Graph Sparse Sampling: Breaking the Curse of the Horizon in Continuous MDP Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.05359","noteUrl":"/review?paper=review-neurips-2026-148152","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148153","title":"Bridge Graphical Models: Coupling, Projection, and Current-Preserving Dynamics for Generative Modeling","authors":["Tiantian Zhang"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Bridge Graphical Models: Coupling, Projection, and Current-Preserving Dynamics for Generative Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.19144","noteUrl":"/review?paper=review-neurips-2026-148153","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148155","title":"FMMI: Flow Matching Mutual Information Estimation","authors":["Ivan Butakov","Alexander Semenenko","Valeriia Kirova","Ivan Oseledets","Alexey Frolov"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Smooth compact support is incompatible with a globally Lipschitz log density","plainExplanation":"Let \\(U=\\{p>0\\}\\), choose \\(y\\in U\\), and take its nearest boundary point \\(b\\). The segment from \\(y\\) to \\(b\\) lies in \\(U\\), while continuity gives \\(p(b)=0\\). A finite log-Lipschitz constant would bound \\(\\log p\\) below on that segment by \\(\\log p(y)-L\\|b-y\\|\\), contradicting its divergence to \\(-\\infty\\). The literal hypothesis class is empty.\n\nReading smoothness only inside a support with a discontinuous boundary does not validate the proof: uniform density on \\([0,1]\\) has interior log-Lipschitz constant zero, but vector field \\(\\epsilon(x)=x\\) gives \\(|\\mathbb E\\operatorname{div}\\epsilon|=1\\). The boundary flux discarded by integration by parts is then nonzero. The literal assumptions make the affected guarantees vacuous; the permissive support-only reading loses the asserted bound. The Fisher-information Cauchy–Schwarz step and empirical estimators are not themselves refuted.","primaryUrl":"https://arxiv.org/pdf/2511.08552v2","noteUrl":"/review?paper=review-neurips-2026-148155","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148156","title":"Sparse Layers are Critical to Scaling Looped Language Models","authors":["Ryan Lee","Jacob Biloki","Edward J Hu","Jonathan May"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Sparse Layers are Critical to Scaling Looped Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.09165v2","noteUrl":"/review?paper=review-neurips-2026-148156","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148157","title":"Selective Safety Steering via Value-Filtered Decoding","authors":["Bat-Sheva Einbinder","Hen Davidov","Yee Whye Teh","Yarin Gal","Yaniv Romano"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Selective Safety Steering via Value-Filtered Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14746","noteUrl":"/review?paper=review-neurips-2026-148157","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148158","title":"Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models","authors":["Junxiang Qiu","Shuo Wang","Zhengsu Chen","Hengheng Zhang","Jinda Lu","Changcheng Li","Xinting Hu","Qi Tian"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.02819","noteUrl":"/review?paper=review-neurips-2026-148158","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148160","title":"RABBiT: Rapidly adaptive BOLD foundation model via brain-tuning for accurate zero-shot and few-shot prediction of speech-elicited responses in the brain","authors":["Omer Moussa","Mariya Toneva"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Averaging responses and averaging principal directions give different initializations","plainExplanation":"Averaging responses and averaging principal directions give different initializations.","primaryUrl":"https://arxiv.org/html/2607.05171","noteUrl":"/review?paper=review-neurips-2026-148160","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148161","title":"Scaling Laws and Tradeoffs in Recurrent Networks of Expressive Neurons","authors":["Aaron Spieler","Georg Martius","Anna Levina"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Scaling Laws and Tradeoffs in Recurrent Networks of Expressive Neurons","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12049","noteUrl":"/review?paper=review-neurips-2026-148161","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148163","title":"Beyond Worst-Case Coreset Bounds for $k$-Clustering via Determinantal Sampling","authors":["Diptarka Chakraborty","Satyaki Mukherjee","Gaurav Vallabhdas Revankar","Hoang Son Tran"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The uniform coreset event needs a union over centers","plainExplanation":"A bound for one fixed center need not bound the probability that some center has a large error. The displayed proposition drops this distinction. Section 3.1 already takes a union over a finite net and only then fixes a center, so the defect is local to the proposition’s wording and normalization. This finding concerns Proposition A.1’s uniform-to-fixed-center inequality, its real-exponent proof, and the normalization in Equation (5). Section 3.1 already retains the finite-net union. The example does not contradict Theorem 3.1; the correction establishes its local reductions, without proving the separate assumptions used in later specializations.","primaryUrl":"https://arxiv.org/html/2609.06394","noteUrl":"/review?paper=review-neurips-2026-148163","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148164","title":"Learning Evidence Highlighting for Frozen LLMs","authors":["Shaoang Li","Yanhang Shi","Yufei Li","Mingfu Liang","Xiaohan Wei","Yuchen Pu","Fei Tian","Chonglin Sun","Frank Shyu","Luke Simon","Sandeep Pandey","Xi Liu","Jian Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Evidence Highlighting for Frozen LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.22565","noteUrl":"/review?paper=review-neurips-2026-148164","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148166","title":"Constrained Code Generation with Discrete Diffusion","authors":["Lize Shao","Michael Cardei","Zichen Xie","Ferdinando Fioretto","Wenxi Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Constrained Code Generation with Discrete Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16829","noteUrl":"/review?paper=review-neurips-2026-148166","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148167","title":"Follow the Winners: Conservative Policy Improvement with the Cross-Entropy Method for Critic-Free RFT","authors":["Joery A de Vries","Neil Lawrence","Zhenwen Dai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Follow the Winners: Conservative Policy Improvement with the Cross-Entropy Method for Critic-Free RFT","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03361","noteUrl":"/review?paper=review-neurips-2026-148167","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148169","title":"Variable-Length Generative Protein Design via Generalized Poisson Flow","authors":["Chaoran Cheng","Zhanghan Ni","Yanru Qu","Yuxin Chen","Ruihan Guo","Jiajun Fan","Ge Liu"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Variable-Length Generative Protein Design via Generalized Poisson Flow","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.09039","noteUrl":"/review?paper=review-neurips-2026-148169","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148170","title":"CanvasMAR: Improving Masked Autoregressive Video Prediction With Canvas","authors":["Zian Li","Muhan Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CanvasMAR: Improving Masked Autoregressive Video Prediction With Canvas","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.13669","noteUrl":"/review?paper=review-neurips-2026-148170","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148172","title":"SyncLight: Single-Edit Multi-View Relighting","authors":["David Serrano-Lozano","Anand Bhattad","Luis Herranz","Jean-Francois Lalonde","Javier Vazquez-Corral"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SyncLight: Single-Edit Multi-View Relighting","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.16981","noteUrl":"/review?paper=review-neurips-2026-148172","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148175","title":"Scaling Causal Reasoning with Increasingly Complex Causal Simulators","authors":["Nicolás Astorga","Anita Kriz","Mihaela van der Schaar"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Enumeration of sampled domains need not recover the full answer support","plainExplanation":"Enumeration of sampled domains need not recover the full answer support.","primaryUrl":"https://arxiv.org/html/2605.09079","noteUrl":"/review?paper=review-neurips-2026-148175","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148177","title":"LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving","authors":["Yuechen Luo","Fang Li","Shaoqing Xu","Yang Ji","Zehan Zhang","Bing Wang","Shen Yuannan","Jianwei Cui","Long Chen","Guang Chen","Hangjun Ye","Zhi-Xin Yang","Fuxi Wen"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.01928","noteUrl":"/review?paper=review-neurips-2026-148177","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148178","title":"Mixture-of-Control: State-Aware Fine-Tuning for Transformer-based Models","authors":["Duc Anh Nguyen","Tien N Luu","Tung Pham","Toan Tran"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Mixture-of-Control: State-Aware Fine-Tuning for Transformer-based Models","plainExplanation":"The frozen blocks can already differ nonlinearly when all weight perturbations are zero. A linear state control cannot generally represent that residual. The appendix’s mixing-closeness constants are independent of \\(\\xi\\), so its residual bound does not vanish merely because \\(\\xi\\) does.","primaryUrl":"https://arxiv.org/html/2606.31397","noteUrl":"/review?paper=review-neurips-2026-148178","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148181","title":"ReGenHuman: Re-Generating Human Appearances for Realistic Full-Body Video Anonymization","authors":["Adam Sun","Eshaan Barkataki","Arnold Milstein","Gordon Wetzstein","Ehsan Adeli"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ReGenHuman: Re-Generating Human Appearances for Realistic Full-Body Video Anonymization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.14972","noteUrl":"/review?paper=review-neurips-2026-148181","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148184","title":"The Value of Covariance Matching in Gaussian DDPMs and the Lanczos Sampler","authors":["Sahil Akhtar","Aymane El Gadarri","Vivek Farias","Adam Jozefiak"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Value of Covariance Matching in Gaussian DDPMs and the Lanczos Sampler","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22723","noteUrl":"/review?paper=review-neurips-2026-148184","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148186","title":"FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning","authors":["Yequan Zhao","Ruijie (Ray) Zhang","Liyan Tan","Niall Moran","Tong Qin","Zheng Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning","plainExplanation":"No issue was found in the checked claims. The original application, lemma heading and proof all specify the Gaussian ensemble. Orthogonal invariance and exchangeability prove the baseline for that ensemble. A non-Gaussian example tests a broader claim obtained by dropping this context; it does not invalidate the result used in the paper.","primaryUrl":"https://arxiv.org/html/2605.22869","noteUrl":"/review?paper=review-neurips-2026-148186","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148189","title":"Markovian Dynamics Enforcer: Feasibility Preserving Correction on Learned Dyanmics Manifolds","authors":["Kevin Yu","Tao Guo","Constantinos Antoniou","Panagiotis Angeloudis"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Markovian Dynamics Enforcer: Feasibility Preserving Correction on Learned Dyanmics Manifolds","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39888","noteUrl":"/review?paper=review-neurips-2026-148189","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148190","title":"B[FM]$^2$: Brain Foundation Model via Flow Matching with SplitUNet","authors":["Jaedong Hwang","Kathleen Zhang","David Dai","Konstantinos Kontras","Maarten De Vos","Ila Fiete","Paul Liang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"B[FM]$^2$: Brain Foundation Model via Flow Matching with SplitUNet","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.20812","noteUrl":"/review?paper=review-neurips-2026-148190","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148191","title":"Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention","authors":["Ali Hatamizadeh","Yejin Choi","Jan Kautz"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22791","noteUrl":"/review?paper=review-neurips-2026-148191","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148192","title":"What Was That Again? Certified Robustness for Automatic Speech Recognition","authors":["Andrew Cullen","Neil Marchant","Jiani Xie","Paul Montague","Benjamin Rubinstein"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The displayed confidence-sequence inversion always returns one half","plainExplanation":"In the last paragraph of Section 3.1, let $m$ be the number of tested hypotheses, $\\alpha\\in(0,1)$ the target false discovery rate, and $E_{(1)}\\ge\\cdots\\ge E_{(m)}\\ge0$ the ordered e-values. The paragraph describes e-BH as selecting the largest index $k$ satisfying the displayed partial-sum threshold. It then explicitly excludes e-BH from the implementation. Because e-values are nonnegative, a qualifying partial sum forces $k=m$ to qualify, so the stated largest-index rule rejects every hypothesis or none. One large e-value from a false hypothesis can then force rejection of all true nulls. The cited e-BH method instead thresholds the $k$th largest e-value.","primaryUrl":"https://arxiv.org/html/2606.27698","noteUrl":"/review?paper=review-neurips-2026-148192","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148194","title":"Smooth Partial Lotteries for Stable Randomized Selection","authors":["Alexander Goldberg","Giulia Fanti","Nihar Shah"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Smooth Partial Lotteries for Stable Randomized Selection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20069","noteUrl":"/review?paper=review-neurips-2026-148194","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148195","title":"Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization","authors":["Parthiv Chatterjee","Kashish Kanjariya","Vashisht Purani","Sourish Dasgupta","Tanmoy Chakraborty"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01375","noteUrl":"/review?paper=review-neurips-2026-148195","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148196","title":"Neuronal Self-Adaptation Enhances Capacity and Robustness of Representation in Spiking Neural Networks","authors":["Zhuobin Yang","Liao Yu","Yeyao Bao","Lv L Fu","Daqing Guo","Jian Zhang","Xiaohong Li","Yunliang Zang"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Neuronal Self-Adaptation Enhances Capacity and Robustness of Representation in Spiking Neural Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.20687","noteUrl":"/review?paper=review-neurips-2026-148196","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148197","title":"Fast Reconstruction of Exact Maxwell Dynamics from Sparse Data","authors":["Dan DeGenaro","Xin Li","Obed Amo","Michael Pokojovy","Sarah Bargal","Markus Lange-Hegermann","Bogdan Raita"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Odd time derivatives need the corresponding Laplacian power","plainExplanation":"The odd-derivative identity uses one Laplacian at every order. Repeated Maxwell differentiation requires \\(\\Delta^j\\) at order \\(2j+1\\), so the printed identity is generally false. This is a local exponent typo. The subsequent differential-operator order \\(2j+1\\) already matches the corrected identity; the argument does not refute the Maxwell construction.","primaryUrl":"https://arxiv.org/html/2605.20514","noteUrl":"/review?paper=review-neurips-2026-148197","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148198","title":"EP-Flow: Disordered Crystal Structure Prediction without Site-level Annotation","authors":["liu qiuliang","Liming Wu","Qi Li","Zhonglong Peng","Chang Chen","Wenbing Huang","Xiaolong Chen","Shifeng Jin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EP-Flow: Disordered Crystal Structure Prediction without Site-level Annotation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01315","noteUrl":"/review?paper=review-neurips-2026-148198","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148199","title":"FedRSPO+: A Heterogeneity-aware Algorithm for Decision-focused Federated Learning","authors":["Konstantinos Ziliaskopoulos","Alexander Vinel","Jiaqi Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FedRSPO+: A Heterogeneity-aware Algorithm for Decision-focused Federated Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.09091","noteUrl":"/review?paper=review-neurips-2026-148199","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148203","title":"How does RL Post-training Induce Skill Composition? A Case Study on Countdown","authors":["Simon Park","Simran Kaur","Sanjeev Arora"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"How does RL Post-training Induce Skill Composition? A Case Study on Countdown","plainExplanation":"No issue was found in the checked claims. The current evaluation count agrees with the stated combinatorial recurrence and resolves the retained mismatch.","primaryUrl":"https://arxiv.org/html/2512.01775","noteUrl":"/review?paper=review-neurips-2026-148203","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148209","title":"Sandboxed Coding Agents are Competitive Omni-modal Task Solvers","authors":["Dongping Chen","Xuanao Huang","Zhihan Hu","Qingyuan Shi","Dianqi Li","Tianyi Zhou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Sandboxed Coding Agents are Competitive Omni-modal Task Solvers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.00579","noteUrl":"/review?paper=review-neurips-2026-148209","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148210","title":"Train the Agent, Not the Expert: Learning to Harness Heterogeneous Experts for Multi-Turn Visual Reasoning","authors":["Yaowu Fan","Tao Han","Dazhao Du","Jinhua Ma","Jia Wan"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Train the Agent, Not the Expert: Learning to Harness Heterogeneous Experts for Multi-Turn Visual Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.29894","noteUrl":"/review?paper=review-neurips-2026-148210","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148211","title":"AdaM-Rec: Adaptive Modality Routing for Multimodal Recommendation","authors":["Honghao Fu","Jiacheng Chen","Manxi Lin","Junjun Zheng","Xiangheng Kong","Yiwei Wang","Xin Yu","Miao Xu","Yuning Jiang","Yujun Cai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AdaM-Rec: Adaptive Modality Routing for Multimodal Recommendation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38455","noteUrl":"/review?paper=review-neurips-2026-148211","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148212","title":"Ensembling Language Models with Sequential Monte Carlo","authors":["Robin Chan","Tianyu Liu","Samuel Kiegeland","Clemente Pasti","Jacob Vigly","Timothy O&amp;#x27;Donnell","Ryan Cotterell","Tim Vieira"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The importance weight uses the updated prefix","plainExplanation":"Appending the sampled symbol before evaluating its weight changes the conditioning prefix required by the proposition. The defect concerns the pseudocode update order; the proposition supplies the correct prefix convention.","primaryUrl":"https://arxiv.org/html/2603.05432","noteUrl":"/review?paper=review-neurips-2026-148212","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148214","title":"Discriminative Score Function: Turning Pretrained Models into Functional Generative Priors","authors":["Junhoo Lee","Hyeonjin Kim","Sangbum Han","Nojun Kwak"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Discriminative Score Function: Turning Pretrained Models into Functional Generative Priors","plainExplanation":"No issue was found in the checked claims. The alleged missing self-similarity term comes from replacing the paper’s alignment objective with a different, squared-distance objective. That replacement is not required by the stated operational definition.","primaryUrl":"https://arxiv.org/html/2610.19743v1","noteUrl":"/review?paper=review-neurips-2026-148214","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148216","title":"Probe Before You Edit: Probing-Guided Molecular Optimization for LLM Agents in Structure-Based Drug Design","authors":["Zaifei Yang","Weiyu Chen","Yaqing Wang","James Kwok"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Probe Before You Edit: Probing-Guided Molecular Optimization for LLM Agents in Structure-Based Drug Design","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.00555","noteUrl":"/review?paper=review-neurips-2026-148216","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148217","title":"Learning Where and What to Restore for Composite Image Restoration","authors":["Jiachen Jiang","Tianyu Ding","Ke Zhang","Jinxin Zhou","Tianyi Chen","Ilya Zharkov","Zhihui Zhu","Luming Liang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Where and What to Restore for Composite Image Restoration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2503.17915","noteUrl":"/review?paper=review-neurips-2026-148217","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148219","title":"Characterizing Learning in Deep Neural Networks using a Tractable Algorithmic Complexity Estimator","authors":["Pedram Bakhtiarifard","Sophia Natasha Wilson","Mahmoud H. A. Afifi","Jonathan Wenshøj","Raghavendra Selvan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Characterizing Learning in Deep Neural Networks using a Tractable Algorithmic Complexity Estimator","plainExplanation":"Being different from a previous plane does not imply being incompressible given the retained planes. The unconditional strict-improvement assertion is stronger than the theorem and its correctly qualified proof.","primaryUrl":"https://arxiv.org/html/2605.15551v1","noteUrl":"/review?paper=review-neurips-2026-148219","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","formula_error","invalid_inference","mathematical_error","notation","unsupported_claim"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148220","title":"LittleLearner: Language Models Under Pedagogically-Controlled Knowledge Exposure","authors":["Fanfei Li","Jana Zeller","Manuel Prada-Corral","Thaddäus Wiedemer","Prasanna Mayilvahanan","Ryan Cotterell","Wieland Brendel"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"LittleLearner: Language Models Under Pedagogically-Controlled Knowledge Exposure","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.13545","noteUrl":"/review?paper=review-neurips-2026-148220","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148222","title":"Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking","authors":["Solomon Messing"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Dispersion across a fixed judge panel is not panel-sampling variance","plainExplanation":"The estimand conditions on the selected finite judge levels but adds their fixed-effect dispersion divided by panel size to the sampling variance of their average. For that conditional estimand, the average fixed judge effect is a constant in the estimator expectation and contributes zero sampling variance.","primaryUrl":"https://arxiv.org/html/2604.11581","noteUrl":"/review?paper=review-neurips-2026-148222","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148225","title":"Visual-Redundancy-Controlled Parallel Decoding for Diffusion-Based Multimodal Large Language Models","authors":["Yulin Yuan","Hongshuo Zhao","Xiangming Meng"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Visual-Redundancy-Controlled Parallel Decoding for Diffusion-Based Multimodal Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25820","noteUrl":"/review?paper=review-neurips-2026-148225","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148226","title":"Efficient Matrix Product State Learning in Logarithmic Depth","authors":["Chia-Ying Lin","Nai-Hui Chia","Shih-Han Hung"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The tomography proof needs a consistent tolerance and a valid hidden-coordinate reduction","plainExplanation":"The stated algorithm permits twice the local error assumed in its guarantee. The proof’s accumulation bound therefore does not establish the advertised total error for the printed tolerance.","primaryUrl":"https://arxiv.org/html/2510.07798","noteUrl":"/review?paper=review-neurips-2026-148226","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","invalid_reduction","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148229","title":"Deep Probabilistic Supervision for Image Classification","authors":["Anton Adelöw","Matteo Gamba","Atsuto Maki"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Deep Probabilistic Supervision for Image Classification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.24162","noteUrl":"/review?paper=review-neurips-2026-148229","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148230","title":"Automated Reformulation of Robust Optimization via Memory-Augmented Large Language Models","authors":["Jinbiao Chen","Shuang Jin","Guoyun Zhang","Junyu Zhang","Guanyi Wang","Hanzhang Qin"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Automated Reformulation of Robust Optimization via Memory-Augmented Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11813","noteUrl":"/review?paper=review-neurips-2026-148230","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148231","title":"CodeMimicry: Exploiting Safety Generalization Lag in Large Language Models via Structured Code Completion","authors":["Liang Zhen","Wentao Chen","Hai Huang"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"CodeMimicry: Exploiting Safety Generalization Lag in Large Language Models via Structured Code Completion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39902","noteUrl":"/review?paper=review-neurips-2026-148231","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148232","title":"LatentRouter: Can We Choose the Right Multimodal Model Before Seeing Its Answer?","authors":["Xueqi Cheng","Yushun Dong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LatentRouter: Can We Choose the Right Multimodal Model Before Seeing Its Answer?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11301","noteUrl":"/review?paper=review-neurips-2026-148232","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148233","title":"PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety","authors":["Ding Jia","Wei Liu","Xianglong Du","Yingjie Li","Yingqing Yang","Huili Yu","Zhangsong Zhan","Chu Zhou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34415","noteUrl":"/review?paper=review-neurips-2026-148233","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148234","title":"SCOPE: Self-Play via Co-Evolving Policies for Open-Ended Tasks","authors":["Wai-Chung Kwan","Aryo Gema","Joshua O Leang","Pasquale Minervini"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An interior mean does not prevent reward-variance collapse","plainExplanation":"An interior mean does not prevent reward-variance collapse.","primaryUrl":"https://arxiv.org/html/2605.31433","noteUrl":"/review?paper=review-neurips-2026-148234","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148236","title":"HACO: Hedged Agent Computing for Reliable LLM Systems","authors":["Enhan Li","Hongyang Du"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"HACO: Hedged Agent Computing for Reliable LLM Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.19215","noteUrl":"/review?paper=review-neurips-2026-148236","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148237","title":"STRIDE: Learnable Stepwise Language Feedback for LLM Reasoning","authors":["Junjie Zhang","Guozheng Ma","Shunyu Liu","Zetian Hu","Yongcheng Jing","Ting-En Lin","Yongbin Li","Dacheng Tao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"STRIDE: Learnable Stepwise Language Feedback for LLM Reasoning","plainExplanation":"Group normalization does not discard failed continuations, and improving a redirected context can harm another context through shared parameters. All-failure groups still have zero task advantages.","primaryUrl":"https://arxiv.org/html/2605.18851","noteUrl":"/review?paper=review-neurips-2026-148237","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-148239","title":"Clapping: Removing Per-sample Storage for Pipeline Parallel Learning with Communication Compression","authors":["Boao Kong","Xu Huang","Yuqi Xu","Yixuan Liang","Bin Wang","Kun Yuan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Clapping: Removing Per-sample Storage for Pipeline Parallel Learning with Communication Compression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.19029v1","noteUrl":"/review?paper=review-neurips-2026-148239","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148240","title":"Christoffel-DPS: Optimal sensor placement in diffusion posterior sampling for arbitrary distributions","authors":["James Rowbottom","Zi Yuan (Nick) Huang","Carola-Bibiane Schönlieb","Ben Adcock"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Christoffel-DPS: Optimal sensor placement in diffusion posterior sampling for arbitrary distributions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06861","noteUrl":"/review?paper=review-neurips-2026-148240","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148241","title":"Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors","authors":["Max McGuinness","Alex Serrano Terre","Luke Bailey","Scott Emmons"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.11949","noteUrl":"/review?paper=review-neurips-2026-148241","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148244","title":"Character Mixing for Video Generation","authors":["Tingting Liao","Chongjian GE","Guangyi Liu","Hao Li","Yi Zhou"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Character Mixing for Video Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2510.05093v1","noteUrl":"/review?paper=review-neurips-2026-148244","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148245","title":"Rigel3D: Rig-aware Latents for Animation-Ready 3D Asset Generation","authors":["Nikitas Chatzis","Marios Loizou","Evangelos Kalogerakis"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Rigel3D: Rig-aware Latents for Animation-Ready 3D Asset Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13129","noteUrl":"/review?paper=review-neurips-2026-148245","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148249","title":"PhysGuard: Fisher-Guided Gradient Projection for Sim-to-Real Neural PDE Surrogates","authors":["Changjian Zhou","Junfeng Fang","Negin Yousefpour","peng wu","Bin Yan","Guillermo A Narsilio"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"PhysGuard: Fisher-Guided Gradient Projection for Sim-to-Real Neural PDE Surrogates","plainExplanation":"No issue was found in the checked claims. The specific formal curvature guarantee is withdrawn in the current manuscript. This does not infer a mathematical reason for the revision or evaluate the empirical preservation results.","primaryUrl":"https://arxiv.org/html/2606.16602","noteUrl":"/review?paper=review-neurips-2026-148249","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148251","title":"VisionCreator-R1: A Reflection-Enhanced Native Visual-Generation Agentic Model","authors":["Jinxiang Lai","Wenzhe Zhao","Zexin Lu","HUALEI Zhang","Qinyu Yang","Rongwei Quan","li zhimin","Shuai Shao","Song Guo","Qinglin Lu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Positive reflection noise does not establish variance dominance","plainExplanation":"Positive reflection noise does not establish variance dominance.","primaryUrl":"https://arxiv.org/html/2603.08812","noteUrl":"/review?paper=review-neurips-2026-148251","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148252","title":"Geometric Analysis of Neural Regression Collapse via Intrinsic Dimension","authors":["George Andriopoulos","Zixuan Dong","Bimarsha Adhikari","Keith Ross"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The regularized factorization threshold misses the feature-loss normalization","plainExplanation":"Theorem I.1 gives a global minimizer for squared reconstruction loss divided by \\(M\\), with separate unnormalized factor penalties. Differentiation gives \\(\\nabla_W L=(WH-Y)H^\\top/M+\\lambda_WW\\) and \\(\\nabla_H L=W^\\top(WH-Y)/M+\\lambda_HH\\).","primaryUrl":"https://arxiv.org/html/2510.01105","noteUrl":"/review?paper=review-neurips-2026-148252","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148253","title":"From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving","authors":["Sining ANG","Yuguang Yang","Chenxu Dang","Canyu Chen","Cheng Chi","Liu Haiyan","Xuanyao Mao","jason bao","Xuliang","Bingchuan Sun","Yan Wang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.10719","noteUrl":"/review?paper=review-neurips-2026-148253","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148254","title":"Hearing is Believing? Evaluating and Analyzing Audio Language Model Sycophancy with SYAUDIO","authors":["Junchi Yao","Lokranjan Lakshmikanthan","Annie Zhao","Danielle Zhao","Shu Yang","Zikang Ding","Di Wang","Lijie Hu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The dB settings do not encode the stated noise ratios","plainExplanation":"The noise experiment associates equal ten-decibel steps with half, equal and double noise volume and the listed physical SNR values. A dB gain \\(g\\) multiplies amplitude by \\(10^{g/20}\\) and power by \\(10^{g/10}\\).","primaryUrl":"https://arxiv.org/html/2601.23149","noteUrl":"/review?paper=review-neurips-2026-148254","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148256","title":"Direction-Aware Offline-to-Online Learning in Linear Contextual Bandits","authors":["Zean Han","Ruihan Lin","Zezhen Ding","Jiheng Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The Gaussian lower-bound model is explicit; its instance-family symbol needs a definition","plainExplanation":"An undeclared family symbol obscures the quantifier in the lower bound. It does not erase the explicit Gaussian and diagonal-covariance conditions already stated in the theorem.","primaryUrl":"https://arxiv.org/html/2604.24016","noteUrl":"/review?paper=review-neurips-2026-148256","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148258","title":"CTM-AI: A Blueprint for General AI Inspired by a Model of Consciousness","authors":["Haofei Yu","Yining Zhao","Lenore Blum","Manuel Blum","Paul Liang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CTM-AI: A Blueprint for General AI Inspired by a Model of Consciousness","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.04097","noteUrl":"/review?paper=review-neurips-2026-148258","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148260","title":"Stop the Sampler! Classifier-Based Adaptive Stopping for Sampling Kernels","authors":["Kirill Korolev","Nikita Morozov","Stepan Pavlenko","Esmeralda S Whitammer","Sergey Samsonov"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Incoming flow has a positive propagated-flow term","plainExplanation":"Eq. (43) subtracts propagated nonterminal flow from injected flow. Incoming flow is their sum; the incorrect sign contradicts the kernel integration used to derive it. The later flow equation already has the intended sign. This local display error does not establish a failure of the sampling theorem or implementation.","primaryUrl":"https://arxiv.org/html/2606.16073","noteUrl":"/review?paper=review-neurips-2026-148260","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148261","title":"Stabilizing the Dynamic Low-Rank Training","authors":["Zhonghan Xu","Ling Wang","Junhao Chen","Jianwei Zhao","Jinwei Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strict subspace inclusion does not imply strict residual contraction","plainExplanation":"Strict subspace inclusion does not imply strict residual contraction.","primaryUrl":"https://arxiv.org/html/2609.32615","noteUrl":"/review?paper=review-neurips-2026-148261","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148262","title":"POME: Post Optimization Model Edit via Muon-Style Projection","authors":["Yong Liu","di fu","Yang Luo","Zirui Zhu","Minhao Cheng","Cho-Jui Hsieh","Yang You"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A scalar rescaling coefficient is equated to a matrix","plainExplanation":"Eq. (12) equates a scalar rescaling coefficient to a matrix-valued expression. Multiplying a normalized matrix by scalar factors does not produce the scalar defined in Eq. (10). This is a type mismatch in the displayed definition. It does not justify replacing the stated scalar rescaling by an elementwise operation or alleging an implementation failure.","primaryUrl":"https://arxiv.org/html/2510.06627","noteUrl":"/review?paper=review-neurips-2026-148262","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148265","title":"Natural Policy Gradient as Doubly Smoothed Policy Iteration: A Bellman-Operator Framework","authors":["Phalguni Nanda","Zaiwei Chen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Natural Policy Gradient as Doubly Smoothed Policy Iteration: A Bellman-Operator Framework","plainExplanation":"The Bellman-operator update leaves a policy unconstrained at states that no transition reaches, so it does not ensure uniform value convergence for every admitted policy sequence. Requiring the statewise update already given in the paper repairs this gap and matches the natural policy gradient specialization.","primaryUrl":"https://arxiv.org/html/2605.10671","noteUrl":"/review?paper=review-neurips-2026-148265","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148266","title":"OneCanvas: 3D Scene Understanding via Panoramic Reprojection","authors":["Bartłomiej Baranowski","Dave Chen","Matthias Niessner"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OneCanvas: 3D Scene Understanding via Panoramic Reprojection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19253","noteUrl":"/review?paper=review-neurips-2026-148266","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148268","title":"CUDABench: Benchmarking LLMs for Text-to-CUDA Generation","authors":["Jiace Zhu","Wentao Chen","Qi Fan","Zhixing Ren","Junying Wu","Xing Z Chai","Chotiwit Rungrueangwutthinon","Yehan Ma","An Zou"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"CUDABench: Benchmarking LLMs for Text-to-CUDA Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2603.02236v1","noteUrl":"/review?paper=review-neurips-2026-148268","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148271","title":"Blocked Gibbs meets Diffusion Transformers: Unsupervised Learning for Constraint Optimization","authors":["Yudong Will Xu","Wenhao Li","Xiaoyu Wang","Scott Sanner","Elias Khalil"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Blocked Gibbs meets Diffusion Transformers: Unsupervised Learning for Constraint Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25129","noteUrl":"/review?paper=review-neurips-2026-148271","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148272","title":"Random-Set Graph Neural Networks","authors":["Tommy Woodley","Matteo Tolloso","Davide Bacciu","Shireen Kudukkil Manchingal","Fabio Cuzzolin"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Full-lattice Möbius coefficients do not invert an arbitrary focal family","plainExplanation":"Full-lattice Möbius coefficients do not invert an arbitrary focal family.","primaryUrl":"https://arxiv.org/html/2605.11987","noteUrl":"/review?paper=review-neurips-2026-148272","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148274","title":"Component-Based Out-of-Distribution Detection","authors":["Wenrui Liu","Hong Chang","RuiBing Hou","Shiguang Shan"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The discrete threshold need not achieve the stated target recall","plainExplanation":"The discrete threshold need not achieve the stated target recall.","primaryUrl":"https://arxiv.org/html/2604.21546","noteUrl":"/review?paper=review-neurips-2026-148274","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148277","title":"CRUMB: Efficient Prior Fitted Network Inference via Distributionally Matched Context Batching","authors":["Jamie Heredge","Mattia Jacopo Villani","Pranav Deshpande","Akshay Seshadri","Niraj Kumar"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"CRUMB: Efficient Prior Fitted Network Inference via Distributionally Matched Context Batching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11473","noteUrl":"/review?paper=review-neurips-2026-148277","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148278","title":"ELF: Embedded Language Flows","authors":["Keya Hu","Linlu Qiu","Yiyang Lu","Hanhong Zhao","Tianhong Li","Yoon Kim","Jacob Andreas","Kaiming He"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ELF: Embedded Language Flows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10938","noteUrl":"/review?paper=review-neurips-2026-148278","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148281","title":"Learning Scenario Reduction for Two-Stage Robust Optimization with Discrete Uncertainty","authors":["TIANJUE LIN","Jianan Zhou","Jieyi Bi","Yaoxin Wu","Wen Song","Zhiguang Cao","Jie Zhang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The zero-valued empty set needs a qualification in the monotonicity statement","plainExplanation":"The pointwise maximum argument proves monotonicity for nonempty scenario sets. It does not justify assigning zero to the empty set when a nonempty problem may have negative optimal value.","primaryUrl":"https://arxiv.org/html/2605.14494","noteUrl":"/review?paper=review-neurips-2026-148281","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148282","title":"SARL: Label-Free Reinforcement Learning by Rewarding Reasoning Topology","authors":["Yifan Wang","Bolian Li","David Cho","Ruqi Zhang","Fanping Sui","Ananth Grama"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Permitted graph cases leave the topology reward undefined","plainExplanation":"The reward averages branching-node and reachable-distinct-pair statistics, while the clustering fallback permits every reasoning step to share one default node. In the all-noise fallback graph, self-loops are omitted, so there are no edges, no degree-at-least-two nodes and no distinct reachable pairs.","primaryUrl":"https://arxiv.org/html/2603.27977","noteUrl":"/review?paper=review-neurips-2026-148282","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148283","title":"CELEUS: Certifiable and Efficient LLM Evaluation via E-Processes","authors":["Zhijian Zhou","Zesheng Ye","Zhaorun Chen","Bo Li","Feng Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A probability floor changes the variance-optimal sampler","plainExplanation":"Theorem 4 claims residual-proportional sampling minimizes conditional variance under \\(q_j\\geq\\beta/M\\) and \\(\\sum_jq_j=1\\). The relevant objective is \\(\\sum_jr_j^2/q_j\\).","primaryUrl":"https://arxiv.org/html/2606.20820","noteUrl":"/review?paper=review-neurips-2026-148283","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148284","title":"UnfoldArt: Zero-Shot Recovery of Full Articulated 3D Objects from Text or Image","authors":["Mohamed El Amine Boudjoghra","Ivan Laptev","Angela Dai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"UnfoldArt: Zero-Shot Recovery of Full Articulated 3D Objects from Text or Image","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30608","noteUrl":"/review?paper=review-neurips-2026-148284","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148286","title":"A 3D Scene is Worth 1K Tokens: 3D-Grounded Representation for Scene Generation at Scale","authors":["Dongxu Wei","QiXu","Zhiqi Li","Hangning Zhou","Cong Qiu","Hailong Qin","Mu Yang","Zhaopeng Cui","Peidong Liu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"A 3D Scene is Worth 1K Tokens: 3D-Grounded Representation for Scene Generation at Scale","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.11331","noteUrl":"/review?paper=review-neurips-2026-148286","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148288","title":"Learning Provable Neural Network Observer for Uncertain Dynamical Systems","authors":["Zhangyi Wang","Jiaxu Liu","Chen Song","Chao Xu","Shengze Cai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The observer formulas need a time convention and a qualified residual certificate","plainExplanation":"Nominal negative drift alone does not establish the stronger residual inequality. If the domain contains the equilibrium, strict negativity already fails there; a nonzero bounded disturbance also creates a linear term that can dominate quadratic decay near the origin.","primaryUrl":"https://arxiv.org/html/2609.30819","noteUrl":"/review?paper=review-neurips-2026-148288","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148290","title":"Pandora's AI Model Routing Box: Efficient Allocation with Costly Value Estimation","authors":["Adam Fisch","Shubhendu Trivedi","Fantine Huot","William Cohen","Michael Kaisers","Mirella Lapata","Kate Larson","Jacob Eisenstein"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Pandora's AI Model Routing Box: Efficient Allocation with Costly Value Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.20316","noteUrl":"/review?paper=review-neurips-2026-148290","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148291","title":"IRIS: Interpolative Rényi Iterative Self-play for Large Language Model Fine-Tuning","authors":["Wenjie Liao","Like Wu","Liangjie Zhao","Shihui Xu","Shigeru Fujimura"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The adaptive order can hit the loss singularity at initialization","plainExplanation":"The Rényi identity is missing its normalization factor. The adaptive order can hit the loss singularity at initialization.","primaryUrl":"https://arxiv.org/html/2604.20933","noteUrl":"/review?paper=review-neurips-2026-148291","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148295","title":"BA-T: An Iterative Transformer for Two-View Bundle Adjustment","authors":["Ganlin Zhang","Weirong Chen","Daniel Cremers","Xi Wang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"BA-T: An Iterative Transformer for Two-View Bundle Adjustment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03287","noteUrl":"/review?paper=review-neurips-2026-148295","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148298","title":"Binding Visual Features Point by Point","authors":["Udith Haputhanthri","Declan Campbell","Rim Assouel","Jonathan D Cohen","Taylor Webb"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Binding Visual Features Point by Point","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25427","noteUrl":"/review?paper=review-neurips-2026-148298","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148302","title":"ProEdit: Inversion-based Editing From Prompts Done Right","authors":["Zhi Ouyang","Dian Zheng","Xiao-Ming Wu","Jian-Jian Jiang","Kun-Yu Lin","Jingke Meng","Wei-Shi Zheng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The decreasing-time update negates the velocity twice","plainExplanation":"The decreasing-time update negates the velocity twice.","primaryUrl":"https://arxiv.org/html/2512.22118","noteUrl":"/review?paper=review-neurips-2026-148302","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148304","title":"Stabilizing Policy Optimization via Logits Convexity","authors":["Hongzhan Chen","Tao Yang","Yuhua Zhu","Shiping Gao","Xiaojun Quan","Ting Yao"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"An epsilon-optimal comparator need not give nonnegative gradient alignment","plainExplanation":"Proposition 4.4 uses convexity to assert \\(\\nabla L(z)^\\top(z-z^*)\\geq0\\) for an epsilon-optimal comparator. Convexity only yields a lower bound by \\(L(z)-L(z^*)\\), whose sign is not fixed for that comparator.","primaryUrl":"https://arxiv.org/html/2603.00963","noteUrl":"/review?paper=review-neurips-2026-148304","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148306","title":"CHASM: Cross-frequency Harmonized Axis-Separable Mixing for Spectral Token Operators","authors":["Pengcheng Fang","Hongli Chen","Yuxia Chen","Tengjiao Sun","Jiaxin Liu","Xiaohao Cai"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"CHASM: Cross-frequency Harmonized Axis-Separable Mixing for Spectral Token Operators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14727","noteUrl":"/review?paper=review-neurips-2026-148306","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148308","title":"Velocity-Space 3D Asset Editing","authors":["Hao Liu","yuxuan lin","jingfeng Guo","Ruihang Chu","junjie wang","Ruotong Li","Yujiu Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Velocity-Space 3D Asset Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07385","noteUrl":"/review?paper=review-neurips-2026-148308","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148310","title":"ABC-Align: Prediction-Powered Alignment with Adaptive Bias Control","authors":["Eric Frankel","Banghua Zhu","Sewoong Oh","Lillian Ratliff"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ABC-Align: Prediction-Powered Alignment with Adaptive Bias Control","plainExplanation":"No issue was found in the checked claims. The calculated mean disagreement rate of approximately $2.61\\%$ concerns a noisy label disagreeing with the noiseless sign. The source does not identify its ambiguity fraction with that event or assert a disagreement probability of $30\\%$. The two numbers therefore do not form a demonstrated contradiction.","primaryUrl":"https://arxiv.org/html/2609.34374","noteUrl":"/review?paper=review-neurips-2026-148310","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148311","title":"Testable and Actionable Calibration for Full Swap Regret","authors":["Konstantina Bairaktari","Lunjia Hu","Huy Nguyen","Jonathan Ullman"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Testable and Actionable Calibration for Full Swap Regret","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17749","noteUrl":"/review?paper=review-neurips-2026-148311","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148313","title":"Hypergraph Representation Learning with Hyperlink Random Effects","authors":["Zimeng Li","Shihao Wu","Gongjun Xu","Ji Zhu"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Hypergraph Representation Learning with Hyperlink Random Effects","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04640","noteUrl":"/review?paper=review-neurips-2026-148313","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148314","title":"Attention Sinks and Outliers in Attention Residuals","authors":["Haozheng Luo","Haoran Dai","Shaoyang Zhang","Xi Chen","Eric Hanchen Jiang","Yijiang Li","Ching-Yuen Huang","Chenghao Qiu","Chenwei Xu","Zhenyu Pan","Haotian Zhang","Binghui Wang","Yan Chen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Attention Sinks and Outliers in Attention Residuals","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17887","noteUrl":"/review?paper=review-neurips-2026-148314","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148315","title":"Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation","authors":["Divya Jyoti Bajpai","Arun Verma","Manjesh Kumar Hanawal"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03202","noteUrl":"/review?paper=review-neurips-2026-148315","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148323","title":"Procedural Memory Distillation: Online Reflection for Self-Improving Language Models","authors":["Ye Liu","Srijan Bansal","Bo Pang","Yang Li","Zeyu L Liu","Yifei Ming","Zixuan Ke","Shafiq Joty","Semih Yavuz"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Procedural Memory Distillation: Online Reflection for Self-Improving Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.01480","noteUrl":"/review?paper=review-neurips-2026-148323","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148324","title":"LACE: Lattice Attention for Cross-thread Exploration","authors":["Yang LI","Zirui Zhang","Yang Liu","Chengzhi Mao"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"LACE: Lattice Attention for Cross-thread Exploration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.15529","noteUrl":"/review?paper=review-neurips-2026-148324","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148326","title":"From Collapse to Improvement:  Statistical Perspectives on the Evolutionary Dynamics of Iterative Training on Contaminated Sources","authors":["Soham Bakshi","Sunrit Chakraborty"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The accumulated-risk additive term needs the total-size squared denominator","plainExplanation":"The accumulated-training risk recurrence uses total accumulated size \\(S\\), new sample count \\(n\\), and squared Euclidean parameter error. Conditional mean equals the previous estimate and conditional trace variance is \\(n(1-\\|\\widehat\\theta\\|^2)/S^2\\).","primaryUrl":"https://arxiv.org/html/2602.10531","noteUrl":"/review?paper=review-neurips-2026-148326","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148328","title":"Entropy Distribution as a Fingerprint for Hallucinations in Generative Models","authors":["Mattia Jacopo Villani","Pranav Deshpande","Akshay Seshadri","Romina Yalovetzky","Niraj Kumar"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Entropy Distribution as a Fingerprint for Hallucinations in Generative Models","plainExplanation":"The strengthened CES power bound applies lower-tail concentration above the alternative mean; a bounded independent example violates the bound. A separate KS calibration claim requires independence or a dependence-preserving null design. The original union bound and the theorems with explicit independence assumptions retain their stated scope.","primaryUrl":"https://arxiv.org/html/2605.28264","noteUrl":"/review?paper=review-neurips-2026-148328","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","lemma_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148331","title":"Co-GRPO: Co-Optimized Group Relative Policy Optimization for Masked Image Generation","authors":["Renping Zhou","Zanlin Ni","Tianyi Chen","Zeyu Liu","Yang Yue","Yulin Wang","Yuxuan Wang","Jingshu Liu","Gao Huang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The remasking softmax favors the tokens marked ineligible","plainExplanation":"The remasking softmax favors the tokens marked ineligible.","primaryUrl":"https://arxiv.org/html/2512.22288v1","noteUrl":"/review?paper=review-neurips-2026-148331","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148332","title":"Lumberjack: Better Differentially Private Random Forests through Heavy Hitter Detection in Trees","authors":["Christian J Lebeda","David Erb","Tudor Cebere","Aurélien Bellet"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Lumberjack: Better Differentially Private Random Forests through Heavy Hitter Detection in Trees","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22756","noteUrl":"/review?paper=review-neurips-2026-148332","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148333","title":"BarrierSteer: LLM Safety via Learning Barrier Steering","authors":["Thanh Q. Tran","Arun Verma","Kiwan Wong","Bryan Kian Hsiang Low","Daniela Rus","Wei Xiao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BarrierSteer: LLM Safety via Learning Barrier Steering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.20102","noteUrl":"/review?paper=review-neurips-2026-148333","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148339","title":"State of Thought Enables Endogenous Reasoning","authors":["Zhiren Gong","Yikun Hou","Zihao Zeng","Ming Xiao","Chau Yuen","Wei Yang Bryan Lim"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Statewise policy suprema erase the claimed strict open-loop disadvantage","plainExplanation":"The proposition defines each state value with its own policy supremum, then claims strictly better expected closed-loop value whenever optimal actions vary by state. At a final binary-state decision with reward one exactly when action equals state, the open class includes either constant action.","primaryUrl":"https://arxiv.org/html/2609.16055","noteUrl":"/review?paper=review-neurips-2026-148339","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148340","title":"The Implicit Bias of Hyperbolic Representation Learning for Multiclass Data: A Busemann Risk Perspective","authors":["Xingrun Li","Sho Kuno","Yusuke Mukuta","Xin Yang","Tatsuya Harada"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The Implicit Bias of Hyperbolic Representation Learning for Multiclass Data: A Busemann Risk Perspective","plainExplanation":"No issue was found in the checked claims. The proposed increasing-template witness conflicts with the monotonicity explicitly used to state and prove the intended classifier. A literal gap in the local transcription of Definition 4.2 does not establish a failure of Theorem 6.4 under the cited, explained PERM assumptions.","primaryUrl":"https://arxiv.org/html/2610.07131","noteUrl":"/review?paper=review-neurips-2026-148340","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148341","title":"DDMS: Discriminative Distillation of Multi-view Foundational Features into Single-view Models","authors":["Jeonggi Kwak","Sho Kagami","Yuki Ono","Kwang Moo Yi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DDMS: Discriminative Distillation of Multi-view Foundational Features into Single-view Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.23850","noteUrl":"/review?paper=review-neurips-2026-148341","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148342","title":"SPHERE-JEPA: Spherical Prediction with Homogeneous Embeddings","authors":["Léo Nicollier","Max Dunitz","Marc Pic","Pablo Muse","Enric Meinhardt-Llopis","Gabriele Facciolo"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"SPHERE-JEPA: Spherical Prediction with Homogeneous Embeddings","plainExplanation":"No issue was found in the checked claims. The old infinity argument is false on the circle, but it is absent from the current compact-manifold theorem. The replacement proof controls the expansion uniformly, localizes high-frequency targets, and handles dimension four with an ambient-coordinate identity.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2605.26900","noteUrl":"/review?paper=review-neurips-2026-148342","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148345","title":"Intrinsic Muon: Spectral Optimization on Riemannian Matrix Manifolds","authors":["Yibang Li","Bihari L Pandey","Ravi Sah","Andi Han","Cyrus Mostajeran","Pratik Kumar Jawanpuria","Bamdev Mishra"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The proposed Ky Fan update violates its norm constraint","plainExplanation":"The proposed matrix has $k$ singular values equal to $\\tau$, so its Ky Fan norm is $k\\tau$. It is infeasible for $k>1$ and need not maximize the objective even at $k=1$. The same paragraph reverses the endpoint norms and assigns this ball the radius of its dual ball. These are errors in the Ky Fan extension. The demonstrated defect is the Ky Fan oracle and radius in Appendix C.3. The Euclidean counterexample satisfies the paper's setting. It leaves the general descent argument with a valid oracle, the separately derived spectral, Frobenius and nuclear cases, and their experiments unchallenged.","primaryUrl":"https://arxiv.org/html/2605.09238","noteUrl":"/review?paper=review-neurips-2026-148345","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","incorrect_bound","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148346","title":"SpecDrop: Parameter-Free Category-Conditioned Routing for Modular Specialization","authors":["Boyao Wang","Zhihan Lei"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SpecDrop: Parameter-Free Category-Conditioned Routing for Modular Specialization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.04084","noteUrl":"/review?paper=review-neurips-2026-148346","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148349","title":"Mirror Descent-Ascent for mean-field min-max problems","authors":["Razvan-Andrei Lascu","Mateusz Majka","Lukasz Szpruch"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The mirror analysis needs boundary control and consistent constants","plainExplanation":"The proof correctly bounds the change in feature expectation by \\(2M_\\varphi\\operatorname{TV}(\\nu',\\nu)\\), then drops one factor of \\(M_\\varphi\\) when squaring it. No unit feature bound is assumed.","primaryUrl":"https://arxiv.org/html/2402.08106","noteUrl":"/review?paper=review-neurips-2026-148349","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148350","title":"PointForward: Feedforward Driving Reconstruction through Point-Aligned Representations","authors":["Cheng Chi_","Xianqi Wang","Hongcheng Luo","Mingfei Tu","Gangwei Xu","Zehan Zhang","Bing Wang","Guang Chen","Hangjun Ye","Sida Peng","Xin Yang","Haiyang Sun"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"PointForward: Feedforward Driving Reconstruction through Point-Aligned Representations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11594v1","noteUrl":"/review?paper=review-neurips-2026-148350","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148354","title":"HOPE: Hand-Object Pressure Estimation from Monocular Videos","authors":["Subin Jeon","Byungjun Kim","Hanbyul Joo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"HOPE: Hand-Object Pressure Estimation from Monocular Videos","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.06192","noteUrl":"/review?paper=review-neurips-2026-148354","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148355","title":"Steering Fields: Adaptive Vector Fields for Safe Image Generation and Beyond","authors":["Simone Facchiano","Jan Eric Lenssen","Bernt Schiele","Wolfgang Stammer","Fabio Galasso","Jonas Fischer"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Steering Fields: Adaptive Vector Fields for Safe Image Generation and Beyond","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39573","noteUrl":"/review?paper=review-neurips-2026-148355","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148357","title":"Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models","authors":["Yifu Yuan","Yaoting Huang","Xianze Yao","Shuoheng Zhang","Linqi Han","Yutong Li","Pengyi Li","Jiangeng Sun","Wenting Jia","Yucheng Hu","YuHao Liu","Ruihao Liao","Qiyu Wu","Yuxiao Li","zhao zhang","Zibin Dong","Fei Ni","YAN ZHENG","Shuyang Gu","Yi Ma","Hongyao Tang","Han Hu","Jianye Hao"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11324","noteUrl":"/review?paper=review-neurips-2026-148357","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148360","title":"Inference-Time Search Using Side Information for Diffusion-Based Image Reconstruction","authors":["Mahdi Farahbakhsh","Vishnu Teja Kunde","Dileep Kalathil","Krishna Narayanan","JF Chamberland"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Global resampling retains an extra parent tilt","plainExplanation":"The procedure starts from already tilted parent particles, proposes with the reference reverse kernel, and globally resamples using only \\(e^{V_t(x)}\\). The Doob kernel contains \\(e^{V_t(x)-V_{t+1}(y)}\\).","primaryUrl":"https://arxiv.org/html/2510.03352","noteUrl":"/review?paper=review-neurips-2026-148360","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148361","title":"Closing the Gap on the Sample Complexity of 1-Identification","authors":["ZITIAN LI","Wang Chi Cheung"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The uniform logarithmic bound fails at its first index","plainExplanation":"The summation lemma bounds positive summands by a universal multiple of \\(\\log j/j\\) for every admissible \\(j\\), including one. At \\(j=1\\), the left side is positive while \\(\\log1=0\\).","primaryUrl":"https://arxiv.org/html/2601.15620","noteUrl":"/review?paper=review-neurips-2026-148361","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148362","title":"Optimal Rates for Pure $\\varepsilon$-Differentially Private Stochastic Convex Optimization with Heavy Tails","authors":["Andrew Lowy"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Optimal Rates for Pure $\\varepsilon$-Differentially Private Stochastic Convex Optimization with Heavy Tails","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.06492","noteUrl":"/review?paper=review-neurips-2026-148362","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148363","title":"How deep is your network? Deep vs. shallow learning of transfer operators","authors":["Mohammad Tabish","Benedict Leimkuhler","Stefan Klus"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The discretization changes the SDE noise variance","plainExplanation":"The discretization changes the SDE noise variance.","primaryUrl":"https://arxiv.org/html/2509.19930","noteUrl":"/review?paper=review-neurips-2026-148363","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148364","title":"LLMs Show No Signs Of Individuated Metacognition","authors":["M. Moran","Mark Whiting"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LLMs Show No Signs Of Individuated Metacognition","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24299","noteUrl":"/review?paper=review-neurips-2026-148364","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148366","title":"AgentBrew: Offline Tool-Use Agent Learning from Raw Real-World Trajectories","authors":["Zhiyi Lyu","Yewen Li","Longtao Zheng","shengtian yang","Lang Feng","Lei Feng","Peng Jiang","Kun Gai","Qingpeng Cai","Bo An"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AgentBrew: Offline Tool-Use Agent Learning from Raw Real-World Trajectories","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.05837","noteUrl":"/review?paper=review-neurips-2026-148366","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148368","title":"Calibrating Generative Models to Feature Distributions with MMD Finetuning","authors":["Nathaniel L. Diamant","Brian L Trippe"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Calibrating Generative Models to Feature Distributions with MMD Finetuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19496","noteUrl":"/review?paper=review-neurips-2026-148368","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148370","title":"FlashMask-3: Efficient and Expressive Mask-Aware Distributed Attention","authors":["Guoxia Wang","Qianyue He","Siming Wu","Haoyang Xie","Qiwen Bao","Jinle Zeng","Jiabin Yang","Dianhai Yu","TIAN WU"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"FlashMask-3: Efficient and Expressive Mask-Aware Distributed Attention","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.17901","noteUrl":"/review?paper=review-neurips-2026-148370","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148371","title":"Where Should Society Draw the Line? A Social Choice Approach to Collective Consent","authors":["Chris Dong","Sonja Kraiczy","Rohit Prashanth Vasishta","Markus Brill","Wesley H Holliday","Niclas Boehmer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Where Should Society Draw the Line? A Social Choice Approach to Collective Consent","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2609.10759","noteUrl":"/review?paper=review-neurips-2026-148371","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148374","title":"Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts","authors":["Sagi Ahrac","Noya Hochwald","Mor Geva"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Softmax cross derivatives change the router coefficient","plainExplanation":"Softmax routing couples all selected probabilities. Omitting their cross derivatives changes the scalar gradient coefficient and can reverse its sign. The correction concerns the scalar coefficient and its interpretation; the multiplication by the input remains valid.","primaryUrl":"https://arxiv.org/html/2605.12476","noteUrl":"/review?paper=review-neurips-2026-148374","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148375","title":"CLoSeR: Closing the Loop for Long-Context Streaming Reconstruction","authors":["Moyang Li","Zihan Zhu","Wei Zhang","Marc Pollefeys","Daniel Barath"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"CLoSeR: Closing the Loop for Long-Context Streaming Reconstruction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.01927","noteUrl":"/review?paper=review-neurips-2026-148375","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148376","title":"On the Overscaling Curse of Parallel Thinking: System Efficacy Contradicts Sample Efficiency","authors":["Yiming Wang","Zhuosheng Zhang","Rui Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The overscaling proof assumes interval dominance absent from the theorem","plainExplanation":"Theorem 1 groups samples by their accuracy-curve shape. Let \\(F_i(N)\\) be mean accuracy at budget \\(N\\) in group \\(i\\), \\(p_i\\) its population fraction, and \\(\\mu_i\\) its mean sample-optimal budget. The theorem uses endpoint gains \\(\\bar\\Delta_i=F_i(N_{\\max})-F_i(1)\\), assumes \\(\\mu_4>\\mu_i\\) for \\(i\\ne4\\), \\(\\bar\\Delta_1=\\bar\\Delta_2=0\\), \\(\\bar\\Delta_5\\approx0\\) and \\(p_4\\bar\\Delta_4>p_3[-\\bar\\Delta_3]_+\\), then concludes expected budget utilization below one. The proof adds an interval-by-interval dominance condition that is absent from the theorem. It also drops the approximately zero Type-5 contribution without bounding its effect on the strict inequality.","primaryUrl":"https://arxiv.org/html/2601.21619","noteUrl":"/review?paper=review-neurips-2026-148376","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148378","title":"Seeing Speech: Learning Visible Articulatory Dynamics for Speech-Driven 3D Facial Animation","authors":["Hyung Kyu Kim","Byungchan Hwang","HAKGU KIM"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Seeing Speech: Learning Visible Articulatory Dynamics for Speech-Driven 3D Facial Animation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30517","noteUrl":"/review?paper=review-neurips-2026-148378","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148381","title":"Optimal sequential tests yield log-optimal e-processes","authors":["Ashwin Ram","Aaditya Ramdas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Optimal sequential tests yield log-optimal e-processes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12720","noteUrl":"/review?paper=review-neurips-2026-148381","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148382","title":"MaRK: Markov-adapted Recurrent Kernels for Dynamic Operator Conditioning in State Space Models","authors":["Syed Ibrahim Omer","Ginny Wong","Xiangyu Zhao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MaRK: Markov-adapted Recurrent Kernels for Dynamic Operator Conditioning in State Space Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.09092","noteUrl":"/review?paper=review-neurips-2026-148382","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148384","title":"Attributions All the Way Down? The Metagame of Interpretability","authors":["Hubert Baniecki","Przemyslaw Prem Biecek","Fabian Fumagalli"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Attributions All the Way Down? The Metagame of Interpretability","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06295v2","noteUrl":"/review?paper=review-neurips-2026-148384","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148385","title":"Iterative Critique-and-Routing Controller for Multi-Agent Systems with Heterogeneous LLMs","authors":["Wenzhi Fang","Liangqi Yuan","Guangchen Lan","Dong-Jun Han","Christopher Brinton"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Replacing expected calls by the horizon changes the constrained problem","plainExplanation":"Replacing expected calls by the horizon changes the constrained problem.","primaryUrl":"https://arxiv.org/html/2605.08686","noteUrl":"/review?paper=review-neurips-2026-148385","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148386","title":"Behavioral Foundation Models for Quality Diversity","authors":["Nazim Bendib","Nicolas Perrin-Gilbert","Olivier Sigaud"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Archive-weighted reward error needs coverage to control a candidate policy’s return","plainExplanation":"A small average projection error under the archive distribution does not control unseen states or their discounted contribution under a new policy. The later unweighted \\(\\|\\cdot\\|_2\\) notation does not supply the missing relation to the explicitly weighted projection residual.","primaryUrl":"https://arxiv.org/html/2609.35615","noteUrl":"/review?paper=review-neurips-2026-148386","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148387","title":"LinearARD: Linear-Memory Attention Distillation for RoPE Restoration","authors":["Ning Yang","Hengyu Zhong","Wentao Wang","Baoliang Tian","Yuan Zhou","Haijun Zhang","Jun Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"LinearARD: Linear-Memory Attention Distillation for RoPE Restoration","plainExplanation":"No issue was found in the checked claims. The row-loss definitions now match the reported derivatives. The earlier single-summand objection no longer applies.","primaryUrl":"https://arxiv.org/html/2604.00004","noteUrl":"/review?paper=review-neurips-2026-148387","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148389","title":"Scaling Reward Modeling without Human Supervision","authors":["Jingxuan Fan","Yueying Li","Zhenting Qi","Dinghuai Zhang","Kianté Brantley","Sham Kakade","Hanlin Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Scaling Reward Modeling without Human Supervision","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.02225","noteUrl":"/review?paper=review-neurips-2026-148389","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148390","title":"Certified Policy Optimisation for Nested Causal Bandits via PAC-Bayes Risk","authors":["Tim Woydt","Paul-David Zuercher"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The handover gate omits the confidence penalty","plainExplanation":"The handover gate omits the confidence penalty. Unit observed denominators do not establish policy overlap. On-policy collection does not make final-policy ratios equal to one.","primaryUrl":"https://arxiv.org/html/2605.29788","noteUrl":"/review?paper=review-neurips-2026-148390","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148391","title":"Text-Vision Co-Instructed Image Editing","authors":["Chenxi Xie","Yuhui Wu","Qiaosi Yi","Lei Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Text-Vision Co-Instructed Image Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.16767","noteUrl":"/review?paper=review-neurips-2026-148391","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148393","title":"Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context","authors":["Zhaowei Wang","Lishu Luo","Haodong Duan","WeiWeiLiu","Sijin Wu","Ji Luo","Shen Yan","Shuai Peng","Sihang Yuan","Chaoyi Huang","Yi Lin","Yangqiu Song"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13831","noteUrl":"/review?paper=review-neurips-2026-148393","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148395","title":"Monitoring Violations of Differential Privacy over Time","authors":["Önder Askin","Tim Kutta","Holger Dette"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Monitoring Violations of Differential Privacy over Time","plainExplanation":"At a valid privacy boundary, studentization creates a positive drift that accumulates over $T=n^2$, making the false-alarm probability tend to one. Separately, the unfloored score can have a zero denominator and a nonzero numerator. The paper’s optional positive floor defines a stabilized variant.","primaryUrl":"https://arxiv.org/html/2509.20283","noteUrl":"/review?paper=review-neurips-2026-148395","collection":"neurips-2026","severity":"critical","issueTypes":["algorithm_specification","false_statement","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-148396","title":"A Unifying Perspective on Language Model Interpretability","authors":["Enyan Zhang","R. Thomas McCoy"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The decoder-bias formula is dimensionally inconsistent and does not invert ReLU","plainExplanation":"The decoder-bias formula is dimensionally inconsistent and does not invert ReLU.","primaryUrl":"https://arxiv.org/html/2608.29034v1","noteUrl":"/review?paper=review-neurips-2026-148396","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148397","title":"Accuracy vs. Accuracy: Computational Tradeoffs Between Classification Rates and Utility","authors":["Noga Amit","Omer Reingold","Guy Rothblum"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Accuracy vs. Accuracy: Computational Tradeoffs Between Classification Rates and Utility","plainExplanation":"The current revision corrects the bin-width mismatch, the loss-weighted corollary’s audit-family premise, the conditional-frequency test, and the treatment of nontrivial threshold losses. Remaining mathematical findings concern aggregate calibration error, within-cell dependence, approximate affinity, finite-domain density, and several narrower proof or computational-cost claims.","primaryUrl":"https://arxiv.org/html/2505.16494","noteUrl":"/review?paper=review-neurips-2026-148397","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","invalid_inference","lemma_error","mathematical_error","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148401","title":"Sample Efficient Generative Molecular Optimization with Joint Self-Improvement","authors":["Serra Korkmaz","Adam Izdebski","Jonathan Pirnay","Rasmus Møller-Larsen","Michal Kmicikiewicz","Pankhil Gawade","Maksim Pavlov","Dominik G Grimm","Ewa Szczurek"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The Gibbs optimizer must tilt the objective’s reference distribution","plainExplanation":"The Gibbs optimizer must tilt the objective’s reference distribution. The squared-loss derivation changes only the reward sign.","primaryUrl":"https://arxiv.org/html/2602.10984","noteUrl":"/review?paper=review-neurips-2026-148401","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148402","title":"Are we really tilting? The mechanics of reward guidance in flow and diffusion models","authors":["Sanjit Dandapanthula","Nicholas Boffi"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Are we really tilting? The mechanics of reward guidance in flow and diffusion models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02884","noteUrl":"/review?paper=review-neurips-2026-148402","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148403","title":"Learning to Discover Iterative Spectral Algorithms","authors":["Zihang Liu","Oleg Balabanov","Yaoqing Yang","Michael Mahoney"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The fixed-degree polynomial diagnostic has a positive baseline","plainExplanation":"The diagnostic divides a polynomial’s affine-window supremum by its leading coefficient and interprets convergence to zero as an optimality gap. For nonzero affine scale and fixed degree \\(d\\), dividing by the transformed leading coefficient yields a monic polynomial on \\([-1,1]\\).","primaryUrl":"https://arxiv.org/html/2602.09530","noteUrl":"/review?paper=review-neurips-2026-148403","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148404","title":"Avoiding Obfuscation with Prover-Estimator Debate","authors":["Jonah Brown-Cohen","Geoffrey Irving","Georgios Piliouras","Lijie Chen","Jiawei Li","Zhiyang Xun"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Correcting the reward telescope","plainExplanation":"The printed polynomial does not equal its claimed telescope. Its extra terms are \\(\\varepsilon_0+\\varepsilon_1\\). This algebra alone is not a reachable dishonest-strategy counterexample.","primaryUrl":"https://arxiv.org/html/2506.13609","noteUrl":"/review?paper=review-neurips-2026-148404","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148405","title":"Back to Blackwell: Closing the Loop on Intransitivity in Multi-Objective Preference Fine-Tuning","authors":["Jiahao Zhang","Lujing Zhang","Keltin Grimes","Zhuohao Yu","Gokul Swamy","Steven Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Back to Blackwell: Closing the Loop on Intransitivity in Multi-Objective Preference Fine-Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.19041","noteUrl":"/review?paper=review-neurips-2026-148405","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148406","title":"Differentiable Bit-Widths: Co-optimizing Pruning and Quantization via SVD for Ultra-Efficient LLM Compression","authors":["Hankyul Kang","Jongbin Ryu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Differentiable Bit-Widths: Co-optimizing Pruning and Quantization via SVD for Ultra-Efficient LLM Compression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06026","noteUrl":"/review?paper=review-neurips-2026-148406","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148407","title":"MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging","authors":["JunXi Wang","Jiayi Zhu","Te Sun","Chen Zhang","Siyuan Li","Xuyang Liu","Zichen Wen","Xiaobing Tu","Jinkui Ren","Xiantao Zhang","Ziqi Yuan","Linfeng Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The total-relevance bound omits a deleted node at every merge","plainExplanation":"The cumulative bound subtracts a local change from one parent to the merged node but omits the other removed parent's score. Even identical embeddings violate the resulting total-score bound. Two additional gaps concern the sign needed by normalization and the unsupported use of the final minimum pair similarity to control earlier merges.","primaryUrl":"https://arxiv.org/html/2609.08273","noteUrl":"/review?paper=review-neurips-2026-148407","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148409","title":"FairMT: Fairness for Heterogeneous Multi-Task Learning","authors":["Guanyu Hu","Tangzheng Lian","Na Yan","Dimitrios Kollias","Xinyu Yang","Oya Celiktutan","Siyang Song","Zeyu Fu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Error parity does not imply prediction-distribution parity","plainExplanation":"Equalizing group error magnitudes can increase the difference between prediction distributions. Separately, differentiating the best-group reference can lower its rate, and the head-space QP has an incorrect fairness cross term. The binary EO selector needs a symbol correction, while the valid simplex projection requires a narrower statement about which tasks receive zero weight.","primaryUrl":"https://arxiv.org/html/2512.00469","noteUrl":"/review?paper=review-neurips-2026-148409","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","interpretation_error","invalid_inference","missing_assumption","notation","unsupported_claim"],"classifications":["major"]},{"slug":"review-neurips-2026-148410","title":"JODA: Composable Joint Dynamics for Articulated Objects","authors":["Tianhong Gao","Cheng Yu","Yinghao Xu","Mengyu Chu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"JODA: Composable Joint Dynamics for Articulated Objects","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09954","noteUrl":"/review?paper=review-neurips-2026-148410","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148412","title":"Seeking the Unfamiliar but Memorable: Conceptual Creativity as Meta-Learning","authors":["Mengye Ren"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Seeking the Unfamiliar but Memorable: Conceptual Creativity as Meta-Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16477","noteUrl":"/review?paper=review-neurips-2026-148412","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148413","title":"A Set-Sequence Model for Time Series","authors":["Elliot Epstein","Apaar Sadhwani","Kay Giesecke"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Invariant quadratics can require a quadratic readout","plainExplanation":"The set module first averages features computed separately for each unit, then applies a readout $\\rho$. Appendix A.2 fixes the number of units $M$ and takes $\\phi_2(x)=[x,\\operatorname{vec}_{\\mathrm{sym}}(xx^\\top)]$, which records first and second empirical moments. Its quadratic-case remark says a linear readout represents every permutation-invariant quadratic of the cross-section. The preceding expressivity proposition permits a continuous readout, and its proof uses a polynomial readout.\n\nThe source states: “Taking $k=2$ recovers first and second empirical moments exactly, so any permutation-invariant quadratic statistic of the cross-section is obtained by a linear $\\rho$.” A linear readout can combine the pooled first and second moments, but it cannot multiply pooled first moments. Those products are needed for quadratic interactions between different units. Permutation invariance does not remove these interactions.","primaryUrl":"https://arxiv.org/html/2505.11243","noteUrl":"/review?paper=review-neurips-2026-148413","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148414","title":"Asymmetric Hierarchical Anchoring for Robust Audio–Visual Cross-Modal Generalization","authors":["Bixing Wu","Yuhong Zhao","Zongli Ye","Jiachen Lian","Xiangyu Yue","Gopala Anumanchipalli"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Velocity sampling has no distribution for a constant sequence","plainExplanation":"The anchor sampler derives probabilities from nonnegative semantic changes \\(\\delta_t\\), adding epsilon only to the normalization denominator. For a constant semantic sequence, every \\(\\delta_t=0\\).","primaryUrl":"https://arxiv.org/html/2602.03570","noteUrl":"/review?paper=review-neurips-2026-148414","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148417","title":"Exploring Starts Are Not Enough: Counterexamples and a Fix for Monte Carlo Exploring Starts","authors":["Octave Oliviers","Glenn Vinnicombe"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The main-text construction and table use a reflected orientation","plainExplanation":"The main-text construction and table use a reflected orientation.","primaryUrl":"https://arxiv.org/html/2606.15247","noteUrl":"/review?paper=review-neurips-2026-148417","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148420","title":"M*: A Modular, Extensible, Serving System for Multimodal Models","authors":["Atindra Jha","Naomi Sagan","Keisuke Kamahori","Irmak Sivgin","Rohan Sanda","Steven Gao","Mark Horowitz","Luke Zettlemoyer","Olivia Hsu","Jure Leskovec","Stephanie Wang","Baris Kasikci"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"M*: A Modular, Extensible, Serving System for Multimodal Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.12688","noteUrl":"/review?paper=review-neurips-2026-148420","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148423","title":"Positional LSH: Binary Block Matrix Approximation for Attention with Linear Biases","authors":["Daniel Wolfson","Tal Wagner"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Positional LSH: Binary Block Matrix Approximation for Attention with Linear Biases","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09472","noteUrl":"/review?paper=review-neurips-2026-148423","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148426","title":"UniSHARP: Universal Sharp Monocular View Synthesis","authors":["Meixi Song","Dizhe Zhang","Ruiyang Zhang","Bo Du","Ming-Hsuan Yang","Lu Qi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"UniSHARP: Universal Sharp Monocular View Synthesis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.07514","noteUrl":"/review?paper=review-neurips-2026-148426","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148427","title":"FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping","authors":["Jongmin Choi","Joon Son Chung"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13071","noteUrl":"/review?paper=review-neurips-2026-148427","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148428","title":"Market Regime Council for Dynamic Credit Assignment in Multi-Agent LLM  Decision Systems","authors":["Yunhua Pei","Zerui Ge","Jin Zheng","John Cartlidge"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Regime switches break continuity, and fixed memory does not imply oracle convergence","plainExplanation":"A hard regime switch produces a nonzero jump in the final feasible allocation, contradicting the global Lipschitz claim. Separately, exponential weighting correctly removes old-data influence but retains recent-data variation; raw Shapley credits need not converge to their distribution-oracle values.","primaryUrl":"https://arxiv.org/html/2605.24490","noteUrl":"/review?paper=review-neurips-2026-148428","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148430","title":"Diffusion LLMs are Natural Adversaries for any LLM","authors":["David Lüdke","Tom Wollschläger","Paul Ungermann","Stephan Günnemann","Leo Schwinn"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Diffusion LLMs are Natural Adversaries for any LLM","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.00203","noteUrl":"/review?paper=review-neurips-2026-148430","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148432","title":"Free Lunch for Pass@$k$? Low Cost Diverse Sampling for Diffusion Language Models","authors":["Sean Lamont","Christian Walder","Paul Montague","Amir Dezfouli","Michael Norrish"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Gram–Schmidt normalization needs a zero-residual branch","plainExplanation":"Gram–Schmidt normalization needs a zero-residual branch. The quality scalar is constant on unmasked tokens.","primaryUrl":"https://arxiv.org/html/2603.04893","noteUrl":"/review?paper=review-neurips-2026-148432","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148433","title":"Learning How to Cube","authors":["Ferhat Erata","Sam Kouteili","Thanos Typaldos","Timos Antonopoulos","Robert B Jones","Byron Cook","Ruzica Piskac"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Minimum-score preference reverses the stated reward ordering","plainExplanation":"Minimum-score preference reverses the stated reward ordering.","primaryUrl":"https://arxiv.org/html/2605.16632","noteUrl":"/review?paper=review-neurips-2026-148433","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148437","title":"Evaluation Awareness in Language Models Has Limited Effect on Behaviour","authors":["Amelie Knecht","Lucas Florin","Thilo Hagendorff"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Evaluation Awareness in Language Models Has Limited Effect on Behaviour","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05835","noteUrl":"/review?paper=review-neurips-2026-148437","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148438","title":"Why Deterministic PRM Guidance Underperforms in Discrete Diffusion Reasoning","authors":["Yan Zhan","Shaobo Liu","Zhijun Gao"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Why Deterministic PRM Guidance Underperforms in Discrete Diffusion Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35472","noteUrl":"/review?paper=review-neurips-2026-148438","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148439","title":"STAPO: Stabilizing Reinforcement Learning for LLMs by Silencing Rare Spurious Tokens","authors":["Shiqi Liu","Zeyu He","Guojian Zhan","Letian Tao","Zhilong Zheng","Jiang Wu","yinuo Wang","Yang Guan","Kehua Sheng","Bo Zhang","Keqiang Li","Jingliang Duan","Shengbo Eben Li"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"STAPO: Stabilizing Reinforcement Learning for LLMs by Silencing Rare Spurious Tokens","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.15620","noteUrl":"/review?paper=review-neurips-2026-148439","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148444","title":"Layerwise LQR for Geometry-Aware Optimization of Deep Networks","authors":["Simon Dufort-Labbé","Pierre-Luc Bacon","Razvan Pascanu","Simon Lacoste-Julien","Aristide Baratin"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The Lagrangian omits the first dynamics constraint","plainExplanation":"The Lagrangian starts its dynamics-constraint sum one index too late, omitting the first transition. Its stationarity equations therefore do not yield the first adjoint/control relations claimed from it. This is an indexing omission in the displayed adjoint derivation. The intended standard dynamics recursion and the reported solver are not refuted by it.","primaryUrl":"https://arxiv.org/html/2605.04230","noteUrl":"/review?paper=review-neurips-2026-148444","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148445","title":"Growing a Neural Network in Breadth, Depth, and Time","authors":["Eivinas Butkus","Kedar Garzón Gupta","Nikolaus Kriegeskorte"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Unit-norm temporal logits do not imply the advertised softmax interval","plainExplanation":"Unit-norm temporal logits do not imply the advertised softmax interval.","primaryUrl":"https://arxiv.org/html/2605.25174","noteUrl":"/review?paper=review-neurips-2026-148445","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148446","title":"Flow-Corrected Shape Optimization: Taming Manifold Drift in High-Dimensional 3D Models","authors":["Emilien Seiler","Nicolas Talabot","Yingxuan You","Federico Stella","Pascal Fua"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Transport to the data law does not make pointwise density approach one","plainExplanation":"Transport to the data law does not make pointwise density approach one.","primaryUrl":"https://arxiv.org/html/2608.07199","noteUrl":"/review?paper=review-neurips-2026-148446","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148447","title":"X2HDR: HDR Image Generation in a Perceptually Uniform Space","authors":["Ronghuan Wu","Wanchao Su","Kede Ma","Jing Liao","Rafal Mantiuk"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"X2HDR: HDR Image Generation in a Perceptually Uniform Space","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04814","noteUrl":"/review?paper=review-neurips-2026-148447","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148449","title":"Test-time Risk Adaptation with Mixture of Agents","authors":["Mohamad Chehade","Amrit Singh Bedi","Souradip Chakraborty","Amy Zhang","Hao Zhu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Test-time Risk Adaptation with Mixture of Agents","plainExplanation":"The localized transfer guarantee can predict zero loss for a strictly suboptimal target policy, even with no risk penalty. Its reward-only repair must include mismatch along the source policy as well as the target-optimal policy. The separate per-step variance formula also uses inconsistent moment normalizations.","primaryUrl":"https://arxiv.org/html/2408.08812","noteUrl":"/review?paper=review-neurips-2026-148449","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","formula_error","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-148452","title":"Instance-Adaptive Online Multicalibration","authors":["Zhiming Huang","Jamie Morgenstern","Aaron Roth","Claire Jie Zhang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Instance-Adaptive Online Multicalibration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09273","noteUrl":"/review?paper=review-neurips-2026-148452","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148453","title":"Social Choice Foundations for Simulation-Augmented Generation","authors":["Sonja Kraiczy","Smitha Milli","Ratip Emin Berker","Avinandan Bose","Brandon Amos","Jamelle Watson-Daniels","Maximilian Nickel","Edith Elkind","Ariel Procaccia"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Social Choice Foundations for Simulation-Augmented Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38287","noteUrl":"/review?paper=review-neurips-2026-148453","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148458","title":"Closed-Form Implicit Neural Representations","authors":["Woojin Cho","Junghwan Park"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Estimator guarantees and partition feasibility","plainExplanation":"Existence of suitable hidden weights does not guarantee that a prescribed random feature span contains such an approximant, or that a finite-sample least-squares fit finds its population approximation.","primaryUrl":"https://arxiv.org/html/2602.07603v1","noteUrl":"/review?paper=review-neurips-2026-148458","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","interpretation_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148459","title":"Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommender Systems","authors":["Jun YIN","Bangguo Zhu","Peng Huo","Ruochen Liu","Hao Chen","Senzhang Wang","Shirui Pan","Chengqi Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"An odds amplification factor does not give the stated absolute suppression factor","plainExplanation":"The definition gives \\(P_\\theta(t)/P_d(t)=\\gamma_j^{-1}P_\\theta(h)/P_d(h)\\). Therefore \\(P_\\theta(t)\\leq\\gamma_j^{-1}P_d(t)\\) additionally requires \\(P_\\theta(h)\\leq P_d(h)\\), which is not assumed. In a binary competition, raising the head odds raises head mass and violates that extra requirement. Multiplication across the path also needs control of nonbranching conditional probabilities, not supplied by the branching odds condition. The absolute \\(\\gamma_{\\min}^{-z}\\) bound is unsupported. Qualitative suppression and the odds-ratio conclusion can survive with correctly stated additional conditions.","primaryUrl":"https://arxiv.org/html/2605.16825v3","noteUrl":"/review?paper=review-neurips-2026-148459","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148461","title":"Theoretical guidelines for annealed Langevin dynamics in compositional simulation-based inference","authors":["Camille Touron","Gabriel Cardoso","Julyan Arbel","Pedro Rodrigues"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Theoretical guidelines for annealed Langevin dynamics in compositional simulation-based inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21253","noteUrl":"/review?paper=review-neurips-2026-148461","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148463","title":"Exact Topological Compliance: Generating Persistence-Equivalent Graphs","authors":["Mattie Ji","Indradyumna Roy","Vikas Garg"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Exact Topological Compliance: Generating Persistence-Equivalent Graphs","plainExplanation":"The printed local graph generator excludes a realizable two-vertex graph because its pivot requires an old endpoint and its merge loop reuses the pivot’s death. Allowing two new endpoints and removing the consumed death repairs the local reachability argument; the original author implementation already avoids this example.","primaryUrl":"https://arxiv.org/html/2610.18856","noteUrl":"/review?paper=review-neurips-2026-148463","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148464","title":"Rethinking Neural Nonlinearity as Gating","authors":["MUHAMMAD SABIH","Frank Hannig","Jürgen Teich"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The local third derivative does not determine the propagation depth scale","plainExplanation":"The depth corollary drops initialization and Gaussian-moment factors that its cited propagation theory retains.","primaryUrl":"https://arxiv.org/html/2607.03148","noteUrl":"/review?paper=review-neurips-2026-148464","collection":"neurips-2026","severity":"major","issueTypes":["citation_hypothesis_mismatch","false_statement","invalid_inference","lemma_error","missing_assumption","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-148465","title":"Text as Partial Constraint: Core–Residual Alignment for Robust Vision–Language Learning","authors":["Chengzhen Yu","Canran Xiao","SiYuan Ma","Yang Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The multi-view concentration rate needs conditional independence","plainExplanation":"The multi-view concentration rate needs conditional independence. The spherical-cap formula needs the alignment hypothesis already stated in the appendix.","primaryUrl":"https://arxiv.org/html/2607.03143","noteUrl":"/review?paper=review-neurips-2026-148465","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148466","title":"Towards Anytime-Valid Statistical Watermarking","authors":["Baihe Huang","Eric Xu","Kannan Ramchandran","Jiantao Jiao","Michael Jordan"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Towards Anytime-Valid Statistical Watermarking","plainExplanation":"Mixing optimal couplings at opposite extreme targets can leave unnecessary off-diagonal mass at an interior target. A maximal coupling for the actual target repairs the inner optimization and recovers the same worst-case growth for the displayed detector, consistent with Remark 4.2.","primaryUrl":"https://arxiv.org/html/2602.17608","noteUrl":"/review?paper=review-neurips-2026-148466","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148467","title":"Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction","authors":["kejun ren","Lei Jin","Tianxin Huang","Lianming Xu","Li Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16981","noteUrl":"/review?paper=review-neurips-2026-148467","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148468","title":"End-to-End Identifiable and Consistent Recurrent Switching Dynamical Systems","authors":["Carles Balsells Rodas","Zhengrui Xiang","Francisco Sumba Toral","Yingzhen Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Pairwise mean separation does not ensure simultaneous component dominance","plainExplanation":"Pairwise mean separation does not ensure simultaneous component dominance. The autonomous reconstruction mixes row and column conventions.","primaryUrl":"https://arxiv.org/html/2605.06315","noteUrl":"/review?paper=review-neurips-2026-148468","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148469","title":"Cheap Per-Component Testing for PLS, Stable Under Rotation","authors":["Paweł Lenartowicz","Hubert Plisiecki"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The operational p-value and sidedness statement use different tests","plainExplanation":"The operational p-value and sidedness statement use different tests.","primaryUrl":"https://arxiv.org/html/2609.36307v1","noteUrl":"/review?paper=review-neurips-2026-148469","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148470","title":"Learning to Trigger:  Reinforcement Learning at the Large Hadron Collider","authors":["Zixin Ding","Shaghayegh Emami","Giovanna Salvi","Cecilia Tosciri","Abhijith Gandrakota","Jennifer Ngadiuba","Nhan Tran","Christian Herwig","David W Miller","Yuxin Chen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Learning to Trigger:  Reinforcement Learning at the Large Hadron Collider","plainExplanation":"Bounded threshold updates can make a non-greedy action optimal even when every cut satisfies the rate tolerance. Immediate rewards and exogenous drift do not establish the claimed exact equivalence of discount choices.","primaryUrl":"https://arxiv.org/html/2606.23993","noteUrl":"/review?paper=review-neurips-2026-148470","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-148475","title":"Scaling Laws for Multimodal Data Mixtures","authors":["Aditi Khandelwal","Ayush Kumar Tarun","Yixuan Xu","Imanol Schlag","Golnoosh Farnadi","Siva Reddy","Luke Zettlemoyer","Antoine Bosselut"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Scaling Laws for Multimodal Data Mixtures","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.14423","noteUrl":"/review?paper=review-neurips-2026-148475","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148476","title":"QuantDemoire: Quantization with Outlier Aware for Image Demoiréing","authors":["郑 陈","Kewei Zhang","Xiaoyang Liu","Kai Liu","Weihang Zhang","Mengfan Wang","Yifan Fu","Yulun Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The dilated filter can preserve Nyquist energy at every level","plainExplanation":"The dilated filter can preserve Nyquist energy at every level.","primaryUrl":"https://arxiv.org/html/2510.04066","noteUrl":"/review?paper=review-neurips-2026-148476","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148477","title":"Statistical Estimation of Adversarial Risk in Large Language Models under Best-of-N Sampling","authors":["Mingqian Feng","Xiaodong Liu","Weiwei Yang","Chenliang Xu","Christopher White","Jianfeng Gao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The fixed-anchor budget interpretation reverses the exponent effect","plainExplanation":"At a fixed measured attack-success rate and a higher target, a smaller fitted exponent $\\widehat\\alpha$ predicts a larger required sampling budget. The explanation after Equation (16) reverses this direction; the inverse formula itself needs no change.","primaryUrl":"https://arxiv.org/html/2601.22636","noteUrl":"/review?paper=review-neurips-2026-148477","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["false_statement"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148484","title":"Q-MMR: Off-Policy Evaluation via Recursive Reweighting and Moment Matching","authors":["Nan Jiang","Xiang Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"One norm formula and one recursion step need correction","plainExplanation":"Positivity within each row does not imply monotonicity across rows with different coefficients. The proof therefore does not justify replacing \\(\\max_{j\\le h}F_j\\) by \\(F_h\\).","primaryUrl":"https://arxiv.org/abs/2605.06474v3","noteUrl":"/review?paper=review-neurips-2026-148484","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148485","title":"Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning","authors":["Byoungwoo Park","Utkarsh Mishra","Jaemoo Choi","Juho Lee","Yongxin Chen"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The stated refinement depth gives a different evaluation count","plainExplanation":"The stated refinement depth gives a different evaluation count. The patch count and overlap imply width 2560. The cosine schedule cannot reach the promised terminal weight.","primaryUrl":"https://arxiv.org/html/2606.00837","noteUrl":"/review?paper=review-neurips-2026-148485","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148487","title":"Temporal Alignment Guidance: On-manifold Sampling in Diffusion Models","authors":["Youngrok Park","Hojung Jung","Sangmin Bae","Se-Young Yun"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The stated TAG potential omits a varying posterior-entropy term","plainExplanation":"The proposed potential omits the gradient of the time-label posterior entropy. An exact replacement repairs the local identity, but does not by itself establish the later convergence claims.","primaryUrl":"https://arxiv.org/html/2510.11057","noteUrl":"/review?paper=review-neurips-2026-148487","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","lemma_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148488","title":"FreqCa: Accelerating Image generation and editing via Frequency-Aware Caching","authors":["Jiacheng Liu","Peiliang Cai","Qinming Zhou","Yuqi Lin","Deyang Kong","Benhao Huang","Yupei Pan","Haowen Xu","Chang Zou","Junshu Tang","Shikang Zheng","Linfeng Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"FreqCa: Accelerating Image generation and editing via Frequency-Aware Caching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.08669v1","noteUrl":"/review?paper=review-neurips-2026-148488","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148490","title":"GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling","authors":["Alireza Dadgarnia","Rush Tabesh","Mahdi Nikdan","Michael Helcig","Eldar Kurtić","Maximilian Kleinegger","Dan Alistarh"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.18556","noteUrl":"/review?paper=review-neurips-2026-148490","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148491","title":"Solaris: Building a Multiplayer Video World Model in Minecraft","authors":["Oscar Michel","Georgy Savva","Daohan Lu","Suppakit Waiwitlikhit","Timothy Meehan","Dhairya P Mishra","Egor Gikalo","Srivats Poddar","Jack Lu","Saining Xie"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Solaris: Building a Multiplayer Video World Model in Minecraft","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.22208","noteUrl":"/review?paper=review-neurips-2026-148491","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148495","title":"TriSearch: Learning to Optimize Triangulations via Bistellar Flips","authors":["Yiran Wang","Guido Montufar"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Coordinate-dependent scalar features remove the encoder’s invariance guarantee","plainExplanation":"Coordinate-dependent scalar features remove the encoder’s invariance guarantee.","primaryUrl":"https://arxiv.org/html/2605.30220","noteUrl":"/review?paper=review-neurips-2026-148495","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148496","title":"Reasoning as Compression: Unifying Budget Forcing via the Conditional Information Bottleneck","authors":["Fabio Valerio Massoli","Andrey Kuzmin","Arash Behboodi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Normalized conditionals do not enforce the data-marginal constraint","plainExplanation":"The exact CIB reparameterization requires \\(P(y\\mid x)=\\sum_z\\pi_\\theta(y\\mid x,z)\\pi_\\theta(z\\mid x)\\), but the practical derivation says normalization makes the constraint unnecessary. Normalization only ensures that the right side is some distribution over \\(y\\), not that it equals the fixed data conditional \\(P(y\\mid x)\\).","primaryUrl":"https://arxiv.org/html/2603.08462","noteUrl":"/review?paper=review-neurips-2026-148496","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148501","title":"Bian Que: An Agentic Framework with Flexible Skill Arrangement for Online System Operations","authors":["bochao liu","Zhipeng Qian","yang zhao","Xinyuan Jiang","Zihan Liang","Yufei Ma","Junpeng Zhuang","Ben Chen","Shuo Yang","Hongen Wan","Yao Wu","Chenyi Lei"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Bian Que: An Agentic Framework with Flexible Skill Arrangement for Online System Operations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.26805","noteUrl":"/review?paper=review-neurips-2026-148501","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148502","title":"Agnostic Language Identification and Generation","authors":["Mikael Møller Høgsgaard","Chirag Pabbaraju"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The countable-language argmax need not exist","plainExplanation":"Theorem 3.2 permits countably infinite language collections, but Algorithm 2 selects a maximizer of \\(r_i\\) over all indices. Let the universe be positive integers, \\(D(k)=2^{-k}\\), and \\(L_i=\\{k:k\\geq i\\}\\).","primaryUrl":"https://arxiv.org/html/2601.23258","noteUrl":"/review?paper=review-neurips-2026-148502","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148503","title":"What’s Holding Back Latent Visual Reasoning?","authors":["Andre G Viveiros","Nuno Gonçalves","André Martins","Matthias Lindemann"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"What’s Holding Back Latent Visual Reasoning?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18445","noteUrl":"/review?paper=review-neurips-2026-148503","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148506","title":"Difference of Convex Programming in the Wasserstein Space with Applications to MMD Optimization","authors":["Clément Bonet","Pierre-Cyril Aubin-Frankowski","Youssef Mroueh"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Radial Hessian bounds need the squared diameter","plainExplanation":"The required range is \\([0,S_*^2]\\). The compact-domain constants are affected. Whole-space results have infinite diameter and are unaffected by this range error; diameter at most one already gives conservative coverage.","primaryUrl":"https://arxiv.org/html/2606.27767","noteUrl":"/review?paper=review-neurips-2026-148506","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148507","title":"Ergodic Trajectory Design by Learned Pushforward Maps: Provable Coverage via Conditional Flow Matching","authors":["Ehsan Aghazadeh","Masoud Malekzadeh","Ahmad Ghasemi","Hossein Pishro-Nik"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The finite-entropy error bridge mixes interpolation measures","plainExplanation":"A triangle inequality under the entropic interpolation cannot control a norm under the different exact-transport interpolation measure. The finite-entropy CFM bridge is affected. Theorem 3, stated directly in terms of exact population error, is not contradicted.","primaryUrl":"https://arxiv.org/html/2605.13063","noteUrl":"/review?paper=review-neurips-2026-148507","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148509","title":"A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models","authors":["Hamid Kazemi","Atoosa Malemir Chegini","Maria Safi"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08513","noteUrl":"/review?paper=review-neurips-2026-148509","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148510","title":"Trust, but Don’t Verify: Epistemic Blind Spots in LLM Source Evaluation","authors":["Rohan N Pradhan","Steve Goley"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Trust, but Don’t Verify: Epistemic Blind Spots in LLM Source Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05403","noteUrl":"/review?paper=review-neurips-2026-148510","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148512","title":"Persistent-Transient Policy Evaluation for Markov Chains via Minimal Peripheral Quotients","authors":["Yang Xu","Vaneet Aggarwal"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The inverse perturbation expansion has reversed signs","plainExplanation":"The factorization and inverse series use signs incompatible with \\(\\Delta=\\widehat B-B\\). The matrix factor must be \\(I+B^{-1}\\Delta\\), with an alternating Neumann series. The norm bounds use absolute values and the same small-perturbation condition. The local sign corrections do not change that condition.","primaryUrl":"https://arxiv.org/html/2602.00474","noteUrl":"/review?paper=review-neurips-2026-148512","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148513","title":"Missing Old Logits in Asynchronous Agentic RL: Semantic Mismatch and Repair Methods for Off-Policy Correction","authors":["Zhong Guan","Yongjian Guo","Haoran Sun","Wen Huang","shuai di","Junwu Xiong","Likang Wu","Hongke Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The PPO ratio upper bound needs a sign-dependent branch","plainExplanation":"The appendix divides \\([1-(1-\\alpha)(1+\\epsilon)]r\\leq\\alpha(1+\\epsilon)\\) to give one upper bound for every \\(0<\\alpha<1\\). The ratio denominator is positive, but the coefficient of \\(r\\) is positive only when \\(\\alpha>\\epsilon/(1+\\epsilon)\\).","primaryUrl":"https://arxiv.org/html/2605.12070","noteUrl":"/review?paper=review-neurips-2026-148513","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148519","title":"Tokenizer-Generator Coupling in Medical Image Generation","authors":["Liam Chalcroft"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Tokenizer-Generator Coupling in Medical Image Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.07713","noteUrl":"/review?paper=review-neurips-2026-148519","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148522","title":"Provable Quantization with Randomized Hadamard Transform","authors":["Ying Feng","Piotr Indyk","Michael Kapralov","Dmitrii Krachun","Boris Prokhorov"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Provable Quantization with Randomized Hadamard Transform","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13810","noteUrl":"/review?paper=review-neurips-2026-148522","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148523","title":"Achieving Better Local Regret Bound for Online Non-Convex Bilevel Optimization","authors":["Tingkai Jia","Haiguang Wang","Ting Wang","Cheng Chen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Regret guarantees need a valid stopping certificate and error bounds","plainExplanation":"The constrained solver can fail to stop at a constrained stationary point. Two additional proof steps use an unjustified contraction rate and an incorrect coefficient in a tracking inequality.","primaryUrl":"https://arxiv.org/html/2602.06457","noteUrl":"/review?paper=review-neurips-2026-148523","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148525","title":"Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation","authors":["Kexuan Shi","Hanxuan Li","Zeju Qiu","Yandong Wen","Simon Buchholz","Weiyang Liu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation","plainExplanation":"The deterministic remainder constant uses a uniform generator bound that is not stated for the stochastic samples. Its reuse in Equation (55) is therefore unproved. The bilateral polynomial update contains fourth-degree sample products and its squared remainder contains products through degree eight.","primaryUrl":"https://arxiv.org/html/2605.12492","noteUrl":"/review?paper=review-neurips-2026-148525","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148526","title":"Diffusion Masked Pretraining for Dynamic Point Cloud","authors":["Zhuoyue Zhang","Yiding Sun","Chaowei Fang","Haozhe Cheng","Jian Liu","Jihua Zhu","Ajmal Mian"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Diffusion Masked Pretraining for Dynamic Point Cloud","plainExplanation":"Mean-regression supervision constrains the prediction head without forcing the encoder to discard class information. A Bayes-optimal mean predictor can coexist with perfectly discriminative features; the narrower obstruction applies to a classifier receiving only the identical mean statistic, as distinguished in the v1 appendix.","primaryUrl":"https://arxiv.org/html/2605.03639","noteUrl":"/review?paper=review-neurips-2026-148526","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148527","title":"GraDE: A Graph Diffusion Estimator for Frequent Subgraph Discovery in Neural Architectures","authors":["Yikang Yang","Zhengxin Yang","Minghao Luo","Luzhou Peng","Hongxiao Li","Wanling Gao","Lei Wang","Jianfeng Zhan"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Edge-weighted graph selection does not sample occurrences uniformly","plainExplanation":"Edge-weighted graph selection does not sample occurrences uniformly. The reverse-rate formula needs a separate diagonal definition. The extra one-way rate does not satisfy detailed balance.","primaryUrl":"https://arxiv.org/html/2602.03257","noteUrl":"/review?paper=review-neurips-2026-148527","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148529","title":"Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models","authors":["Runqian (Ray) Wang","Yilun Du"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models","plainExplanation":"Conditional targets can cancel away from the data. The resulting stationary-point example exposes a proof gap, while the stronger claim about all local minima remains unresolved.","primaryUrl":"https://arxiv.org/html/2510.02300","noteUrl":"/review?paper=review-neurips-2026-148529","collection":"neurips-2026","severity":"major","issueTypes":["proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148530","title":"Asymptotically Exact Negative Guidance of Diffusion Models via Positive-Unlabeled Learning","authors":["Hirohane Takagi","Masashi Sugiyama"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Asymptotically Exact Negative Guidance of Diffusion Models via Positive-Unlabeled Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.19802","noteUrl":"/review?paper=review-neurips-2026-148530","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148531","title":"LLM Agents Already Know When to Call Tools - Even Without Reasoning","authors":["Chung-En Sun","Linbo Liu","Ge Yan","Zimo Wang","Lily Weng"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The threshold tradeoff runs in the opposite direction","plainExplanation":"The prose reverses the threshold’s effect on predicted tool use. Because direct answering is selected by \\(p<\\tau\\), increasing \\(\\tau\\) selects more direct answers, not more predicted tool calls. The binary rule itself is coherent. This correction concerns its monotonicity explanation and does not guarantee monotonic realized calls or accuracy.","primaryUrl":"https://arxiv.org/html/2605.09252","noteUrl":"/review?paper=review-neurips-2026-148531","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148533","title":"Functional Gradient Descent with Adaptive Representations","authors":["Daniel Csillag","Rodrigo Schuller","Pedro Dall’Antonia","Leonidas Guibas","Luiz Velho","Tiago Novello"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Absolute gradient accuracy need not yield relative accuracy at stationarity","plainExplanation":"Absolute errors can converge to zero at a stationary point while their relative errors remain one, preventing the stated inner-loop termination. The unconditional inner-loop termination claim fails at stationary points. Descent statements conditional on successful steps are not disproved.","primaryUrl":"https://arxiv.org/html/2606.16926","noteUrl":"/review?paper=review-neurips-2026-148533","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148534","title":"Unveiling Memorization–Generalization Coexistence: A Case Study on Arithmetic Tasks with Label Noise","authors":["Linyu Liu","Pinyan Lu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Unveiling Memorization–Generalization Coexistence: A Case Study on Arithmetic Tasks with Label Noise","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18022","noteUrl":"/review?paper=review-neurips-2026-148534","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148539","title":"FunctionEvolve: Structure-Guided Symbolic Regression with LLMs","authors":["Zeyu Xia","Jun Zhu","Dong Yan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FunctionEvolve: Structure-Guided Symbolic Regression with LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.07704","noteUrl":"/review?paper=review-neurips-2026-148539","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148540","title":"Fixed-Point Reasoning: Stable and Adaptive Deep Looped Models","authors":["Sajad Movahedi","Shlomo Libo Feigin","Vera Milovanović","Alexander Theus","Thomas Hofmann","Valentina Boeva","T. Konstantin Rusch","Antonio Orvieto"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The intermediate norm estimate omits half the sublayer terms","plainExplanation":"The intermediate norm estimate omits half the sublayer terms. Normalizing a vector does not bound its image under an arbitrary matrix by one.","primaryUrl":"https://arxiv.org/html/2606.18206v1","noteUrl":"/review?paper=review-neurips-2026-148540","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148541","title":"Symplectic Neural Operators for Learning Infinite Dimensional Hamiltonian Systems","authors":["Makara Yeang","Yusuke Tanaka","Takashi Matsubara","Takaharu Yaguchi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Symplectic Neural Operators for Learning Infinite Dimensional Hamiltonian Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15881","noteUrl":"/review?paper=review-neurips-2026-148541","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148542","title":"DiffPTS: Rethinking Diffusion ELBO for Probabilistic Time Series Forecasting","authors":["Weiwei Ye","Dongyuan Li","Hangchen Liu","Haotong Jiang","Yoshihide Sekimoto","Renhe Jiang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DiffPTS: Rethinking Diffusion ELBO for Probabilistic Time Series Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.32363","noteUrl":"/review?paper=review-neurips-2026-148542","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148544","title":"Assessing Sample Quality in Conditional Generation under Compositional Shift","authors":["Berker Demirel","Valentino Maiorca","Marco Fumero","Theofanis Karaletsos","Francesco Locatello"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Assessing Sample Quality in Conditional Generation under Compositional Shift","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.09601","noteUrl":"/review?paper=review-neurips-2026-148544","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148546","title":"Intrinsic-dimension empirical Bernstein inequalities for bounded self-adjoint operators","authors":["Diego Martinez Taboada","Aaditya Ramdas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Intrinsic-dimension empirical Bernstein inequalities for bounded self-adjoint operators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15278","noteUrl":"/review?paper=review-neurips-2026-148546","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148548","title":"Do Coding Agents Deceive Us? Detecting and Preventing Cheating via Capped Evaluation with Randomized Tests","authors":["Thanawat Lodkaew","Johannes Ackermann","Soichiro Nishimori","Nontawat Charoenphakdee","Masashi Sugiyama","Takashi Ishida"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Do Coding Agents Deceive Us? Detecting and Preventing Cheating via Capped Evaluation with Randomized Tests","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.07379","noteUrl":"/review?paper=review-neurips-2026-148548","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148549","title":"ReSCUE: Re-translation with Sentence Commitment for Unsegmented Long-Form Simultaneous Sign Language Translation","authors":["Sihan Ren","Gaozheng Li","Yuanshang Quan","Yiming Qin","Fuyi Yang","Chang Liu","Lan Xu","Minye Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ReSCUE: Re-translation with Sentence Commitment for Unsegmented Long-Form Simultaneous Sign Language Translation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03022","noteUrl":"/review?paper=review-neurips-2026-148549","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148550","title":"ASAP: Assembly-Source Aligned Pseudocode Refinement For Binary Decompilation","authors":["Yujian Zhuang","Dehong Gao","Qichao Zhang","QiJing Lai","Jiaxin Wang","Libin Yang","Xiaoyan Cai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ASAP: Assembly-Source Aligned Pseudocode Refinement For Binary Decompilation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06900","noteUrl":"/review?paper=review-neurips-2026-148550","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148552","title":"Heterogeneous Agent Collaborative Reinforcement Learning","authors":["Zhixia Zhang","Zixuan Huang","Gonxun Li","Huaiyang Wang","Chengyi Yuan","Xin Xia","deqing wang","Fuzhen Zhuang","Shuai Ma","Ning Ding","Yaodong Yang","Yikun Ban"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The empirical ratio needs a value on zero-reward batches","plainExplanation":"A positive population mean does not ensure a positive finite-sample mean. The ratio is therefore not a real-valued random variable on all outcomes allowed by its definition, although the theorem’s conditional ratio algebra is valid.","primaryUrl":"https://arxiv.org/html/2603.02604","noteUrl":"/review?paper=review-neurips-2026-148552","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148554","title":"Learning under Localized Minority Imbalance","authors":["Amin Hosseininasab","Steven M Shugan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Cumulative union sampling overshoots the target damage count","plainExplanation":"Cumulative union sampling overshoots the target damage count.","primaryUrl":"https://arxiv.org/html/2610.04936","noteUrl":"/review?paper=review-neurips-2026-148554","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148557","title":"Retrieve What’s Missing: Coverage-Maximizing Retrieval for Consistent Long Video Generation","authors":["Minseok Joo","Dogyun Park","Taehoon Lee","Kyujin Lee","Hyunwoo J. Kim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Retrieve What’s Missing: Coverage-Maximizing Retrieval for Consistent Long Video Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02479","noteUrl":"/review?paper=review-neurips-2026-148557","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148559","title":"Generalization Dynamics of Linear Diffusion Models","authors":["Claudia Lioba Merger","Sebastian Goldt"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The data-prediction gap has an incorrect covariance coefficient","plainExplanation":"Equation (12) gives the test-minus-training mean-squared error for the optimal unregularized affine predictor of clean data. The losses are normalized by $dT$, where $d$ is dimension and $T$ is the number of noise levels. Here $\\Sigma$ and $\\Sigma_0$ are the population and empirical covariance matrices, $(\\lambda_\\nu^0,e_\\nu^0)$ are the eigenpairs of $\\Sigma_0$, and $\\bar\\alpha_t$ is the signal fraction at noise level $t$. The paper omits mean-mismatch terms in this display; a centered example tests the formula without them. The variance mismatch must be multiplied by the square of the residual prediction coefficient. The printed expression does not reduce to that square, even in one dimension. Its leading trace also needs the normalization from the loss definition.","primaryUrl":"https://arxiv.org/html/2505.24769","noteUrl":"/review?paper=review-neurips-2026-148559","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148560","title":"EDISCO: Equivariant Discrete Diffusion for Euclidean Combinatorial Optimization","authors":["Ruogu Chen","Jie Han"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"EDISCO: Equivariant Discrete Diffusion for Euclidean Combinatorial Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04953","noteUrl":"/review?paper=review-neurips-2026-148560","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148562","title":"Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery","authors":["Chi Nguyen Tran","Dao S Minh","Trung-Kiet Huynh","Phu-Hoa Pham","Phu-Quy Nguyen-Lam","Long Tran-Thanh"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11654","noteUrl":"/review?paper=review-neurips-2026-148562","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148563","title":"From Inexact Gradients to Byzantine Robustness: Acceleration and Optimization under Similarity","authors":["Renaud Gaucher","Aymeric Dieuleveut","Hadrien Hendrikx"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The printed accelerated update is algebraically wrong but has an explicit repair","plainExplanation":"The printed recurrence moves away from a nonzero minimizer even with an exact gradient. Its coefficients do not preserve a constant sequence. The corresponding theorem therefore fails as printed, although the underlying three-sequence accelerated method supplies a direct correction. The defects concern the recurrence in Algorithms 2 and 6, its elimination in Appendix B.3, and the transferred bound in Theorem 3.3. They do not refute the underlying accelerated method, the separate similarity method or the experiments.","primaryUrl":"https://arxiv.org/html/2602.03329","noteUrl":"/review?paper=review-neurips-2026-148563","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement","formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148564","title":"Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks","authors":["Young-Jun Lee","Seungone Kim","Minki Kang","Alistair Liang Chuen Cheong","Zerui Chen","Seungho Han","Taehee Jung","Dongyeop Kang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.29082","noteUrl":"/review?paper=review-neurips-2026-148564","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148567","title":"PR-Smoother: Simulator-Preserving Non-Gaussian Smoothing for Data Assimilation","authors":["Yuta Tarumi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PR-Smoother: Simulator-Preserving Non-Gaussian Smoothing for Data Assimilation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.26890","noteUrl":"/review?paper=review-neurips-2026-148567","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148568","title":"RefDecoder: Enhancing Visual Generation with Conditional Video Decoding","authors":["Xiang Fan","Yuheng Wang","Bohan Fang","Jason Ren","Ranjay Krishna"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"RefDecoder: Enhancing Visual Generation with Conditional Video Decoding","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15196","noteUrl":"/review?paper=review-neurips-2026-148568","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148569","title":"Discovering Phase Space Structure in Learned Hamiltonian Systems","authors":["Jiayin Liu","Yulong Yang","Vineet Bansal","Christine Allen-Blanchette"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Entrywise squaring does not ensure dissipative positive semidefiniteness","plainExplanation":"The port-Hamiltonian model parameterizes its dissipation matrix by elementwise squares and claims it is positive semidefinite. Elementwise squaring \\(\\begin{pmatrix}1&2\\\\2&1\\end{pmatrix}\\) gives \\(D=\\begin{pmatrix}1&4\\\\4&1\\end{pmatrix}\\), with eigenvalues \\(-3,5\\).","primaryUrl":"https://arxiv.org/html/2509.23003","noteUrl":"/review?paper=review-neurips-2026-148569","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148571","title":"Trust Region Continual Learning as an Implicit Meta-Learner","authors":["Zekun Wang","Anant Gupta","Christopher MacLellan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Trust Region Continual Learning as an Implicit Meta-Learner","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.02417","noteUrl":"/review?paper=review-neurips-2026-148571","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148572","title":"Zero-Shot Quantization via Weight-Space Arithmetic","authors":["Daniele Solombrino","Antonio Andrea Gargiulo","Alessandro Zirilli","Luca Zhou","Robert Adrian Minut","Emanuele Rodolà"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Avoiding exact permuted equality does not prove gauge alignment","plainExplanation":"Theorem 1 excludes a measure-zero set where two trained endpoints are exactly related by a nonidentity permutation, then interprets the remaining endpoints as sharing a coordinate gauge. The proved exceptional set only concerns exact equality \\(\\Phi_{\\rm QAT}=g\\Phi_{\\rm FT}\\).","primaryUrl":"https://arxiv.org/html/2604.03420","noteUrl":"/review?paper=review-neurips-2026-148572","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148576","title":"HoTS: Homophily-Aware Temperature Scaling for Graph Neural Network Calibration","authors":["Inwoo Tae","Yoontae Hwang","Yongjae Lee"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"HoTS: Homophily-Aware Temperature Scaling for Graph Neural Network Calibration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.32426","noteUrl":"/review?paper=review-neurips-2026-148576","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148582","title":"PithTrain: A Compact and Agent-Native MoE Training System","authors":["Ruihang Lai","Hao Kang","Haozhan Tang","Akaash Raghunathan Parthasarathy","Zichun Yu","Junru Shao","Todd Mowry","Chenyan Xiong","Tianqi Chen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PithTrain: A Compact and Agent-Native MoE Training System","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31463","noteUrl":"/review?paper=review-neurips-2026-148582","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148583","title":"TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning","authors":["Alliot Nagle","Jakhongir Saydaliev","Dhia Garbaya","Michael Gastpar","Ashok Vardhan Makkuva","Hyeji Kim"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The stopping labels are reversed relative to answer arrival","plainExplanation":"The prose and inference stop rule treat label one as “answer already arrived,” but Equation (3) assigns one strictly before the answer-arrival index \\(i^*\\). Binary cross-entropy with those labels rewards high stopping probability before arrival and low probability afterward.","primaryUrl":"https://arxiv.org/html/2603.12529","noteUrl":"/review?paper=review-neurips-2026-148583","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148585","title":"Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning","authors":["Yu-Ang Lee","Ching-Yun Ko","Pin-Yu Chen","Mi-Yen Yeh"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04998","noteUrl":"/review?paper=review-neurips-2026-148585","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148587","title":"Opening the Black Box of Classifier-Free Guidance via Information Bottleneck","authors":["Jiayang Gao","Tianyi Zheng","Jiayang Zou","Fengxiang Yang","Luyao Fan","Lv Tang","Bo Li","Jia Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Nondegeneracy in the Fisher-balance theorem","plainExplanation":"Stationarity makes this product zero. It cannot justify cancelling \\(g(t)^2\\) when the diffusion has stopped.","primaryUrl":"https://arxiv.org/html/2610.13369","noteUrl":"/review?paper=review-neurips-2026-148587","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148592","title":"HESTIA: A Hessian-Guided Differentiable Quantization-Aware Training Framework for Extremely Low-Bit LLMs","authors":["Guoan Wang","Feiyu Wang","Zongwei Lv","Yikun Zong","Zhewen Tan","Tong Yang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The multi-probe residual trace lacks averaging","plainExplanation":"The multi-probe residual trace lacks averaging.","primaryUrl":"https://arxiv.org/html/2601.20745","noteUrl":"/review?paper=review-neurips-2026-148592","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148594","title":"When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding","authors":["Xu Yang","Lunyiu Nie","Ethan Chandra","Stanislav Gannutin","Fangru Lin","Swarat Chaudhuri"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A local lift can leave the global bottleneck unchanged and add communication cost","plainExplanation":"A local lift can leave the global bottleneck unchanged and add communication cost.","primaryUrl":"https://arxiv.org/html/2606.00953","noteUrl":"/review?paper=review-neurips-2026-148594","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148595","title":"DarkVGGT: Seeing Through Darkness Using Thermal Geometry without Daylight Tax","authors":["Minseong Kweon","Wenyuan Zhao","Nuo Chen","Lulin Liu","Tracy Han","Zihao Zhu","Srinivas Shakkottai","Chao Tian","Zhiwen Fan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DarkVGGT: Seeing Through Darkness Using Thermal Geometry without Daylight Tax","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11326","noteUrl":"/review?paper=review-neurips-2026-148595","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148598","title":"Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs","authors":["Robert Graham","Edward Stevinson","Yariv Barsheshat"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.14888","noteUrl":"/review?paper=review-neurips-2026-148598","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148599","title":"Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators","authors":["Chenming Zhu","Jingli Lin","Yilin Long","Peizhou Cao","Tai WANG","Jiangmiao Pang","Xihui Liu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06476","noteUrl":"/review?paper=review-neurips-2026-148599","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148600","title":"Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning","authors":["Xuan Gong","HanBo Huang","Hao Zheng","Yiran Zhang","Wenbin Dai","Weishu Zhao","Shiyu Liang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09614","noteUrl":"/review?paper=review-neurips-2026-148600","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148601","title":"Inadvertent Context Leakage in Language Models","authors":["Jaiden Fairoze","Neal Mangaokar","Kamalika Chaudhuri","Sanjam Garg","Saeed Mahloujifar"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Inadvertent Context Leakage in Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.19857","noteUrl":"/review?paper=review-neurips-2026-148601","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148602","title":"Eliciting Secret Knowledge from Language Models","authors":["Bartosz Cywiński","Emil Ryd","Rowan Wang","Senthooran Rajamanoharan","Neel Nanda","Arthur Conmy","Samuel Marks"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Eliciting Secret Knowledge from Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.01070","noteUrl":"/review?paper=review-neurips-2026-148602","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148605","title":"DocAtlas: Long-Document Understanding as Mutable-State Interaction","authors":["Hongchen Wei","Yuanzhe Wang","Bei Liu","Yifan Yang","Qi Dai","Kai Qiu","Yunsheng Li","Dongdong Chen","Chong Luo","Zhenzhong Chen","Baining Guo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DocAtlas: Long-Document Understanding as Mutable-State Interaction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.07527","noteUrl":"/review?paper=review-neurips-2026-148605","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148606","title":"Log-Likelihood, Simpson’s Paradox, and the Detection of Machine-Generated Text","authors":["Tom Kempton","Viktor Drobnyi","Maeve Madigan","Stuart Burrell"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Log-Likelihood, Simpson’s Paradox, and the Detection of Machine-Generated Text","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06294","noteUrl":"/review?paper=review-neurips-2026-148606","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148608","title":"Know What You Need: Efficient Activation Checkpointing","authors":["Danyal Ahmed","Edouard Oyallon"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Know What You Need: Efficient Activation Checkpointing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://hal.science/hal-05708820v1/file/main.pdf","noteUrl":"/review?paper=review-neurips-2026-148608","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148612","title":"Uncovering and Shaping the Latent Representation of 3D Scene Topology in Vision-Language Models","authors":["Haoming Wang","Wei Gao"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The unpenalized constant mode prevents the claimed uniqueness","plainExplanation":"The Laplacian annihilates the constant vector. The singular value assigned to that direction therefore has zero coefficient in the energy and can exceed its lower bound without increasing the objective. This produces minimizers with different singular values, which cannot be related by a right-orthogonal transformation. The defect is the forced saturation of the constant singular mode and the uniqueness claim in Theorem 1(a). Mean-centering removes that free component from the minimizing family, leaving its nonconstant eigenmap directions. Uniqueness of individual directions also requires spectral gaps; the corrected uncentered optimization does not establish the separate continuum-limit or statistical theorems.","primaryUrl":"https://arxiv.org/html/2605.07148","noteUrl":"/review?paper=review-neurips-2026-148612","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148615","title":"Task-Induced Riemannian Metrics for Vision Transformer Feature Spaces","authors":["Andrew Bond","Ege E Özlü","Tuna Çimen","Ilkin U Melanlioglu","Tolga Birdal","Erkut Erdem","Aykut Erdem"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Task-Induced Riemannian Metrics for Vision Transformer Feature Spaces","plainExplanation":"Decoder error can come entirely from a kept token while the claimed bound is zero. The corrected bound must include changes to every final token feature.","primaryUrl":"https://arxiv.org/html/2609.27988","noteUrl":"/review?paper=review-neurips-2026-148615","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-148616","title":"L$^2$EAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks","authors":["Po-Nien Kung","Linfeng Song","Dawsen Hwang","Jinsung Yoon","Chun-Liang Li","Simone Severini","Miroslav Olšák","Edward Lockhart","Quoc V Le","Burak Gokturk","Thang Luong","Tomas Pfister","Nanyun Peng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"L$^2$EAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03303","noteUrl":"/review?paper=review-neurips-2026-148616","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148617","title":"Factor Augmented High-Dimensional SGD","authors":["Shubo Li","Yuefeng Han","Xiufan Yu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Sub-Gaussian tails do not give the assumed almost-sure covariance bound","plainExplanation":"Sub-Gaussianity controls tail probabilities. It does not bound every possible covariance update by a finite deterministic constant. The claimed implication fails even for a one-dimensional Gaussian factor, with no idiosyncratic noise. The error is the asserted applicability of the bounded-update analysis to general sub-Gaussian data, including the substitution used in Remark 3.7. It does not contradict an Oja or FSGD theorem that retains Assumption 2.3 as a separate hypothesis.","primaryUrl":"https://arxiv.org/html/2605.19291","noteUrl":"/review?paper=review-neurips-2026-148617","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148618","title":"When Are Teacher Tokens Reliable? Position-Weighted On-Policy Self-Distillation for Reasoning","authors":["Xiaogeng Liu","Xinyan Wang","Yingzi Ma","Yechao Zhang","Chaowei Xiao"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"When Are Teacher Tokens Reliable? Position-Weighted On-Policy Self-Distillation for Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21606","noteUrl":"/review?paper=review-neurips-2026-148618","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148620","title":"Feature-Context Consistency: Unsupervised Adversarial Detection with Drift Stability on Attributed Graphs","authors":["Firas Bayram","Maxime Cordy"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Feature-Context Consistency: Unsupervised Adversarial Detection with Drift Stability on Attributed Graphs","plainExplanation":"Positive coordinate covariances do not imply positive mean cosine alignment: large aligned observations can dominate covariance while most normalized pairs are opposed. The separation theorem remains conditional on its separate positive-cosine assumption.","primaryUrl":"https://arxiv.org/html/2610.16694","noteUrl":"/review?paper=review-neurips-2026-148620","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148621","title":"MARS: Harmonizing Multimodal Convergence via Adaptive Rank Search","authors":["Minkyoung Cho","Insu Jang","Shuowei Jin","Zesen Zhao","Adityan Jothi","Ethem F Can","Min-Hung Chen","Zhuoqing Morley Mao"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"MARS: Harmonizing Multimodal Convergence via Adaptive Rank Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.00720","noteUrl":"/review?paper=review-neurips-2026-148621","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148624","title":"Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules","authors":["Riccardo De Santi","Bruce D Lee","Cristian Jensen","Kimon Protopapas","Sophia Tang","Chenghao Liu","Pranam Chatterjee","Yisong Yue","Andreas Krause"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The sample condition puts the logarithmic factor on the wrong side","plainExplanation":"The sample condition puts the logarithmic factor on the wrong side.","primaryUrl":"https://arxiv.org/html/2606.08802","noteUrl":"/review?paper=review-neurips-2026-148624","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148625","title":"CLEF: EEG Foundation Models for Learning Clinical Semantics","authors":["Peng Cao","Ali Mirzazadeh","Jong W Lee","Aleksandar Videnovic","Dina Katabi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CLEF: EEG Foundation Models for Learning Clinical Semantics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10817","noteUrl":"/review?paper=review-neurips-2026-148625","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148627","title":"CounterFlowNet: From Minimal Changes to Meaningful Counterfactual Explanations","authors":["Oleksii Furman","Patryk Marszałek","Jan Masłowski","Marek Śmieja","Maciej Zieba","Piotr Gaiński"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CounterFlowNet: From Minimal Changes to Meaningful Counterfactual Explanations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.17244","noteUrl":"/review?paper=review-neurips-2026-148627","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148630","title":"When Does Trimming Help Conformal Prediction? A Retained-Law Diagnostic under Calibration Contamination","authors":["Congye Wang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"When Does Trimming Help Conformal Prediction? A Retained-Law Diagnostic under Calibration Contamination","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06204","noteUrl":"/review?paper=review-neurips-2026-148630","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148631","title":"Advectra: Asymmetric Latent Transport for Non-Stationary Physics","authors":["Nodens Koren","Thomas Hofmann","Georgios Kissas"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Advectra: Asymmetric Latent Transport for Non-Stationary Physics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.05098","noteUrl":"/review?paper=review-neurips-2026-148631","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148632","title":"Neural Networks With Dense Weights Are Not Universal Approximators","authors":["Levi Rauchwerger","Stefanie Jegelka","Ron Levie"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neural Networks With Dense Weights Are Not Universal Approximators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.07618","noteUrl":"/review?paper=review-neurips-2026-148632","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148636","title":"Sub-Gaussian Confidence Intervals for Heavy-Tailed Data: Characterizing the Limits of Inference","authors":["Ilyes Hammouda","Stanislav Minsker","Mohamed Ndaoud"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Sub-Gaussian Confidence Intervals for Heavy-Tailed Data: Characterizing the Limits of Inference","plainExplanation":"The proposed variance estimator can collapse inside the claimed distribution class. The test and interval also use a radius too small for their stated mean-error bound.","primaryUrl":"https://arxiv.org/html/2610.15799","noteUrl":"/review?paper=review-neurips-2026-148636","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-148639","title":"Learning to Persuade Exposes How Easily LLMs Abandon Correct Beliefs","authors":["Nimet Beyza Bozdag","Emre Can Acikgoz","Gokhan Tur","Dilek Hakkani-Tur"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning to Persuade Exposes How Easily LLMs Abandon Correct Beliefs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.11624","noteUrl":"/review?paper=review-neurips-2026-148639","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148640","title":"Can LLMs Take Retrieved Information with a Grain of Salt?","authors":["Behzad Shayegh","Mohamed Osama Ahmed","Fred Tung","Leo Feng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Can LLMs Take Retrieved Information with a Grain of Salt?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06919","noteUrl":"/review?paper=review-neurips-2026-148640","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148644","title":"Breakeven complexity: A new perspective on neural partial differential equation solvers","authors":["Yijing Zhang","Nicholas Roberts","Tanya Marwah","Misha Khodak"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Breakeven complexity: A new perspective on neural partial differential equation solvers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15399","noteUrl":"/review?paper=review-neurips-2026-148644","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148649","title":"Geometry-Calibrated Conformal Abstention for Language Models","authors":["Rui Xu","Yi Chen","Sihong Xie","Hui Xiong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The token-stability certificate omits the change in the competing logit","plainExplanation":"Exchangeability gives \\(\\Pr(J_{n+1}=1)=\\mathbb E[c]/n\\), not \\(c/n\\). The identity and calibration-conditional interpretation of Equation (6) fail. A marginal expectation-based replacement may be possible but requires a new argument; the empirical abstention procedure is not refuted.","primaryUrl":"https://arxiv.org/html/2604.27914","noteUrl":"/review?paper=review-neurips-2026-148649","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148651","title":"Continuous Audio Thinking for Large Audio Language Models","authors":["Gyojin Han","DongJae Lee","Changho Choi","Jongsuk Kim","Junmo Kim"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Continuous Audio Thinking for Large Audio Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.18273","noteUrl":"/review?paper=review-neurips-2026-148651","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148653","title":"Flow Annealing Posterior Sampling for Function-Space Regression and Inverse Problems","authors":["Yaozhong Shi","Zachary Ross","Yisong Yue"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Flow Annealing Posterior Sampling for Function-Space Regression and Inverse Problems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.22346","noteUrl":"/review?paper=review-neurips-2026-148653","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148654","title":"Is Dimensionality a Barrier for Retrieval Models?","authors":["Kiril Bangachev","Guy Bresler","Jonathan Kogan","Yury Polyanskiy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Is Dimensionality a Barrier for Retrieval Models?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23556","noteUrl":"/review?paper=review-neurips-2026-148654","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148657","title":"Internal Safety Collapse in Frontier Large Language Models","authors":["Oscar W","Xiao Liu","Hanxun Huang","Yige Li","Xiang Zheng","Yifeng Gao","Cong Wang","Bo Li","Xingjun Ma","Yu-Gang Jiang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Internal Safety Collapse in Frontier Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.23509","noteUrl":"/review?paper=review-neurips-2026-148657","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148658","title":"Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering","authors":["Gerasimos Chatzoudis","Zhuowei Li","Gemma Moran","Hao Wang","Dimitris Metaxas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2506.01247","noteUrl":"/review?paper=review-neurips-2026-148658","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148660","title":"A Composite Activation Function for Learning Stable Binary Representations","authors":["Seokhun Park","Choeun Kim","Kwanho Lee","Sehyun Park","Insung Kong","Yongdai Kim"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"A Composite Activation Function for Learning Stable Binary Representations","plainExplanation":"The three-unit square approximation has the output sign reversed; flipping that sign restores its local error estimate. Version 3 already removes the earlier explicit tail prefactor and adds the parameter restriction that excludes the proposed negative-risk example.","primaryUrl":"https://arxiv.org/html/2605.11558","noteUrl":"/review?paper=review-neurips-2026-148660","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148663","title":"Otter Weather: Skillful and computationally-efficient medium-range weather forecasting","authors":["Cristiana Diaconu","Jonas Scholz","Aliaksandra Shysheya","Stratis Markou","Payel Mukhopadhyay","Miles Cranmer","Richard Turner"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Otter Weather: Skillful and computationally-efficient medium-range weather forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.26421","noteUrl":"/review?paper=review-neurips-2026-148663","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148664","title":"LINC: Decoupling Local Consequence Scoring from Hidden Matching in Constructive Neural Routing","authors":["ShaoFeng Qin","Li Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LINC: Decoupling Local Consequence Scoring from Hidden Matching in Constructive Neural Routing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06332","noteUrl":"/review?paper=review-neurips-2026-148664","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148667","title":"Metropolis-Adjusted Diffusion Models","authors":["Kevin H. Lam","Tyler Farghly","Christopher Williams","Jun Yang","Yee Whye Teh","Arnaud Doucet"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Metropolis-Adjusted Diffusion Models","plainExplanation":"The MH–Barker asymptotic-variance comparison needs an additive stationary-variance term. An exact-target Gaussian proposal gives Barker variance three times the MH variance, while the cited bound with the additive term holds.","primaryUrl":"https://arxiv.org/html/2605.09654","noteUrl":"/review?paper=review-neurips-2026-148667","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148668","title":"Simulating Human Memory with Language Models","authors":["Qihan Wang","Nicholas Tomlin","Michael Hu","Brian Dillon","Tal Linzen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Simulating Human Memory with Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25680","noteUrl":"/review?paper=review-neurips-2026-148668","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148670","title":"TagBO: LLM-Driven Task-Aware Graph Bayesian Optimization for Scientific Discovery","authors":["Xinzhe Yuan","Zimu Zeng","Zhuo Chen","Jianshu Zhang","Jinzong Dong","Jingyi Chen","Nanyang Ye","Huan Xiong","Qinying Gu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The scalarization conclusion needs a nonzero mixture coefficient","plainExplanation":"The scalarization conclusion needs a nonzero mixture coefficient.","primaryUrl":"https://arxiv.org/html/2610.14626","noteUrl":"/review?paper=review-neurips-2026-148670","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148676","title":"Focused Forcing: Content-Aware Per-Frame KV Selection for Efficient Autoregressive Video Diffusion","authors":["Peiliang Cai","Evelyn Zhang","Jiacheng Liu","Hao Lin","Ruiqi Zhang","Weile Mo","Yue Ma","Shikang Zheng","Jiehang Huang","Dongrui Liu","Linfeng Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Focused Forcing: Content-Aware Per-Frame KV Selection for Efficient Autoregressive Video Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18346","noteUrl":"/review?paper=review-neurips-2026-148676","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148677","title":"InfoFlow: A Framework for Multi-Layer Transformer Analysis","authors":["Penghao Yu","Haotian Jiang","Zeyu Bao","Qianxiao Li"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Distinct retrieval matrices do not establish the claimed head lower bound","plainExplanation":"Distinct matrices can repeat the same retrieval task, so their number alone does not give the asserted minimum number of heads. The encoding-cost argument also uses a necessary parameter lower bound as an upper rate.","primaryUrl":"https://arxiv.org/html/2605.17930","noteUrl":"/review?paper=review-neurips-2026-148677","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148679","title":"Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching","authors":["Junwan Kim","Jiho Park","Seonghu Jeon","Seungryong Kim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.05951","noteUrl":"/review?paper=review-neurips-2026-148679","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148680","title":"dRAE: Representation Autoencoder with Hyper-Spherical Codes","authors":["Tianren Ma","Lin Long","Chuyan Chen","Mu Zhang","Junbo Zhao","Tong Zhang","Qixiang Ye"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"dRAE: Representation Autoencoder with Hyper-Spherical Codes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.22148","noteUrl":"/review?paper=review-neurips-2026-148680","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148681","title":"EVA-0: Test-Time Model Evolution with Only Two Forward Passes per Sample","authors":["Guohao Chen","Shuaicheng Niu","Geng Li","Yunbei Zhang","Shilin Shan","Chunyan Miao","Jianfei Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EVA-0: Test-Time Model Evolution with Only Two Forward Passes per Sample","plainExplanation":"Zero-mean loss noise can depend on the Gaussian perturbation and cancel the entire gradient estimator. Independence or suitable conditional centering supplies the missing assumption for Proposition 1’s mean and variance formulas.","primaryUrl":"https://arxiv.org/html/2605.18867","noteUrl":"/review?paper=review-neurips-2026-148681","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148684","title":"PanoWorld: Geometry-Consistent Panoramic Video World Modeling","authors":["Le Jiang","Xiangyu Bai","Bishoy Galoaa","Shayda Moezzi","Caleb J Lee","Tooba Imtiaz","Edmund Yeh","Jennifer Dy","Yanzhi Wang","Sarah Ostadabbas"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The finite-difference loss terms are not latitude weighted","plainExplanation":"The finite-difference loss terms are not latitude weighted.","primaryUrl":"https://arxiv.org/html/2605.15391","noteUrl":"/review?paper=review-neurips-2026-148684","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148686","title":"SRL-MPC: Shape-Aware Reinforcement Learned Model Predictive Control","authors":["Ruihua Han","Rui Gao","Zhe Liu","Xinyi WANG","Chang Chen","Shuai Wang","Qi Hao","Jia Pan","Hengshuang Zhao"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The shortest-segment unit direction is undefined when occupied sets intersect","plainExplanation":"The shortest-segment unit direction is undefined when occupied sets intersect.","primaryUrl":"https://arxiv.org/html/2608.21175","noteUrl":"/review?paper=review-neurips-2026-148686","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148687","title":"Towards On-Policy Data Evolution for Visual-Native Multimodal Deep Search Agents","authors":["Shijue Huang","Hangyu Guo","Guanting Dong","Chenxin Li","Junting Lu","Xinyu Geng","Zhaochen Su","Zhenyu Li","Shuang Chen","Hongru WANG","Yi R. (May) Fung"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Towards On-Policy Data Evolution for Visual-Native Multimodal Deep Search Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10832","noteUrl":"/review?paper=review-neurips-2026-148687","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148688","title":"TAC: Timestamped Audio Captioning","authors":["Sonal Kumar","Prem Seetharaman","Ke Chen","Oriol Nieto","Jiaqi Su","Zhepei Wang","Rithesh Kumar","Dinesh Manocha","Nicholas J. Bryan","Zeyu Jin","Justin Salamon"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TAC: Timestamped Audio Captioning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.15766","noteUrl":"/review?paper=review-neurips-2026-148688","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148689","title":"Causal pieces: analysing and improving spiking neural networks piece by piece","authors":["Dominik Dold","Philipp Petersen"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Causal pieces: analysing and improving spiking neural networks piece by piece","plainExplanation":"No issue was found in the checked claims. The asymmetric mixture does not satisfy the theorem’s symmetry assumption. The original v1 already contains that assumption, so its presence is not a later correction to a missing condition.","primaryUrl":"https://arxiv.org/html/2504.14015","noteUrl":"/review?paper=review-neurips-2026-148689","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148692","title":"Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer","authors":["Yulun Wu","Sravan Kumar Ankireddy","Samuel Sharpe","Nikita Seleznev","Dehao Yuan","Hyeji Kim","Nam Nguyen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Clipping the advantage product does not implement PPO ratio clipping","plainExplanation":"Clipping the advantage product does not implement PPO ratio clipping.","primaryUrl":"https://arxiv.org/html/2603.26097","noteUrl":"/review?paper=review-neurips-2026-148692","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148693","title":"SnapMLA: Efficient Long-Context MLA Decoding via Hardware-Aware FP8 Quantized Pipelining","authors":["Yifan Zhang","Zunhai Su","Shuhao Hu","YangRui","Wei Wu","Yulei Qian","Yuchen Xie","Xunliang Cai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SnapMLA: Efficient Long-Context MLA Decoding via Hardware-Aware FP8 Quantized Pipelining","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.10718","noteUrl":"/review?paper=review-neurips-2026-148693","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148695","title":"DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models","authors":["Amin Karimi Monsefi","Dominic Culver","Nikhil Bhendawade","Lokesh Boominathan","Manuel R Ciosici","Yizhe Zhang","Irina Belousova"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.16342","noteUrl":"/review?paper=review-neurips-2026-148695","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148696","title":"DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts","authors":["Jiading Gai","Shuai Zhang","Xiang song","Yuyang (Bernie) Wang","George Karypis"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15422","noteUrl":"/review?paper=review-neurips-2026-148696","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148697","title":"Rationale-Guided Policy Optimization: Learning to Reason with Adaptive Rationale Scaffolding","authors":["Hoang Phan","Minh Pham","Chau Pham","Chinmay Hegde","Trung Le","Qi Lei"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Rationale-Guided Policy Optimization: Learning to Reason with Adaptive Rationale Scaffolding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07342","noteUrl":"/review?paper=review-neurips-2026-148697","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148698","title":"When and Why is Optimistic Multiplicative Weights Slow? The Geometry of Energy Dissipation","authors":["John Lazarsfeld","Anas Barakat","Georgios Piliouras","Antonios Varvitsiotis","Andre Wibisono"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"When and Why is Optimistic Multiplicative Weights Slow? The Geometry of Energy Dissipation","plainExplanation":"The proofs establish one-sided bounds, not equality. In particular, restricting a diagonal quadratic form to the zero-sum subspace can change its extreme eigenvalues. Its unrestricted largest diagonal entry need not be attainable in that subspace.","primaryUrl":"https://arxiv.org/html/2605.13242","noteUrl":"/review?paper=review-neurips-2026-148698","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","incorrect_bound","invalid_inference","invalid_reduction","lemma_error","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148700","title":"Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS","authors":["Chaeeun Kim","Seungone Kim"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.16409","noteUrl":"/review?paper=review-neurips-2026-148700","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148701","title":"Selectivity and Shape in the Design of Forward-Forward Goodness Functions","authors":["Talha Rüzgar Akkuş","Şuayp T Kocabay","Kamer A Yuksel","Hassan Sawaf"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Selectivity and Shape in the Design of Forward-Forward Goodness Functions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.13081","noteUrl":"/review?paper=review-neurips-2026-148701","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148703","title":"Perturb and Correct: Post-Hoc Ensembles using Affine Redundancy","authors":["Eleanor Quint"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Perturb and Correct: Post-Hoc Ensembles using Affine Redundancy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.01632","noteUrl":"/review?paper=review-neurips-2026-148703","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148705","title":"The Shape of Events: Edge-Based Inductive Biases via Cross-Domain Distillation","authors":["Shunsuke Yasuki","Masato Taki","Soshun Kihara"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Shape of Events: Edge-Based Inductive Biases via Cross-Domain Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30478","noteUrl":"/review?paper=review-neurips-2026-148705","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148709","title":"Optimal Hidden-Target Learning for Online Inventory Optimization on General Convex Sets","authors":["Anthony Pineci","Yunzong Xu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Optimal Hidden-Target Learning for Online Inventory Optimization on General Convex Sets","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.14679","noteUrl":"/review?paper=review-neurips-2026-148709","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148713","title":"Uncertainty Estimation for Pretrained Medical Image Registration Models via Transformation Equivariance","authors":["Lin Tian","Xiaoling Hu","Juan Eugenio Iglesias"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The perturbation Jacobian must be evaluated before the diffeomorphism","plainExplanation":"The derivation expands \\(\\tau(v+\\epsilon)\\), with \\(v=\\tau^{-1}(\\phi)\\), but evaluates the Jacobian at \\(\\phi\\) rather than \\(v\\). For \\(\\tau(x)=x+x^3\\), \\(\\phi=2\\) and \\(v=1\\), the expansion is \\(2+4\\epsilon+3\\epsilon^2+\\epsilon^3\\).","primaryUrl":"https://arxiv.org/html/2509.23355","noteUrl":"/review?paper=review-neurips-2026-148713","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148714","title":"QT-Net: Rethinking Evaluation of AI Models in Atomic Chemical Space","authors":["Pablo Martínez Crespo","Stefano Ribes","Martin Rahm","Richard Johannes Maximilian Beckmann","Robert Jordan","Marisa Gliege","Santiago Miret","Vijay K Narasimhan","Rocío Mercado"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The displayed atom weights do not give equal species totals","plainExplanation":"The displayed atom weights do not give equal species totals. A small within-repeat ICC does not establish independence across repeated folds. Fold ICC and marginal variance tests do not establish sphericity.","primaryUrl":"https://arxiv.org/html/2605.10458","noteUrl":"/review?paper=review-neurips-2026-148714","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148715","title":"Useful Memories Become Faulty When Continuously Updated by LLMs","authors":["Dylan Zhang","Yanshan Lin","Zhengkun Wu","Yihang Sun","Bingxuan Li","Dianqi Li","Hao Peng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Useful Memories Become Faulty When Continuously Updated by LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12978","noteUrl":"/review?paper=review-neurips-2026-148715","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148716","title":"Compute Efficiency and Serial Runtime Tradeoffs for Stochastic Momentum Methods","authors":["Depen Morwani","Alexandru Meterez","Pranav Nair","Sham Kakade"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The covariance recurrence and secular equation use different noise operators","plainExplanation":"The covariance recurrence and secular equation use different noise operators. The finite-dimensional power-law tail has no uniform stated lower comparison.","primaryUrl":"https://arxiv.org/html/2606.19179","noteUrl":"/review?paper=review-neurips-2026-148716","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148717","title":"Refinement as a Service: Algorithmic Predictor Refinement","authors":["Wei Tang","Hanrui Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Two introductory notions need alignment with the paper’s formal definitions","plainExplanation":"At a particular boundary observation, the probability restrictions \\(0\\le y_i\\le\\lambda_i\\) can uniquely determine more quantities than the linear equations alone. The equivalence needs to specify that these additional inequalities are not part of its structural observability test.","primaryUrl":"https://arxiv.org/html/2610.11415","noteUrl":"/review?paper=review-neurips-2026-148717","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error","missing_assumption","notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148719","title":"Spherical Interpolation for Backward-Compatible Multimodal Representations","authors":["Simone Ricci","Niccolò Biondi","Federico Pernici"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Spherical Interpolation for Backward-Compatible Multimodal Representations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39836","noteUrl":"/review?paper=review-neurips-2026-148719","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148720","title":"Working with AI: Measuring the Applicability of Generative AI to Occupations","authors":["Kiran Tomlinson","Sonia Jaffe","Will Wang","Scott Counts","Siddharth Suri"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Working with AI: Measuring the Applicability of Generative AI to Occupations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2507.07935","noteUrl":"/review?paper=review-neurips-2026-148720","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148722","title":"Aligning Language Models with Selective Prediction","authors":["Gaoxiang Luo","Yifan Wu","Sinian Zhang","Aryan Deshwal","Ju Sun"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Finite temperature scaling preserves the full probability range","plainExplanation":"Appendix C.6.5 applies temperature scaling to a probability $s$. It defines $f_T(s)=\\sigma(\\operatorname{logit}(s)/T)$, where $\\sigma(z)=1/(1+e^{-z})$ and $T>0$. After reporting $T\\approx47.9$, the text says the full domain $[0,1]$ is compressed into approximately $[0.38,0.62]$. The claimed full-domain compression is false: every finite positive temperature leaves the interior probability range equal to \\((0,1)\\), with endpoint limits zero and one. A narrow output band requires restricted inputs or an explicit clipping convention.","primaryUrl":"https://arxiv.org/html/2607.03528","noteUrl":"/review?paper=review-neurips-2026-148722","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["false_statement"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148723","title":"Oracle Supervision Transfers for Hyperparameter Prediction in Model-Based Image Denoising","authors":["Jianmin Liao","Lixin Shen","Yuesheng Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Oracle Supervision Transfers for Hyperparameter Prediction in Model-Based Image Denoising","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20479","noteUrl":"/review?paper=review-neurips-2026-148723","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148724","title":"On the Nature of Attention Sink that Shapes Decoding Strategy in Omni-LLMs","authors":["Suho Yoo","Youngjoon Jang","Joon Son Chung"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Shared sink values do not give a uniform additive vector","plainExplanation":"The sink values are shared across queries, but their weighted sums need not be identical. Strong attention to the same sink is insufficient for a uniform additive bias.","primaryUrl":"https://arxiv.org/html/2603.14337","noteUrl":"/review?paper=review-neurips-2026-148724","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148726","title":"Structure-Semantic Co-optimized Latent Diffusion Model for Fast Visual Anagram Synthesis","authors":["Xiang Gao","Yunpeng Jia"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Two formula-level corrections are needed","plainExplanation":"An inclusive maximum threshold retains at least one maximizer. Its selected objective can have a nonzero latent gradient, so this endpoint does not algebraically disable the update.","primaryUrl":"https://arxiv.org/html/2606.16241","noteUrl":"/review?paper=review-neurips-2026-148726","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148730","title":"Aegis: Generative Gradient Masking for Privacy-Preserving Medical Federated Learning","authors":["Chaoyu Zhang","Shanghao Shi","Heng Jin","Ning Wang","Thomas Hou","Wenjing Lou"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Mixed-gradient moment assumptions","plainExplanation":"The missing term is \\(2\\mathbb E\\langle u,v\\rangle\\). The discrepancy budget does not control its sign, and a matched synthetic sample can make it positive.","primaryUrl":"https://arxiv.org/html/2609.38339","noteUrl":"/review?paper=review-neurips-2026-148730","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148731","title":"Conformal Selective Acting: Anytime-Valid Risk Control for RLVR-Trained LLMs","authors":["Hamed Khosravi","Xiaoming Huo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Pointwise surrogate convergence conflicts with the fixed risk-slope neighborhood","plainExplanation":"For Algorithm 1’s score and a deterministic verifier, uniform pointwise surrogate convergence creates a central zero-risk plateau incompatible with a fixed-width strictly increasing risk neighborhood. The existing coarsened-feature repair restores compatible premises and the stated risk-approximation bound; it does not establish every downstream utility claim.","primaryUrl":"https://arxiv.org/html/2605.20270","noteUrl":"/review?paper=review-neurips-2026-148731","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148732","title":"OASES: Outcome-Aligned Search-Evaluation Co-Training for Agentic Search","authors":["Erhan Zhang","Yiqun Chen","Zechun Niu","Wei Yang","Xiaochi Wei","Yan Gao","YIWU","Yao Hu","Jiaxin Mao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OASES: Outcome-Aligned Search-Evaluation Co-Training for Agentic Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.03675","noteUrl":"/review?paper=review-neurips-2026-148732","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148733","title":"Efficient Adjoint Matching for Fine-tuning Diffusion Models","authors":["Jeongwoo Shin","Dongsoo Shin","Yuchen Zhu","Wei Guo","Yongxin Chen","Joonseok Lee","Jaewoong Choi","Jaemoo Choi"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Endpoint proof and singular-boundary convention","plainExplanation":"The endpoint bridge identity is tautological. It supplies neither the marginal variance of \\(X_1\\) nor its independence from an initial variable. The singular diffusion schedule also requires an explicit boundary interpretation.","primaryUrl":"https://arxiv.org/html/2605.11480","noteUrl":"/review?paper=review-neurips-2026-148733","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148735","title":"Going Down Memory Lane: Scaling Tokens for Video Stream Understanding with Dynamic KV-Cache Memory","authors":["Vatsal Agarwal","Saksham Suri","Matthew Gwilliam","Pulkit Kumar","Abhinav Shrivastava"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The inclusive cache window contains one extra frame","plainExplanation":"The stated \\(\\omega\\)-frame cache includes \\(\\omega+1\\) frame indices. Its inclusive lower endpoint is one frame too early for the claimed \\(\\omega N\\)-token size. This is an off-by-one in the mathematical window definition. It does not establish an error in the executed cache.","primaryUrl":"https://arxiv.org/html/2602.18434","noteUrl":"/review?paper=review-neurips-2026-148735","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148736","title":"Refining Compositional Diffusion for Reliable Long-Horizon Planning","authors":["Kyowoon Lee","Yunhao Luo","Anh Tong","Jaesik Choi"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The stated density guarantees do not follow from the guidance objectives","plainExplanation":"The unweighted DDPM loss is not the claimed likelihood upper bound, and adding a static energy gradient does not generally produce the stated Gibbs endpoint distribution.","primaryUrl":"https://arxiv.org/html/2605.03075","noteUrl":"/review?paper=review-neurips-2026-148736","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148737","title":"Online Localized Conformal Prediction","authors":["Yuheng Lai","Garvesh Raskutti"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Online Localized Conformal Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05497","noteUrl":"/review?paper=review-neurips-2026-148737","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148738","title":"How Much is Left? LLMs Linearly Encode Their Remaining Output Length","authors":["Mohamed Amine Merzouk","Dmitri Carpov","Mirko Bronzi","Damiano Fornasiere","Adam Oberman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The countdown subtracts one too many under the stated token indexing","plainExplanation":"The countdown subtracts one too many under the stated token indexing. Clamping invalidates the asserted countdown MAE identity.","primaryUrl":"https://arxiv.org/html/2607.05316","noteUrl":"/review?paper=review-neurips-2026-148738","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148740","title":"SpatialFlow-GRPO: Where Spatial Credit Drives Image Editing","authors":["Yankai Yang","Yancheng Long","Wei Chen","Xingyu Lu","Hongyang Wei","Bin Wen","Fan Yang","Tingting Gao","Han Li","Shuo Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SpatialFlow-GRPO: Where Spatial Credit Drives Image Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.26872","noteUrl":"/review?paper=review-neurips-2026-148740","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148742","title":"Efficient bias mitigation in T2I diffusion models using Concept Graphs","authors":["Mansi -","Avinash Kori","Francesco Leofante"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Self-comparisons shift the soft rank outside its stated range","plainExplanation":"The soft-rank formula includes a self-comparison worth \\(1/2\\), adding a uniform offset. For well-separated scores and small temperature, its maximum exceeds the claimed upper bound one. The error is a rank offset and range claim. Pairwise sigmoid complements make each row sum constant, so this uniform shift changes the squared rank loss by a constant and preserves its training gradients.","primaryUrl":"https://arxiv.org/html/2607.03397","noteUrl":"/review?paper=review-neurips-2026-148742","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148744","title":"SteerVTE: Seamless Video Text Editing with Style and Glyph Control","authors":["Kai Zeng","Moran Li","Zhengwei Wang","Yingchen Yu","Yiheng Lin","Ruichuan An","Ming Lu","Qi She","Wentao Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SteerVTE: Seamless Video Text Editing with Style and Glyph Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.23254","noteUrl":"/review?paper=review-neurips-2026-148744","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148746","title":"Robust Instruction Compliance in Cooperative Multi-Agent Reinforcement Learning","authors":["Wo Wei Lin","Ethan Rathbun","Enrico Marchesini","Xiang Zhi Tan"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Value correction does not undo a new macro-action decision boundary","plainExplanation":"Interrupting a macro-action changes when the policy may choose its next action. Cancelling the incoming instruction value does not by itself recover the value of the uninterrupted process.","primaryUrl":"https://arxiv.org/html/2605.12655","noteUrl":"/review?paper=review-neurips-2026-148746","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","formula_error","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148751","title":"DRTriton: Large-Scale Synthetic Data Driven Reinforcement Learning for Triton Kernel Generation","authors":["Siqi Guo","Ming Lin","Tianbao Yang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The sampler does not cover valid programs with repeated operands","plainExplanation":"Selecting operator inputs without replacement excludes valid programs that use the same tensor twice. Shape solving cannot recover graph structures that the sampler never generates.","primaryUrl":"https://arxiv.org/html/2603.21465","noteUrl":"/review?paper=review-neurips-2026-148751","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","formula_error"],"classifications":["major"]},{"slug":"review-neurips-2026-148755","title":"Leviathan: Decoupling Input and Output Representations in Language Models","authors":["Reza T Batley","Sourav Saha"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Leviathan: Decoupling Input and Output Representations in Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.22040","noteUrl":"/review?paper=review-neurips-2026-148755","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148756","title":"Understanding and Mitigating Under-Confidence in GNNs from the Final Layer","authors":["Jincheng Huang","Jie Xu","Ping Hu","Xiaoshuang Shi","Lei Feng","Xiaofeng Zhu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Two confidence interpretations need additional conditions","plainExplanation":"The loss-gradient offset is not scaled with the old logits. Consequently the update is not ordinary temperature scaling of a fixed logit vector, and increasing the decay can increase the separation between the updated logits.","primaryUrl":"https://arxiv.org/html/2505.11335","noteUrl":"/review?paper=review-neurips-2026-148756","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148758","title":"Consistent 3D Surface Flow Model with Global State","authors":["Antoine Guedon","Shu Nakamura","Nicolas Dufour","Jiahui Lei","Ko Nishino","Angjoo Kanazawa"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Consistent 3D Surface Flow Model with Global State","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.13644","noteUrl":"/review?paper=review-neurips-2026-148758","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148761","title":"When the Same Coefficients Reach Different Places: Asymmetric Realizability in Transplanting Tokenizers across Large Language Models","authors":["Xiaoze Liu","Weichen Yu","Matt Fredrikson","Xiaoqian Wang","Jing Gao"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"When the Same Coefficients Reach Different Places: Asymmetric Realizability in Transplanting Tokenizers across Large Language Models","plainExplanation":"Appending a malicious vocabulary row can reduce the probability of emitting any breaker under nucleus sampling, even when old logits are unchanged. The bundling monotonicity argument needs a specified decoder and conditions that control its retained set.","primaryUrl":"https://arxiv.org/html/2601.00065","noteUrl":"/review?paper=review-neurips-2026-148761","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148762","title":"When Dynamics Shift, Robust Task Inference Wins: Offline Imitation Learning with Behavior Foundation Models Revisited","authors":["Rishabh Agrawal","Rahul Jain","Ashutosh Nayyar"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The robust dual needs support and finite-weight conditions","plainExplanation":"A total-variation perturbation can put mass outside nominal support, where a nominal density ratio cannot represent it. SoftTV also permits fixed-critic dual evaluations with no finite maximizing weight.","primaryUrl":"https://arxiv.org/html/2605.17017","noteUrl":"/review?paper=review-neurips-2026-148762","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","formula_error","invalid_reduction","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-148764","title":"IFDECORATOR: Wrapping Instruction Following Reinforcement Learning with Verifiable Rewards","authors":["Xu Guo","Tianyi Liang","Jian Tong","Xiaogui Yang","Ling-I Wu","Chenhui Li","Zhihui Lu","Qipeng Guo","Kai Chen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"IFDECORATOR: Wrapping Instruction Following Reinforcement Learning with Verifiable Rewards","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2508.04632","noteUrl":"/review?paper=review-neurips-2026-148764","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148765","title":"Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic","authors":["Yewon Jeong","Nayoung Jung","Hyeri Roh","Woo-Seok Choi"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11827","noteUrl":"/review?paper=review-neurips-2026-148765","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148766","title":"Motion-o: Trajectory-Grounded Video Reasoning","authors":["Bishoy Galoaa","Shayda Moezzi","Xiangyu Bai","Sarah Ostadabbas"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Matched prompts do not remove stochastic decoding disagreement","plainExplanation":"The paired-video analysis attributes any change in sampled motion attributes to temporal evidence because the prompts are identical. An input-independent policy can independently sample among nine directions, four speeds and three scales.","primaryUrl":"https://arxiv.org/html/2603.18856","noteUrl":"/review?paper=review-neurips-2026-148766","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148770","title":"Phase Transitions in Attention: A Bayesian Theory of Copy Head Emergence","authors":["Itay Lavie","Kirsten Fischer","Andrey Lekov","Frederic Van Maele","Zohar Ringel","Moritz Helias"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The Laplace approximation has the wrong action sign","plainExplanation":"The Laplace approximation has the wrong action sign. The projected prior couples the uniform and copy-head coordinates.","primaryUrl":"https://arxiv.org/html/2606.12058","noteUrl":"/review?paper=review-neurips-2026-148770","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148778","title":"Surprises in Proper Positive-Only Learning","authors":["Shai Ben-David","Farnam Mansouri","Anay Mehrotra","Manolis Zampetakis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The exponential-tail threshold needs the reciprocal logarithm","plainExplanation":"The displayed logarithmic sample-size condition has the reciprocal reversed. It is vacuous for positive sample sizes and cannot imply the asserted exponential-tail bound on its own. This is a local sufficient-condition typo. The earlier sample-size bound can absorb the corrected logarithm; learnability and its stated asymptotic rate are not refuted.","primaryUrl":"https://arxiv.org/html/2606.28309","noteUrl":"/review?paper=review-neurips-2026-148778","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148780","title":"FlowR2A: Learning Reward-to-Action Distribution for Multimodal Driving Planning","authors":["Xirui Li","Zhe Liu","Xiaoqing Ye","Wenhua Han","Yifeng Pan","Junyu han","Hengshuang Zhao"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"FlowR2A: Learning Reward-to-Action Distribution for Multimodal Driving Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.24231","noteUrl":"/review?paper=review-neurips-2026-148780","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148781","title":"Regret–Oracle Complexity Tradeoffs in Agnostic Online Learning","authors":["Idan Attias","Steve Hanneke","Arvind Ramaswami"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Regret–Oracle Complexity Tradeoffs in Agnostic Online Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07155","noteUrl":"/review?paper=review-neurips-2026-148781","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148791","title":"Log-Averaged Mirror Prox for Fast, Large-Scale Optimal Transport in Linear Space","authors":["Matthew Burns","Jiaming Liang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The online log-sum-exp update raises the old sum to a power","plainExplanation":"When a new maximum \\(m_{\\mathrm{new}}\\) arrives, old terms must be rescaled by \\(e^{m_{\\mathrm{old}}-m_{\\mathrm{new}}}\\), giving \\(s_{\\mathrm{new}}=1+s_{\\mathrm{old}}e^{m_{\\mathrm{old}}-m_{\\mathrm{new}}}\\). The v4 typeset page 27 instead raises the old sum to the exponent. After values \\(0,1\\), it returns \\(2\\), whereas the invariant requires \\(1+e^{-1}\\). The otherwise branch correctly adds an exponential and confirms the intended invariant. This is a typeset recurrence error verified on the exact v4 PDF, not evidence of an implementation failure.","primaryUrl":"https://arxiv.org/pdf/2511.11359v4","noteUrl":"/review?paper=review-neurips-2026-148791","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148795","title":"Alleviating Hallucination with Training-Free Uncertainty-Guided Steering","authors":["Litian Liu","Yubing Jian","Qiqi Hou","Reza Pourreza","Mohammad Ghavamzadeh","Yao Qin","Roland Memisevic","Herbert Cai"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The final steering direction needs a zero-gradient convention","plainExplanation":"The RMSNorm derivative equality is inexact although its normalized direction agrees. The final steering direction needs a zero-gradient convention.","primaryUrl":"https://arxiv.org/html/2609.38962","noteUrl":"/review?paper=review-neurips-2026-148795","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148796","title":"StarWM:  Self-Supervised Trained Attention Routing for Robust World Models","authors":["Zeqiang Zhang","Fabian Wurzberger","Maximilian Otte","Daniel Schmid","Sebastian Gottwald","Arne P Raulf","Daniel A Braun"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"StarWM:  Self-Supervised Trained Attention Routing for Robust World Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30667","noteUrl":"/review?paper=review-neurips-2026-148796","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148800","title":"Sparse Attention as Compact Kernel Regression","authors":["Saul Santos","Nuno Gonçalves","Daniel McNamee","Marcos Treviso","André Martins"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Sparse Attention as Compact Kernel Regression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.22766","noteUrl":"/review?paper=review-neurips-2026-148800","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148807","title":"Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior","authors":["Tuc Nguyen","Thai Le"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.08454","noteUrl":"/review?paper=review-neurips-2026-148807","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148808","title":"SimpliHuMoN: Simplifying Human Motion Prediction","authors":["Aadya Agrawal","Alex Schwing"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SimpliHuMoN: Simplifying Human Motion Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.04399","noteUrl":"/review?paper=review-neurips-2026-148808","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148810","title":"Is Text All You Need? Text as a Universal Information Bottleneck for Speech LLMs","authors":["Ming-Hao Hsu","Yuxuan Hu","Shujie LIU","Jinyu Li","Zhizheng Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Is Text All You Need? Text as a Universal Information Bottleneck for Speech LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.09366","noteUrl":"/review?paper=review-neurips-2026-148810","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148811","title":"UniT:  Toward a Unified Physical Language for Human-to-Humanoid Policy Learning and World Modeling","authors":["Boyu Chen","Yi Chen","Lu QIU","Jerry Bai","Yuying Ge","Yixiao Ge"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"UniT:  Toward a Unified Physical Language for Human-to-Humanoid Policy Learning and World Modeling","plainExplanation":"Predicting another modality does not require discarding information useful for reconstructing one’s own. An allowed reconstruction optimum retains private noise.","primaryUrl":"https://arxiv.org/html/2604.19734","noteUrl":"/review?paper=review-neurips-2026-148811","collection":"neurips-2026","severity":"major","issueTypes":["invalid_inference","unsupported_claim"],"classifications":["major"]},{"slug":"review-neurips-2026-148812","title":"gHAWK: Structural Encoding for Scalable Training of Graph Neural Networks on Knowledge Graphs","authors":["Humera Sabir","Fatima Farooq","Ashraf Aboulnaga"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"gHAWK: Structural Encoding for Scalable Training of Graph Neural Networks on Knowledge Graphs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.08274","noteUrl":"/review?paper=review-neurips-2026-148812","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148813","title":"Private Prediction via Shrinkage","authors":["Chao Yan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Private Prediction via Shrinkage","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.05219","noteUrl":"/review?paper=review-neurips-2026-148813","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148814","title":"Verifier-Backed Hard Problem Generation for Mathematical Reasoning","authors":["Yuhang Lai","Jiazhan Feng","Yee Whye Teh","Ning Miao"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Verifier-Backed Hard Problem Generation for Mathematical Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06660","noteUrl":"/review?paper=review-neurips-2026-148814","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148816","title":"Controllable User Simulation","authors":["Guy Tennenholtz","Ofer Meshi","Amir Globerson","Uri Shalit","Jihwan Jeong","Craig Boutilier"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A policy shift need not produce belief-ratio variance","plainExplanation":"A policy shift need not produce belief-ratio variance.","primaryUrl":"https://arxiv.org/html/2605.11519","noteUrl":"/review?paper=review-neurips-2026-148816","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148817","title":"Analyzing and Guiding Zero-Shot Posterior Sampling in Diffusion Models","authors":["Roi Benita","Miki Elad","Joseph Keshet"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Gaussian transport uses the absolute transfer as its standard deviation","plainExplanation":"The analytic Wasserstein objective uses a signed linear sampling transfer \\(D_1\\) directly as a Gaussian standard deviation. The generated covariance is \\(D_1^2\\), so the scalar covariance contribution is \\((\\sqrt\\lambda-|D_1|)^2\\).","primaryUrl":"https://arxiv.org/html/2602.07715","noteUrl":"/review?paper=review-neurips-2026-148817","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148820","title":"In-Context Learning Can Help Vision Language Models Overcome Training Prior","authors":["Kun Wang","Xindi Wu","Sanghyuk Chun","Olga Russakovsky","Esin Tureci"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"In-Context Learning Can Help Vision Language Models Overcome Training Prior","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.12942","noteUrl":"/review?paper=review-neurips-2026-148820","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148821","title":"AdaCodec: A Predictive Visual Code for Video MLLMs","authors":["Haowen Hou","Zhen Huang","Zheming Liang","Qingyi Si","Chenglin Li","Shuai Dong","Kele Shao","Ruilin Li","Dianyi Wang","Nan Duan","Jiaqi Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AdaCodec: A Predictive Visual Code for Video MLLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02569","noteUrl":"/review?paper=review-neurips-2026-148821","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148822","title":"Bounding Global and Local Compression Error of Signal Parameterizations","authors":["Quang Luong Nhat Nguyen","Sara Fridovich-Keil"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Bounding Global and Local Compression Error of Signal Parameterizations","plainExplanation":"The model-specific compression bounds need stronger hypotheses: radial decay does not ensure diminishing shell gains, collision upper bounds do not yield positive lower envelopes, and hash distinguishability alone does not control a fixed decoder’s error. The FFN minimax rate also does not lower-bound every signal, while the finite-endpoint envelope argument needs either a vanishing majorant or the proved endpoint remainder.","primaryUrl":"https://arxiv.org/html/2606.00126","noteUrl":"/review?paper=review-neurips-2026-148822","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","lemma_error","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148825","title":"TS-ICL: A Flexible Time-Indexed Foundation Model for Time Series via In-Context Learning","authors":["Etienne Le Naour","Tahar Nabil","Adrien Petralia"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"TS-ICL: A Flexible Time-Indexed Foundation Model for Time Series via In-Context Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.05878","noteUrl":"/review?paper=review-neurips-2026-148825","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148827","title":"Phantom Transfer: Data Poisoning can Survive Data-Level Defences","authors":["Andrew Draganov","Tolga Hasan Dur","Anandmayi (Andi) Bhongade","Mary Phuong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Phantom Transfer: Data Poisoning can Survive Data-Level Defences","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2602.04899","noteUrl":"/review?paper=review-neurips-2026-148827","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148828","title":"Replay-buffer engineering for noise-robust quantum circuit optimization","authors":["Akash Kundu","Sebastian Feld"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Replay-buffer engineering for noise-robust quantum circuit optimization","plainExplanation":"Reliability weighting need not reduce the fully corrected update variance. The theoretical bias, contraction and noise-floor bounds also need correction; the implementation already uses consecutive cached costs for intermediate rewards.","primaryUrl":"https://arxiv.org/html/2604.21863","noteUrl":"/review?paper=review-neurips-2026-148828","collection":"neurips-2026","severity":"major","issueTypes":["documentation_error","false_statement","formula_error","incorrect_bound","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148829","title":"MonarchRT: Efficient Attention for Real-Time Video Generation","authors":["Krish Agarwal","Zhuoming Chen","cheng Luo","Yongqi Chen","Haizhong Zheng","Xun Huang","Atri Rudra","Beidi Chen"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Nontrivial tiling need not strictly enlarge an already unrestricted matrix class","plainExplanation":"Nontrivial tiling need not strictly enlarge an already unrestricted matrix class.","primaryUrl":"https://arxiv.org/html/2602.12271","noteUrl":"/review?paper=review-neurips-2026-148829","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148831","title":"MMSkills: Towards Multimodal Skills for General Visual Agents","authors":["Kangning Zhang","Shuai Shao","Wenxiang Jiao","Qingyao Li","Jianghao Lin","Lingyue Fu","Shijian Wang","Yuan Lu","Weiwen Liu","Weinan Zhang","Yong Yu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MMSkills: Towards Multimodal Skills for General Visual Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13527","noteUrl":"/review?paper=review-neurips-2026-148831","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148833","title":"Flow Mismatching: Unsupervised Anomaly Detection via Velocity Discrepancies in Flow Matching Models","authors":["Shengzhe Chen","Mehrdad Moradi","Kamran Paynabar","Hao Yan"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Flow Mismatching: Unsupervised Anomaly Detection via Velocity Discrepancies in Flow Matching Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23070","noteUrl":"/review?paper=review-neurips-2026-148833","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148834","title":"Spherical Boltzmann machines: a solvable theory of learning and generation in energy-based models","authors":["Thomas Tulinski","Simona Cocco","Remi Monasson","Jorge FERNANDEZ-DE-COSSIO-DIAZ"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Spherical Boltzmann machines: a solvable theory of learning and generation in energy-based models","plainExplanation":"The Gaussian comparison assigns spectra outside the normalizable domain of its stated density and needs a corrected model and new comparison calculation. A separate sign sentence in the temperature-tuning appendix is reversed; the main-text interpretation and later phase formulas already use the correct direction.","primaryUrl":"https://arxiv.org/html/2605.09031","noteUrl":"/review?paper=review-neurips-2026-148834","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-148835","title":"Normative Networks for Source Separation via Local Plasticity and Dendritic Computation","authors":["Bariscan Bozkurt","Efe Ali Gorguner","Francesco Innocenti","Rafal Bogacz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Normative Networks for Source Separation via Local Plasticity and Dendritic Computation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19965","noteUrl":"/review?paper=review-neurips-2026-148835","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148836","title":"ManipShield: A Unified Framework for Image Manipulation Detection, Localization and Explanation","authors":["Zitong Xu","Huiyu Duan","Xiaoyu Wang","Zhaolin Cai","Kaiwei Zhang","Xiongkuo Min","Guangtao Zhai"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Weight decay is scaled by the learning rate","plainExplanation":"The weight-decay explanation mistakes the coefficient for a per-update fractional shrinkage. In the ordinary decoupled update, shrinkage is scaled by the learning rate, so coefficient \\(0.1\\) does not mean ten percent per update. The error concerns the stated mathematical effect of a hyperparameter. It does not identify which optimizer update was executed or challenge measured performance.","primaryUrl":"https://arxiv.org/html/2511.14259","noteUrl":"/review?paper=review-neurips-2026-148836","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148838","title":"When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search","authors":["John Robertson","Jianing Zhu","Haris Vikalo","Zhangyang &amp;quot;Atlas&amp;quot; Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16362","noteUrl":"/review?paper=review-neurips-2026-148838","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148839","title":"A Provably Convergent and Practical Algorithm for Gromov–Wasserstein Optimal Transport","authors":["Ling Liang","Lei Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Provably Convergent and Practical Algorithm for Gromov–Wasserstein Optimal Transport","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.04175","noteUrl":"/review?paper=review-neurips-2026-148839","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148840","title":"Safe Score Matching: Diffusion Policies with Hamilton-Jacobi Reachability for Online Safe Reinforcement Learning","authors":["Boyang Li","Matthew Kim","Sylvia Herbert"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Safe Score Matching: Diffusion Policies with Hamilton-Jacobi Reachability for Online Safe Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33337","noteUrl":"/review?paper=review-neurips-2026-148840","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148841","title":"Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap","authors":["Tianlang Chen","Shirley Wu","Jure Leskovec"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24432","noteUrl":"/review?paper=review-neurips-2026-148841","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148846","title":"Scalable Token-Level Hallucination Detection in Large Language Models","authors":["Rui Min","Tianyu Pang","Chao Du","Minhao Cheng","Yi R. (May) Fung"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Scalable Token-Level Hallucination Detection in Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12384","noteUrl":"/review?paper=review-neurips-2026-148846","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148848","title":"TGPO: Temporal Grounded Policy Optimization for Signal Temporal Logic Tasks","authors":["Yue Meng","Fei Chen","Chuchu Fan"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The Until decomposition omits the initial left-predicate interval","plainExplanation":"The Until decomposition omits the initial left-predicate interval. Endpoint flags do not certify an uninterrupted stay interval. Body angular velocity is not the Euler-angle derivative.","primaryUrl":"https://arxiv.org/html/2510.00225","noteUrl":"/review?paper=review-neurips-2026-148848","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148849","title":"Parallel-in-Time Training of Recurrent Neural Networks for Dynamical Systems Reconstruction","authors":["Florian Hess","Florian Götz","Daniel Durstewitz"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The readout regularizer includes one extra coordinate","plainExplanation":"The displayed regularizer includes \\(M_r+1\\) coordinates while claiming to select the final \\(M_r\\). At the full-width boundary it includes the nonexistent zero index. This is an index typo in the displayed regularizer. The nearby correct range establishes the intended convention; implementation behavior is not determined here.","primaryUrl":"https://arxiv.org/html/2605.12683","noteUrl":"/review?paper=review-neurips-2026-148849","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148850","title":"SkillGen: Verified Inference-Time Agent Skill Synthesis","authors":["Yuchen Ma","Yue Huang","Han Bao","Haomin Zhuang","Swadheen Shukla","Michel Galley","Xiangliang Zhang","Stefan Feuerriegel"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"SkillGen: Verified Inference-Time Agent Skill Synthesis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10999","noteUrl":"/review?paper=review-neurips-2026-148850","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148851","title":"Align $\\&$ Invert: Solving Inverse Problems with Diffusion and Flow-based Models via Representation Alignment","authors":["Loukas Sfountouris","Giannis Daras","Paris Giampouras"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The cosine-to-Euclidean bound needs normalized projected features too","plainExplanation":"Equation (3) converts a probability-flow ODE velocity \\(v\\) to a reverse-clock SDE with diffusion \\(g\\), adding a full \\(g^2\\) score correction. For the same marginals, reverse-clock drift is \\(-v+(g^2/2)\\nabla\\log p\\).","primaryUrl":"https://arxiv.org/html/2511.16870","noteUrl":"/review?paper=review-neurips-2026-148851","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148853","title":"Time-Sensitive Anytime-Valid Testing","authors":["Eugenio Clerico","Tobias Wegel","Iskander Azangulov","Patrick Rebeschini"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The displayed optimal bet loses a normalization factor","plainExplanation":"Path atoms have unequal null probabilities when $p_0\\ne1/2$. Ranking by likelihood ratio solves the fractional problem, but deterministic events cannot split atoms. A lower-ranked path can be better than several higher-ranked paths combined. Failure to fit it into the leftover budget does not prove the greedy selection optimal. The greedy shortcut is false for the stated deterministic problem. Solving the existing integer programme repairs the terminal-event selection without changing the level constraint, objective or subsequent martingale construction. The general optimal-control formulation is not refuted.","primaryUrl":"https://arxiv.org/html/2605.06521","noteUrl":"/review?paper=review-neurips-2026-148853","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","definition_error","false_statement","formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148856","title":"Normalizing Trajectory Models","authors":["Jiatao Gu","Tianrong Chen","Ying Shen","David Berthelot","Shuangfei Zhai","Joshua Susskind"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Normalizing Trajectory Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08078","noteUrl":"/review?paper=review-neurips-2026-148856","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148857","title":"DC-DiT: Adaptive Compute and Elastic Inference for Visual Generation via Dynamic Chunking","authors":["Akash Haridas","Utkarsh Saxena","Parsa Ashrafi Fashi","Mehdi Rezagholizadeh","Vikram Appia","Emad Barsoum"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DC-DiT: Adaptive Compute and Elastic Inference for Visual Generation via Dynamic Chunking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.06351","noteUrl":"/review?paper=review-neurips-2026-148857","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148859","title":"Duality Models: An Embarrassingly Simple One-step Generation Paradigm","authors":["PENG SUN","Xinyi Shang","Zhenglin Cheng","deyuan liu","Tao Lin","Zhiqiang Shen"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A flow-matching summand does not prevent opposing gradients","plainExplanation":"A flow-matching summand does not prevent opposing gradients. The earlier-state consistency term uses the wrong time argument.","primaryUrl":"https://arxiv.org/html/2602.17682","noteUrl":"/review?paper=review-neurips-2026-148859","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148860","title":"Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry","authors":["Pablo Marcos Manchón","Rishi Jha","Lluís Fuentemilla"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.20496","noteUrl":"/review?paper=review-neurips-2026-148860","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148861","title":"DiPO: Disentangled Perplexity Policy Optimization for Fine-grained Exploration-Exploitation Trade-Off","authors":["Xiaofan Li","Ming Yang","Zhiyuan Ma","Shichao Ma","Jintao Du","Yu Cheng","Weiqiang Wang","zhizhong zhang","Xin Tan","Yanyun Qu","Lizhuang Ma","Yuan Xie"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"DiPO: Disentangled Perplexity Policy Optimization for Fine-grained Exploration-Exploitation Trade-Off","plainExplanation":"Maximum perplexity within a sampled group does not guarantee the stated policy-entropy direction. Algorithm 2 also retains nonzero auxiliary rewards for mixed groups, although Equation (10) already defines those rewards as zero; an explicit mixed-group reset aligns the two maps.","primaryUrl":"https://arxiv.org/html/2604.13902","noteUrl":"/review?paper=review-neurips-2026-148861","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148863","title":"MAST: Label-Efficient, Robust, and Generalizable Sound Detection for Biodiversity Monitoring via Masked Audio Pretraining and Self-Training","authors":["Tianyi Xu","Daniel Pimentel-Alarcón","Zuzana Buřivalová","Claudia Solis-Lemus"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MAST: Label-Efficient, Robust, and Generalizable Sound Detection for Biodiversity Monitoring via Masked Audio Pretraining and Self-Training","plainExplanation":"Detection precision and coverage do not control total variation between complete annotation laws. Separately, the convolution appendix’s rank inequality is correct, but sequential axis filters can still produce full-rank nonseparable kernels for one channel pair; the meaningful restriction is their joint channel/offset factorization.","primaryUrl":"https://arxiv.org/html/2609.15221","noteUrl":"/review?paper=review-neurips-2026-148863","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","incorrect_bound","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-148865","title":"LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss","authors":["Euntae Choi","Sumin Song","Sungjoo Yoo"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08755","noteUrl":"/review?paper=review-neurips-2026-148865","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148867","title":"Causal Representation Learning for Generalisable Recommendation","authors":["Yorgos Felekis","Michael O&amp;#x27;Riordan","Oriol Corcoll Andreu","Ciarán Gilligan-Lee"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Causal Representation Learning for Generalisable Recommendation","plainExplanation":"No issue was found in the checked claims. The arbitrary-critic guarantee in v1 fails because a constant critic gives zero penalty for dependent variables. Version 2, uploaded 23 October 2026, requires an optimal InfoNCE critic recovering the log-density ratio, removes the associated tightness sentence, and distinguishes a learned approximation from that guarantee. The constant-critic example does not satisfy the revised premise.","primaryUrl":"https://arxiv.org/html/2605.27043","noteUrl":"/review?paper=review-neurips-2026-148867","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148870","title":"Rethinking Ratio-Based Trust Regions for Policy Optimization in Multi-Agent Reinforcement Learning","authors":["Chulabhaya Wijesundara","Andrea Baisero","Zhongheng Li","Gregory D Castanon","Alan S Carlin","Christopher Amato"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Rethinking Ratio-Based Trust Regions for Policy Optimization in Multi-Agent Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09212","noteUrl":"/review?paper=review-neurips-2026-148870","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148871","title":"A Structured LLM Framework for Inorganic Material Synthesis Planning","authors":["Heewoong Noh","Gyoung S. Na","Namkyeong Lee","Chanyoung Park"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Auxiliary precursor prediction does not force operation conditioning","plainExplanation":"The argument says predicting precursor output prevents precursor-agnostic operation prediction, although the input already contains the precursor information. Both objectives observe \\((X,Z_{\\rm in})\\), with output precursor \\(Z\\) deterministic from that input.","primaryUrl":"https://arxiv.org/html/2602.07543","noteUrl":"/review?paper=review-neurips-2026-148871","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148872","title":"Risk-Averse Online POMDP Planning via CVaR of the Immediate Cost with Performance Guarantees","authors":["Yaacov Pariente","Vadim Indelman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Risk-Averse Online POMDP Planning via CVaR of the Immediate Cost with Performance Guarantees","plainExplanation":"Three mathematical issues concern policy disagreement, different trajectory probabilities, and the equality condition for the marginal CVaR certificate.","primaryUrl":"https://arxiv.org/html/2609.35874","noteUrl":"/review?paper=review-neurips-2026-148872","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-148873","title":"LR-V2X: Loss Resilient Collaborative Perception under Low-Bandwidth Communication","authors":["Kang Yang","Tianci Bu","Peng Wang","Deying Li","Yongcai Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LR-V2X: Loss Resilient Collaborative Perception under Low-Bandwidth Communication","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.11264","noteUrl":"/review?paper=review-neurips-2026-148873","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148874","title":"ProDG: Prototypes for Data-Free Generative Post-Hoc Explainability","authors":["Piotr Borycki","Magdalena Trędowicz","Jacek Tabor","Łukasz Struski","Przemysław Spurek"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"ProDG: Prototypes for Data-Free Generative Post-Hoc Explainability","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08858","noteUrl":"/review?paper=review-neurips-2026-148874","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148876","title":"Learning with Synthetic Data via SGD in High-Dimensional Linear Regression","authors":["Jichu Li","Difan Zou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning with Synthetic Data via SGD in High-Dimensional Linear Regression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.09572","noteUrl":"/review?paper=review-neurips-2026-148876","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148878","title":"Spherical Flows for Sampling Categorical Data","authors":["Jannis Chemseddine","Gregor Kornhardt","Gabriele Steidl"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Spherical Flows for Sampling Categorical Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05629","noteUrl":"/review?paper=review-neurips-2026-148878","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148879","title":"Literati: Towards Anytime Optimal Shape Generalized Trees via AO*","authors":["Nakul Upadhya","Eldan Cohen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Pruning must retain actions whose bound equals the incumbent","plainExplanation":"Algorithm 5’s strict retention test prunes equality cases, including a tight incumbent. This contradicts the main-text rule that only actions strictly worse than the incumbent bound are safely pruned. This is a boundary-sign typo in the pseudocode. The main-text pruning principle and optimality argument use the intended comparison.","primaryUrl":"https://arxiv.org/html/2609.09299","noteUrl":"/review?paper=review-neurips-2026-148879","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148880","title":"A Unified Perturbation Framework for Analyzing Leaderboard Stability and Manipulation","authors":["Hosna Oyarhoseini","Jimmy Lin","Amir-Hossein Karimi"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Duplicating each vote halves the rowwise covariance estimate","plainExplanation":"Duplicating each vote halves the rowwise covariance estimate.","primaryUrl":"https://arxiv.org/html/2605.15761","noteUrl":"/review?paper=review-neurips-2026-148880","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148881","title":"Hypergraph Generation via Structured Stochastic Diffusion","authors":["Christopher Nemeth"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The reverse OU term has the expansive sign","plainExplanation":"Continuity of a time-dependent drift does not control how quickly a left-endpoint quadrature converges. Even a bounded drift independent of the state can violate both stated rates while satisfying the printed regularity assumptions. This is a missing assumption in the functional statement of the numerical rate. The example is a continuous drift, not a claim that a particular finite trained network realizes the infinite series. It does not show that the reported sampler fails, or refute the separate ideal reverse-SDE result.","primaryUrl":"https://arxiv.org/html/2605.05024","noteUrl":"/review?paper=review-neurips-2026-148881","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148882","title":"AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace","authors":["Seonglae Cho","Donghyun Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace","plainExplanation":"Pairwise-independent file choices do not imply the birthday collision formula for three or more agents. The product formula is valid for mutually independent uniform choices; the two-agent uniform illustration remains valid.","primaryUrl":"https://arxiv.org/html/2608.23740","noteUrl":"/review?paper=review-neurips-2026-148882","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148883","title":"Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models","authors":["Bryce Grant","Xijia Zhao","Peng Wang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.19233","noteUrl":"/review?paper=review-neurips-2026-148883","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148884","title":"Approximate Envy-Free Allocations  up to any k Goods","authors":["Aris Filos-Ratsikas","Georgios Kalantzis","Fangxiao Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Weak envy cycles can persist after cycle resolution","plainExplanation":"Weak envy cycles can persist after cycle resolution. The positive perturbation parameter may be required to lie below zero.","primaryUrl":"https://arxiv.org/html/2605.10371","noteUrl":"/review?paper=review-neurips-2026-148884","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148886","title":"Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models","authors":["Xin Xu","Clive Bai","Kai Yang","Tianhao Chen","Yuchen Cai","Yangkun Chen","Weijie Liu","Hao CHEN","Yang Wang","Saiyong Yang","Can Yang"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.12036","noteUrl":"/review?paper=review-neurips-2026-148886","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148887","title":"ORCAID: Oblique Rule-Based Continuous-Action Interpretation for Deep RL Policies","authors":["Ignacio D. Lopez-Miguel","Ezio Bartocci","Thomas Eiter","Martin Tappler"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ORCAID: Oblique Rule-Based Continuous-Action Interpretation for Deep RL Policies","plainExplanation":"Binary branches can be complementary on their feature domain without having opposite affine thresholds. The universal sibling-adjacency claim therefore needs a separate binary-split rule or a restriction to the stated continuous geometric setting.","primaryUrl":"https://arxiv.org/html/2607.07235","noteUrl":"/review?paper=review-neurips-2026-148887","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148889","title":"Learning the Signature of Memorization in Autoregressive Language Models","authors":["David Ilić","Kostadin Cvejoski","David Stanojević","Evgeny Grigorenko"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Learning the Signature of Memorization in Autoregressive Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.03199","noteUrl":"/review?paper=review-neurips-2026-148889","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148891","title":"From SGD to Muon: Adaptive Optimization via Schatten-p Norms","authors":["Thomas Massena","Corentin Friedrich","Mathieu Serrurier"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The surrogate geometry search needs a positive-part constraint","plainExplanation":"Squaring a negative alignment rewards an ascent direction even though its best admissible nonnegative step is zero. Momentum need not align with the current gradient, so this case is allowed.","primaryUrl":"https://arxiv.org/abs/2605.19781v2","noteUrl":"/review?paper=review-neurips-2026-148891","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148893","title":"PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models","authors":["Xinyu Guo","Bin Xie","Wei Chai","Xianchi Deng","Tiancai Wang","Zhengxing Wu","Xingyu Chen"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10925","noteUrl":"/review?paper=review-neurips-2026-148893","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148896","title":"AgentOdyssey: Open-Ended Long-Horizon Text Game Generation for Test-Time Continual Learning Agents","authors":["Zheyuan Zhang","Zehao Wen","Bowei Zhang","Andrew Wang","Jianwen Xie","Daniel Khashabi","Tianmin Shu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"AgentOdyssey: Open-Ended Long-Horizon Text Game Generation for Test-Time Continual Learning Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.24893","noteUrl":"/review?paper=review-neurips-2026-148896","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148897","title":"Adapting in the Dark: Efficient and Stable Test-Time Adaptation for Black-Box Models","authors":["Yunbei Zhang","Shuaicheng Niu","Chengyi Cai","Feng Liu","Jihun Hamm"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Adapting in the Dark: Efficient and Stable Test-Time Adaptation for Black-Box Models","plainExplanation":"At an interior mixture weight, zero disagreement requires the steering and target distributions to agree. A nonuniform detached target can leave both positive divergence and a nonzero steering gradient at uniform output; the underlying entropy identity remains valid.","primaryUrl":"https://arxiv.org/html/2604.15609","noteUrl":"/review?paper=review-neurips-2026-148897","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148898","title":"Understanding the Challenges in Iterative Generative Optimization with LLMs","authors":["Allen Nie","Xavier Daull","Zhiyi Kuang","Abhinav Akkiraju","Anish Chaudhuri","Max Piasevoli","Ryan Rong","YuCheng Yuan","Prerit Choudhary","Shannon Xiao","Rasool Fakoor","Adith Swaminathan","Ching-An Cheng"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Understanding the Challenges in Iterative Generative Optimization with LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.23994","noteUrl":"/review?paper=review-neurips-2026-148898","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148899","title":"SierpinskiCam: Camera-Controlled Video Retaking with Sierpinski Triangle Pattern Cues","authors":["Suttisak Wizadwongsa","Hyelin Nam","Supasorn Suwajanakorn","Jeong Joon Park"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SierpinskiCam: Camera-Controlled Video Retaking with Sierpinski Triangle Pattern Cues","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.17310","noteUrl":"/review?paper=review-neurips-2026-148899","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148900","title":"A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning","authors":["Haibo Chen","Xin Wang","Jiaheng Chao","Ling Feng","Wenwu Zhu"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12197","noteUrl":"/review?paper=review-neurips-2026-148900","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148902","title":"Triggering Generalist Reasoning via Predictive Uncertainty for Dual-System VLA","authors":["Hyemin Yang","Wooseong Jeong","Giwon Lee","Kuk-Jin Yoon"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The buffer boundary needs the drift term one step earlier","plainExplanation":"At elapsed time exactly \\(E\\), the anchor timestamp is already outside an \\(E\\)-entry consecutive-time buffer. Recent agreement can then coexist with a different cached anchor.","primaryUrl":"https://arxiv.org/html/2610.05025","noteUrl":"/review?paper=review-neurips-2026-148902","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148903","title":"Test-Time Conditioning with Representation-Aligned Visual Features","authors":["Nicolas Sereyjol-Garros","Ellington Kirby","Victor Letzelter","Victor Besnier","Nermin Samet"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Test-Time Conditioning with Representation-Aligned Visual Features","plainExplanation":"The unit-feature optimum follows the normalized conditional mean, rather than the unconstrained conditional mean. The high-temperature SPA scalar limit also needs subtraction of its additive temperature-dependent constant; this normalization leaves the correctly derived guidance gradient unchanged.","primaryUrl":"https://arxiv.org/html/2602.03753","noteUrl":"/review?paper=review-neurips-2026-148903","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","missing_assumption","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-148905","title":"Existence Precedes Value: Joint Modeling of Observational Existence and Evolving States in Time Series Forecasting","authors":["Yifan Hu","Hongzhou Chen","Peiyuan Liu","Yiding Liu","Zewei Dong","Jiang-Ming Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Unrestricted reliability outputs do not support real geometric-mean weights","plainExplanation":"Reliability is an unrestricted affine output \\(\\rho\\), and attention uses the entrywise square root of \\(\\rho\\rho^\\top\\). Two permitted patch outputs \\((-1,1)\\) give negative off-diagonal outer-product entries, so their entrywise square roots are not real.","primaryUrl":"https://arxiv.org/html/2606.13571","noteUrl":"/review?paper=review-neurips-2026-148905","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148906","title":"Capturing LLM Capabilities via Evidence-Calibrated Query Clustering","authors":["Fangzhou Wu","Sandeep Silwal","Qiuyi (Richard) Zhang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Convexity alone does not ensure an attained or descending profile update","plainExplanation":"The monotonicity proof treats unconstrained convex Bradley–Terry profile optimization as automatically yielding a finite global minimum from finite initialization. One one-sided comparison has gap loss \\(\\log(1+e^{-d})\\), with unattained infimum zero as \\(d\\to\\infty\\).","primaryUrl":"https://arxiv.org/html/2605.17110","noteUrl":"/review?paper=review-neurips-2026-148906","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148909","title":"Simple Projection-Free Algorithm for Contextual Recommendation with Logarithmic Regret and Robustness","authors":["Shinsaku Sakaue"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Simple Projection-Free Algorithm for Contextual Recommendation with Logarithmic Regret and Robustness","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.20826","noteUrl":"/review?paper=review-neurips-2026-148909","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148910","title":"RotMoLE: Enhancing Mixture of Low-Rank Experts through Rotational Gating Mechanism","authors":["Mengyang Sun","MaoChuan Dou","Tao Feng","Dan Zhang","Yihao Wang","Junpeng Liu","Yifan Zhu","Jie Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Arccosine loses the orientation needed for planar reconstruction","plainExplanation":"Arccosine loses the orientation needed for planar reconstruction. The ambient planar operator is not a full-dimensional rotation.","primaryUrl":"https://arxiv.org/html/2605.25565","noteUrl":"/review?paper=review-neurips-2026-148910","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148912","title":"Learning Agentic Policy from Action Guidance","authors":["Yuxiang Ji","Zengbin Wang","Yong Wang","Yang","Ziyu Ma","Guanhua Chen","Zonghua Sun","Liaoni Wu","Xiangxiang Chu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Monotone success probabilities do not make sampled successes monotone","plainExplanation":"Monotone success probabilities do not make sampled successes monotone.","primaryUrl":"https://arxiv.org/html/2605.12004","noteUrl":"/review?paper=review-neurips-2026-148912","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148915","title":"Predictive–Generative Drift Decomposition for Speech Enhancement and Separation","authors":["Julius Richter","Yoshiki Masuyama","Christoph Boeddeker","Takahiro Edo","Gordon Wichern","Jonathan Le Roux"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Predictive–Generative Drift Decomposition for Speech Enhancement and Separation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06189","noteUrl":"/review?paper=review-neurips-2026-148915","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148916","title":"DiPhon: Diffusion on Graphons for Scalable Graph Generation","authors":["Sergio Rozada","Yiming QIN","Manuel Madeira","Pascal Frossard","Alejandro Ribeiro"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The diffusion construction needs admissible parameters and the correct reverse law","plainExplanation":"The resolution-dependent parameters do not always define a bounded diffusion. The reverse equation also omits a covariance-derivative term, and a fixed Jacobi eigenexpansion does not describe the general time-dependent target.","primaryUrl":"https://arxiv.org/html/2607.07232","noteUrl":"/review?paper=review-neurips-2026-148916","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148919","title":"Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression","authors":["Tue M Cao","Nguyen Do","My T. Thai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression","plainExplanation":"No issue was found in the checked claims. The earlier direction-of-improvement objection does not apply to the revised definition.","primaryUrl":"https://arxiv.org/html/2605.28567","noteUrl":"/review?paper=review-neurips-2026-148919","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148920","title":"Regularized Large Neighborhood Search","authors":["Germain Vivier-Ardisson","Laurent Demonet","Axel Parmentier","Mathieu Blondel"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The projection and perturbation bounds need stronger conditions","plainExplanation":"The general projection identity fails at an allowed sparse optimum, and a cut-only bias bound misses dependence in the perturbation noise. The routing appendix also has a local depot-indexing error.","primaryUrl":"https://arxiv.org/html/2606.02294","noteUrl":"/review?paper=review-neurips-2026-148920","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","incorrect_bound","lemma_error","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-148922","title":"Self-Supervised Goal-Reaching Results in Multi-Agent Cooperation and Exploration","authors":["Chirayu Nimonkar","Shlok Shah","Catherine Ji","Benjamin Eysenbach"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The unsquared log-sum-exp penalty makes the objective unbounded below","plainExplanation":"The objective adds a positive multiple of an unsquared log-sum-exp regularizer to symmetric contrastive cross-entropy, without bounding encoder scale. Set constant encoders \\(\\phi=0\\), \\(\\psi=ce_1\\), \\(c\\geq0\\), so every score is \\(-c\\).","primaryUrl":"https://arxiv.org/html/2509.10656","noteUrl":"/review?paper=review-neurips-2026-148922","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148923","title":"Causal learning with the invariance principle","authors":["Francesco Montagna","Francesco Locatello"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Causal learning with the invariance principle","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13589","noteUrl":"/review?paper=review-neurips-2026-148923","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148926","title":"TEMPO: Temporal Enforcement via Mode-Separated Policy Optimization for Trustworthy LLM Backtesting","authors":["Zeyu Zhang","Bradly Stadie"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"TEMPO: Temporal Enforcement via Mode-Separated Policy Optimization for Trustworthy LLM Backtesting","plainExplanation":"A decreasing reward need not make its parameter gradient oppose the cost gradient. The missing covariance-centering term can reverse the claimed alignment.","primaryUrl":"https://arxiv.org/html/2605.18843","noteUrl":"/review?paper=review-neurips-2026-148926","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference"],"classifications":["major"]},{"slug":"review-neurips-2026-148929","title":"ProxyPose: 6-DoF Pose Tracking via Video-to-Video Translation","authors":["Ruihang Zhang","Felix Taubner","Pooja Ravi","Kyros Kutulakos","David Lindell"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The multi-face refinement is outside the stated square solver’s domain","plainExplanation":"The multi-face refinement is outside the stated square solver’s domain.","primaryUrl":"https://arxiv.org/html/2607.06555","noteUrl":"/review?paper=review-neurips-2026-148929","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148930","title":"GVCC: Zero-Shot Video Compression via Codebook-Driven Stochastic Rectified Flow","authors":["Ziyue Zeng","Xun Su","Haoyuan Liu","Bingyu Lu","Yui Tatsumi","Hiroshi Watanabe"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GVCC: Zero-Shot Video Compression via Codebook-Driven Stochastic Rectified Flow","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.26571","noteUrl":"/review?paper=review-neurips-2026-148930","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148932","title":"A Statistical Framework for Algorithmic Collective Action with Multiple Collectives","authors":["Claudio Battiloro","Pietro Greiner","Dario Rancati","Bret Nestor","Oumaima Amezgar","Francesca Dominici"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Feature-only planting needs a label-conditional pushforward","plainExplanation":"Feature-only planting needs a label-conditional pushforward.","primaryUrl":"https://arxiv.org/html/2605.06749","noteUrl":"/review?paper=review-neurips-2026-148932","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148933","title":"On Neural Scaling Laws for Weather Emulation through Continual Training","authors":["Shashank Subramanian","Alexander Kiefer","Arnur Nigmetov","Amir Gholami","Dmitriy Morozov","Michael Mahoney"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The coherence numerator and power spectra use different normalizations","plainExplanation":"The coherence numerator and power spectra use different normalizations. The distributed shift slice has the wrong length.","primaryUrl":"https://arxiv.org/html/2603.25687","noteUrl":"/review?paper=review-neurips-2026-148933","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148934","title":"Revisiting Diffusion Fine-Tuning for Unsupervised Domain Adaptation","authors":["Xuan Qi","Yi Wei","Daniele Berardini","Vito Paolo Pastore","Vittorio Murino"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Revisiting Diffusion Fine-Tuning for Unsupervised Domain Adaptation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33716","noteUrl":"/review?paper=review-neurips-2026-148934","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148937","title":"Adaptive Conditional Gradient Sliding: Projection-Free and Line-Search-Free Acceleration","authors":["Shota Takahashi"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The minimum-gap proof applies a post-update bound to the initial iterate","plainExplanation":"The conditional-gradient analysis uses an estimate for \\(u_{t+1}\\) to bound a minimum including \\(u_1\\), producing an immediate one-call gap claim. On \\([0,1]\\), take \\(u_1=1\\), gradient term \\(100\\), coefficient one and \\(\\psi(z)=100z+(z-1)^2/2\\).","primaryUrl":"https://arxiv.org/html/2601.20443","noteUrl":"/review?paper=review-neurips-2026-148937","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148938","title":"On-Policy Consistency Training Improves LLM Safety with Minimal Capability Degradation","authors":["Andy Q Han","Kristina Fujimoto","Avidan Shah","Kiet Nguyen","Kai Xu","Yueh-Han Chen","Ilia Sucholutsky","Rico Angell"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"On-Policy Consistency Training Improves LLM Safety with Minimal Capability Degradation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.21834","noteUrl":"/review?paper=review-neurips-2026-148938","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148940","title":"Resilience Matters for Embodied Agents System: New Metrics, Systematic Evaluation, and Optimization","authors":["Yapeng Liu","Yuanzhao Zhai","Xudong Gong","Feng Dawei","Bo Ding","Lin Wang","Huaimin Wang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Resilience Matters for Embodied Agents System: New Metrics, Systematic Evaluation, and Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.23839","noteUrl":"/review?paper=review-neurips-2026-148940","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148942","title":"Beyond Augmented-Action Surrogates for Multi-Expert Learning-to-Defer","authors":["Yannis Montreuil","Axel Carlier","Lai Xing Ng","Wei Ooi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Logit separability does not decouple a shared representation","plainExplanation":"The zero mixed Hessian between class and expert output logits is interpreted as guaranteeing that expert errors cannot perturb class probabilities, including end-to-end shared-backbone training. Separate logit losses have zero mixed logit Hessian, but an expert loss contributes \\(J_s(\\theta)^\\top(u-t)\\) to shared parameters.","primaryUrl":"https://arxiv.org/html/2604.09414","noteUrl":"/review?paper=review-neurips-2026-148942","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148943","title":"Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model for Cell-Level Dense Prediction","authors":["Weiming Chen","Xitong Ling","Zhenyang Cai","Xidong Wang","Jiawen Li","Tian Guan","Benyou Wang","Yonghong He"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model for Cell-Level Dense Prediction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08276","noteUrl":"/review?paper=review-neurips-2026-148943","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148944","title":"What should post-training optimize? A test-time scaling law perspective","authors":["Muheng Li","Jian Qian","Wenlong Mou"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The stochastic ascent remainder loses the smoothness factor","plainExplanation":"The stochastic ascent remainder loses the smoothness factor.","primaryUrl":"https://arxiv.org/html/2605.10716","noteUrl":"/review?paper=review-neurips-2026-148944","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148947","title":"Reward-free Pretraining for Reinforcement Learning via Occupancy Coverage Maximization","authors":["Marco Pratticò","Pietro Novelli","Massimiliano Pontil","Carlo Ciliberto"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Reward-free Pretraining for Reinforcement Learning via Occupancy Coverage Maximization","plainExplanation":"The declared lattice kernel has infinite self-interactions. Removing the diagonal gives a different, well-defined sum whose scaling changes at exponent one; the continuous exponential calculation has the claimed powers only in its integrable range. This corrects the toy comparison without changing the general bounded-kernel objective.","primaryUrl":"https://arxiv.org/html/2606.21271","noteUrl":"/review?paper=review-neurips-2026-148947","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148948","title":"Scaling Multi-Teacher Distillation for Digital Pathology","authors":["Sofiène Boutaj","Pierre Marza","Varun Belagali","Dimitris Samaras","Maria Vakalopoulou","Stergios Christodoulidis"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Scaling Multi-Teacher Distillation for Digital Pathology","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.12574","noteUrl":"/review?paper=review-neurips-2026-148948","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148949","title":"OpenCoF: Learning to Reason Through Video Generation","authors":["Xinyan Chen","Renrui Zhang","Ziyu Guo","Dongzhi JIANG","Hongsheng Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"OpenCoF: Learning to Reason Through Video Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.08763","noteUrl":"/review?paper=review-neurips-2026-148949","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148952","title":"Dynin-Robotics: Omnimodal Unified Diffusion Vision-Language-Action Model","authors":["Hoeun Lee","Hyeonggeun KIM","Jaeik Kim","Jusang Oh","Geon Choi","Jinhyeok Kim","Jaeyoung Do"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Dynin-Robotics: Omnimodal Unified Diffusion Vision-Language-Action Model","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.13053","noteUrl":"/review?paper=review-neurips-2026-148952","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148954","title":"3D-PLOT-LLM: Part-Level Object Tokens for 3D Large Language Models","authors":["Jintang Xue","Xinyu Wang","Yixing Wu","Jingwen Chen","C.-C. J Kuo"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"3D-PLOT-LLM: Part-Level Object Tokens for 3D Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19828","noteUrl":"/review?paper=review-neurips-2026-148954","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148955","title":"On the Price of Privacy for Language Identification and Generation","authors":["Xiaoyu Li","Andi Han","Jiaojiao Jiang","Junbin Gao"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Privacy schedules and valid error-rate deductions","plainExplanation":"The advertised range permits noise that dominates the bounded score differences. For a fixed language family and fixed distribution, Algorithm 7 can have a positive error lower bound for every allowed horizon choice. A fixed positive mass floor cannot be used to rescue this broad claim while simultaneously requiring an unbounded window in an infinite supported language.","primaryUrl":"https://arxiv.org/html/2604.07238","noteUrl":"/review?paper=review-neurips-2026-148955","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","invalid_inference","notation","numerical_error"],"classifications":["major"]},{"slug":"review-neurips-2026-148959","title":"Reject, Resample, Repeat: Understanding Parallel Reasoning in Language Model Inference","authors":["Noah Golowich","Fan Chen","Dhruv Rohatgi","Raghav Singhal","Carles Domingo i Enrich","Dylan J Foster","Akshay Krishnamurthy"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Reject, Resample, Repeat: Understanding Parallel Reasoning in Language Model Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.07887","noteUrl":"/review?paper=review-neurips-2026-148959","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148960","title":"Self-Consuming Generative Models with Co-Evolving Human Preferences","authors":["Xiukun Wei","Tian Xie","Ding Zhu","Xueru Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Repairing the expectation-ratio step","plainExplanation":"The pointwise ratio identity is correct, but averaging it does not produce a ratio of separately averaged quantities. The appropriate averaging measure is weighted by \\(g_i\\).","primaryUrl":"https://arxiv.org/html/2610.09415","noteUrl":"/review?paper=review-neurips-2026-148960","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148963","title":"A Generalized Tikhonov Layer for Interpretable-by-design Graph Neural Networks","authors":["Nicolas Tremblay","Filippo Maria Bianchi","Benjamin Ricaud"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"A Generalized Tikhonov Layer for Interpretable-by-design Graph Neural Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28578","noteUrl":"/review?paper=review-neurips-2026-148963","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148967","title":"MetaCluster: Enabling Deep Compression of Kolmogorov-Arnold Network","authors":["Matthew Raffel","Adwaith Renjith","Lizhong Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MetaCluster: Enabling Deep Compression of Kolmogorov-Arnold Network","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.19105","noteUrl":"/review?paper=review-neurips-2026-148967","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148968","title":"Learning Fractional-Order Dynamics from a Single Trajectory","authors":["Xiaole Zhang","Ziyi Zhang","zehao zhao","Stephen Tu","Guannan Qu","Yorie Nakahira","Paul Bogdan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Fractional-Order Dynamics from a Single Trajectory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.18127","noteUrl":"/review?paper=review-neurips-2026-148968","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148969","title":"Stepwise Penalization for Length-Efficient Chain-of-Thought Reasoning","authors":["Xintong Li","Sha Li","Rongmei Lin","Hongye Jin","Linwei Li","Hejie Cui","Sarah Zhang","Chia-Yuan Chang","Kewei Cheng","Yuwei Zhang","Qingyu Yin","Jingbo Shang","Bing Yin"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Stepwise Penalization for Length-Efficient Chain-of-Thought Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.00296","noteUrl":"/review?paper=review-neurips-2026-148969","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148970","title":"Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL","authors":["Darshan Deshpande"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.16204","noteUrl":"/review?paper=review-neurips-2026-148970","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148971","title":"From Zero to Hero: Training-Free Custom Concept Spawning in World Models","authors":["Kiymet Akdemir","Pinar Yanardag"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"From Zero to Hero: Training-Free Custom Concept Spawning in World Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02575","noteUrl":"/review?paper=review-neurips-2026-148971","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148973","title":"SplitZip: Ultra Fast Lossless KV Compression for Disaggregated LLM Serving","authors":["Yipin Guo","Siddharth Joshi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SplitZip: Ultra Fast Lossless KV Compression for Disaggregated LLM Serving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.01708","noteUrl":"/review?paper=review-neurips-2026-148973","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148974","title":"LoReC: Rethinking Large Language Models for Graph Data Analysis","authors":["Hongyu Zhan","Qixin Wang","Yusen Tan","Hai-tao Yu","Jingbo Zhou","Shuai Chen","Jia Li","Xiao Tan","Jun Xia"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The edge-dropout expression can be negative and reverses its stated ordering","plainExplanation":"The edge-dropout expression can be negative and reverses its stated ordering.","primaryUrl":"https://arxiv.org/html/2604.17897","noteUrl":"/review?paper=review-neurips-2026-148974","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148975","title":"PACE: A Proxy for Agentic Capability Evaluation","authors":["Yueqi Song","Lintang Sutawika","Jiarui Liu","Lindia Tjuatja","Jiayi Geng","Yunze Xiao","Daniel Lee","Aditya B Soni","Vincent Lo","Xiang Yue","Graham Neubig"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"PACE: A Proxy for Agentic Capability Evaluation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.02032","noteUrl":"/review?paper=review-neurips-2026-148975","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148976","title":"PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion","authors":["Yifan Lu","Qi Wu","Jay Zhangjie Wu","Zian Wang","Huan Ling","Sanja Fidler","Xuanchi Ren"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23902","noteUrl":"/review?paper=review-neurips-2026-148976","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148977","title":"Polynomial-Time Algorithm for Thiele Voting Rules with Voter Interval Preferences","authors":["Pasin Manurangsi","Krzysztof Sornat"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The optimality display mixes penalized and unpenalized values","plainExplanation":"The proof comparator does not order crossing intervals. The optimality display mixes penalized and unpenalized values.","primaryUrl":"https://arxiv.org/html/2604.05953","noteUrl":"/review?paper=review-neurips-2026-148977","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148979","title":"Rethinking Contrastive Loss in CLIP Post-training: A Complementary Framework with Frozen Text Encoder","authors":["Zidan Wang","Yaqian Li","xiaokai zhang","Kun He","Kaiwen Long","Hanpeng Liu"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Rethinking Contrastive Loss in CLIP Post-training: A Complementary Framework with Frozen Text Encoder","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.11374","noteUrl":"/review?paper=review-neurips-2026-148979","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148982","title":"Sharpness of Minima in Deep Matrix Factorization","authors":["Anıl Kamber","Rahul Parhi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Sharpness of Minima in Deep Matrix Factorization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.25783","noteUrl":"/review?paper=review-neurips-2026-148982","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148983","title":"Learning to play with spikes. Characterizing, predicting, and engineering unsupervised plasticity rules for spiking reservoir computing","authors":["Maciej Kania","Basile Confavreux","Tim Vogels"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning to play with spikes. Characterizing, predicting, and engineering unsupervised plasticity rules for spiking reservoir computing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://www.biorxiv.org/content/biorxiv/early/2026/09/21/2026.09.15.751573.full.pdf","noteUrl":"/review?paper=review-neurips-2026-148983","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-148984","title":"NEvo: Neural-Guided Evolutionary Video Synthesis for Dynamic Visual Selectivity","authors":["Yingtian Tang","Sogand Salehi","Ming Zhou","Amir Zamir","Leyla Isik","Martin Schrimpf"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"NEvo: Neural-Guided Evolutionary Video Synthesis for Dynamic Visual Selectivity","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.02317","noteUrl":"/review?paper=review-neurips-2026-148984","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148986","title":"S2MDF: A Plug-And-Play Layer for Intersection-Free Multi-Object Signed Distance Fields","authors":["Deniz Sayin Mercadier","Federico Stella","Aurel Bizeau","Nicolas Talabot","Pascal Fua"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The strict constraint excludes the claimed quadratic-program optimum","plainExplanation":"The printed strict QP constraint excludes the boundary point returned by Shift-All. In an admissible example the infimum is approached but no optimum is attained under that strict condition. The closed-QP Shift-All result survives. The failure is confined to the strict sign in this local display.","primaryUrl":"https://arxiv.org/html/2605.29761","noteUrl":"/review?paper=review-neurips-2026-148986","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-148987","title":"Native Audio-Visual Alignment for Generation","authors":["Longbin Ji","Guan Wang","xuan wei","Chenye Yang","Zhenyu Zhang","Shuohuan Wang","YU SUN","Jingzhou HE"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Native Audio-Visual Alignment for Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.30073","noteUrl":"/review?paper=review-neurips-2026-148987","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-148988","title":"Intrinsic Riemannian Cross-covariance for Manifold-valued Random Objects","authors":["Carlos Soto","Cheng Wang","Yujing Huang","Xiaoyu Chen"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The plug-in Kronecker factor order conflicts with column vectorization","plainExplanation":"The asymptotic expansion fixes column stacking but represents the perturbation \\(U(D\\delta)^\\top\\) using \\(U\\otimes D\\). Column vectorization gives \\(\\operatorname{vec}(U(D\\delta)^\\top)=(D\\otimes U)\\delta\\).","primaryUrl":"https://arxiv.org/html/2606.10212","noteUrl":"/review?paper=review-neurips-2026-148988","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148989","title":"ONE-SHOT: Compositional Human-Environment Video Synthesis via Spatial-Decoupled Motion Injection and Hybrid Context Integration","authors":["Fengyuan Yang","Luying Huang","Jiazhi Guan","Quanwei Yang","Dongwei Pan","Jianglin Fu","Haocheng Feng","Wei He","Kaisiyuan Wang","Hang Zhou","Angela Yao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Canonical box coordinates require translation as well as scaling","plainExplanation":"Canonical box coordinates require translation as well as scaling.","primaryUrl":"https://arxiv.org/html/2604.01043","noteUrl":"/review?paper=review-neurips-2026-148989","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148995","title":"VisInteract: Towards Dynamic Interactive Text-to-Visualization under Imperfect Queries","authors":["Wenxin XU","Jinwei Lu","Hwanhee Kim","Chen J Zhang","Xiao-Yong Wei","Haoyang LI","Yuanfeng SONG"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"VisInteract: Towards Dynamic Interactive Text-to-Visualization under Imperfect Queries","plainExplanation":"The multiplicative rescaling preserves the weighted total before clipping. If a positively weighted component exceeds one, clipping removes part of that total.","primaryUrl":"https://arxiv.org/html/2609.15182","noteUrl":"/review?paper=review-neurips-2026-148995","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-148996","title":"Automata from Agent Traces: Failure and Next-Step Prediction","authors":["Seonglae Cho","Franklin Cardenoso Fernandez","Umar Mohammed","Zekun Wu","Kleyton da Costa","Ilham Wicaksono","Adriano Koshiyama"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The statistical guarantees need assumptions beyond independent traces","plainExplanation":"Singleton filtering, repeated visits within a trace, variable trace lengths and fixed alarm thresholds each require conditions missing from the corresponding guarantee. A compact FSM also does not by itself encode the trace corpus.","primaryUrl":"https://arxiv.org/html/2608.23670","noteUrl":"/review?paper=review-neurips-2026-148996","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","incorrect_bound","invalid_inference","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-148997","title":"Colored Noise Diffusion Sampling","authors":["Hadar Davidson","Noam Issachar","Sagie Benaim"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The distribution and energy identities omit effects of the changed dynamics","plainExplanation":"Conserving total injected noise variance does not preserve its covariance matrix or the target diffusion law. The terminal energy-gap identity also omits changes in the sampling distributions.","primaryUrl":"https://arxiv.org/html/2605.30332","noteUrl":"/review?paper=review-neurips-2026-148997","collection":"neurips-2026","severity":"major","issueTypes":["invalid_inference","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-148999","title":"On Observation Time for Recovering Latent Hawkes Networks","authors":["Jonas Linkerhägner","Michele Bortolasi","Lorenzo Baldassari","Maarten V. de Hoop","Ivan Dokmanić"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"On Observation Time for Recovering Latent Hawkes Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08400","noteUrl":"/review?paper=review-neurips-2026-148999","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149000","title":"What Should a Streaming Video Model Remember?","authors":["Haonan Ge","Yiwei Wang","Hang Wu","Yujun Cai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"What Should a Streaming Video Model Remember?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.16353","noteUrl":"/review?paper=review-neurips-2026-149000","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149001","title":"Beyond the Training Distribution: Evaluating Predictions Under Distribution Shift and Selection Bias","authors":["Annie Ulichney","Amanda L Coston"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Beyond the Training Distribution: Evaluating Predictions Under Distribution Shift and Selection Bias","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.14506","noteUrl":"/review?paper=review-neurips-2026-149001","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149004","title":"World Action Verifier: Self-Improving World Models via Forward-Inverse Asymmetry","authors":["Yuejiang Liu","Fan Feng","Lingjing Kong","Weifeng Lu","Jinzhou Tang","Kun Zhang","Kevin Murphy","Chelsea Finn","Yilun Du"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Structural inverse identifiability does not guarantee exact finite-sample learning","plainExplanation":"Theorem F.8 concludes exact recovery of every on-support action by an inverse trained on sampled transitions, using structural alignment and identifiability conditions. Let true dynamics be \\(y=x+a\\) for \\(x,a\\in[0,1]\\), with identity observed representation.","primaryUrl":"https://arxiv.org/html/2604.01985","noteUrl":"/review?paper=review-neurips-2026-149004","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149005","title":"Kernel-based guarantees for nonlinear parametric models in Bayesian optimization","authors":["Rafael Oliveira"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A representing section need not have minimum norm","plainExplanation":"The canonical section represents the function but need not minimize that infimum. The norm and distance equalities, and conclusions using them as equalities, require the projection correction or a totality assumption.","primaryUrl":"https://arxiv.org/html/2605.13160","noteUrl":"/review?paper=review-neurips-2026-149005","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149006","title":"DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model","authors":["Shurui Liu","Weide Chen","Changwang Yi","Ancong Wu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.35242","noteUrl":"/review?paper=review-neurips-2026-149006","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149007","title":"Sign-Based Optimizers Are Effective Under Heavy-Tailed Noise","authors":["Dingzhi Yu","Hongyi Tao","Yuanyu Wan","Lijun Zhang"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Sign-Based Optimizers Are Effective Under Heavy-Tailed Noise","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.07425","noteUrl":"/review?paper=review-neurips-2026-149007","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149009","title":"Bifurcation Models: Learning Set-Valued Solution Maps with Weight-Tied Dynamics","authors":["Caleb Jore","Jialin Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Bifurcation Models: Learning Set-Valued Solution Maps with Weight-Tied Dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07277","noteUrl":"/review?paper=review-neurips-2026-149009","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149015","title":"Impacts of Aggregation on Model Diversity and Consumer Utility","authors":["Kate Donahue","Manish Raghavan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Task caps invalidate the universal greedy allocation rule","plainExplanation":"With unequal task caps, saturating the first preferred task can be worse than splitting the budget. A separate definition formula also fails to express the stronger substitutability condition stated in prose.","primaryUrl":"https://arxiv.org/html/2602.23293","noteUrl":"/review?paper=review-neurips-2026-149015","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","invalid_inference","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-149018","title":"PCBSchemaGen: Reward-Guided LLM Code Synthesis for Printed Circuit Boards (PCB) Schematic Design with Structured Verification","authors":["Huanghaohe Zou","Peng Han","Emad Nazerian","Mafu Zhang","Zhicheng Guo","Alex Huang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"PCBSchemaGen: Reward-Guided LLM Code Synthesis for Printed Circuit Boards (PCB) Schematic Design with Structured Verification","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.00510","noteUrl":"/review?paper=review-neurips-2026-149018","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149021","title":"Leech Lattice Vector Quantization for Efficient LLM Compression","authors":["Tycho F van der Ouderaa","Mart van Baalen","Paul Whatmough","Markus Nagel"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The row-wise compensation uses the wrong Hessian block order","plainExplanation":"The quadratic correction partitions coordinates into fixed quantized block \\(Q\\) and adjustable block \\(R\\), then gives a row-wise update claimed equivalent to the correct column formula. Differentiating gives \\(\\delta_R=-H_{RR}^{-1}H_{RQ}\\delta_Q\\).","primaryUrl":"https://arxiv.org/html/2603.11021","noteUrl":"/review?paper=review-neurips-2026-149021","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149022","title":"To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent Systems","authors":["Pengfei He","Zhenwei Dai","Xianfeng Tang","Yue XING","Hui Liu","Jingying Zeng","Qiankun Peng","Shrivats Agrawal","Samarth Varshney","Suhang Wang","Jiliang Tang","Qi He"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The inclusive violation window has one more flag than its denominator","plainExplanation":"The trust record describes a proportion of violating messages and sums Bernoulli flags from index \\(k_1\\) through \\(k_2\\), dividing by \\(k_2-k_1\\). The inclusive window contains \\(k_2-k_1+1\\) flags.","primaryUrl":"https://arxiv.org/html/2506.02546","noteUrl":"/review?paper=review-neurips-2026-149022","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149023","title":"I-Perceive: A Foundation Model for Vision-Language Active Perception","authors":["Yongxi Huang","Wang","Wenjing Tang","Xinyu He","Cewu Lu","Panpan Cai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"I-Perceive: A Foundation Model for Vision-Language Active Perception","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.00600","noteUrl":"/review?paper=review-neurips-2026-149023","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149024","title":"NOFE – Neural Operator Function Embedding","authors":["Lars Uebbing","Harald Lykke Joakimsen","Siyan Chen","Georgios Leontidis","Kristoffer Wickstrøm","Michael Kampffmeyer","Sébastien Lefèvre","Arnt B Salberg","Robert Jenssen"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Target exclusion makes the affinities odds rather than probabilities","plainExplanation":"The affinity denominator excludes the target neighbor \\(j\\), although the resulting values are used as a probability row in a KL loss. With two outgoing neighbors having kernels \\(a,b\\), the displayed row is \\((a/b,b/a)\\), which does not sum to one.","primaryUrl":"https://arxiv.org/html/2605.11970","noteUrl":"/review?paper=review-neurips-2026-149024","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149029","title":"Self-Rewarded Multimodal Coherent Reasoning Across Diverse Visual Domains","authors":["jusheng zhang","Ningyuan Liu","Kaitong Cai","Sidi Liu","Ziliang Chen","Wenhao Wang","Keze Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The cooling rule, objective and gradient analysis do not agree","plainExplanation":"The main and supplementary cooling gates favor opposite confidence regimes. The expected log-ratio loss also differs from the group-centered policy-gradient estimator, whose coefficients are not individually scaled by each sample’s cooling weight.","primaryUrl":"https://arxiv.org/html/2512.22545","noteUrl":"/review?paper=review-neurips-2026-149029","collection":"neurips-2026","severity":"major","issueTypes":["definition_ambiguity","false_statement","formula_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149032","title":"Quantifying Concentration Phenomena of Mean-Field Transformers in the Low-Temperature Regime","authors":["Albert Alcalde","Leon Bungert","Konstantin Riedl","Tim Roith"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Quantifying Concentration Phenomena of Mean-Field Transformers in the Low-Temperature Regime","plainExplanation":"The printed prefactor can be much smaller than the initial Wasserstein distance, including under the theorem’s positive-density assumptions. Restoring the square-root prefactor corrects the quantitative bound while retaining its exponential decay rate.","primaryUrl":"https://arxiv.org/html/2605.10931","noteUrl":"/review?paper=review-neurips-2026-149032","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149033","title":"Generating Financial Time Series by Matching Random Convolutional Features","authors":["Konrad Mueller","Nikita Zozoulenko","Ben Wood","Thomas Cass","Lukas Gonon"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Generating Financial Time Series by Matching Random Convolutional Features","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.05138","noteUrl":"/review?paper=review-neurips-2026-149033","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149035","title":"One Language-Free Foundation Model Is Enough for Universal Vision Anomaly Detection","authors":["Bin-Bin Gao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"One Language-Free Foundation Model Is Enough for Universal Vision Anomaly Detection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.05552","noteUrl":"/review?paper=review-neurips-2026-149035","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149036","title":"A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering","authors":["Zhanliang Wang","Jiancong Xiao","Ruochen Jin","Shu Yang","Bojian Hou","Li Shen"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering","plainExplanation":"An overall overconfidence gap can hide underconfident bins. Confidence inflation can reduce their calibration error and reverse the claimed binned-ECE ordering.","primaryUrl":"https://arxiv.org/html/2605.08432","noteUrl":"/review?paper=review-neurips-2026-149036","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149038","title":"Local Intrinsic Dimension Unveils Hallucinations in Diffusion Models","authors":["Bartlomiej Sobieski","Matthew Tivnan","Dawid Płudowski","Michał J Włodarczyk","Pengfei Jin","Przemyslaw Prem Biecek","Quanzheng Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The operational guidance is not an exact finite-time true-LID tilt","plainExplanation":"The operational guidance is not an exact finite-time true-LID tilt.","primaryUrl":"https://arxiv.org/html/2605.05026","noteUrl":"/review?paper=review-neurips-2026-149038","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149039","title":"Safe Few-Step Generation via Velocity Editing","authors":["Yujin Choi","Jaehong Yoon"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The clean-estimate safety gradient needs the denoiser Jacobian","plainExplanation":"The scorer approximation replaces the unsafe posterior by \\(g(m(x_t))\\), with \\(m(x_t)=x_t-tv_t(x_t)\\), then uses its clean-estimate gradient as the noisy-state gradient. The chain rule gives \\(-\\nabla_x\\log(1-g(m(x)))=(I-tDv_t)^\\top\\nabla_mg(m)/(1-g(m))\\).","primaryUrl":"https://arxiv.org/html/2606.23267","noteUrl":"/review?paper=review-neurips-2026-149039","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149040","title":"DriveFuture: Future-Aware Latent World Models for Autonomous Driving","authors":["Yufeng hong","Xiaotian Zhou","Yingyan Li","Xiangpo Zhou","Lin Liu","Yadan Luo","Shaoqing Xu","Lei Yang","Ziying Song"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"DriveFuture: Future-Aware Latent World Models for Autonomous Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09701","noteUrl":"/review?paper=review-neurips-2026-149040","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149041","title":"Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control","authors":["Zunzhe Zhang","Runhan Huang","Yicheng Liu","Shaoting Zhu","Linzhan Mou","Hang Zhao"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.17834","noteUrl":"/review?paper=review-neurips-2026-149041","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149042","title":"LogSig-SSM: Time-Series Modelling with Multi-Scale Log-Signature Compression for State-Space Models","authors":["Felix Oury","Nicolas C Peiro","Reiko J Tanaka"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LogSig-SSM: Time-Series Modelling with Multi-Scale Log-Signature Compression for State-Space Models","plainExplanation":"A compact family containing one path admits only constant restricted functionals, and the comparator class already represents constants. Part (ii) therefore needs an explicit richer path family and a separation proof on that family; the separate token-approximation statement is unaffected.","primaryUrl":"https://arxiv.org/html/2610.05051","noteUrl":"/review?paper=review-neurips-2026-149042","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149043","title":"LoRIF: Low-Rank Influence Functions for Scalable Training Data Attribution","authors":["Shuangqi Li","Hieu Minh Le","Jingyi Xu","Mathieu Salzmann"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The general-rank gradient product needs a scalar Frobenius contraction","plainExplanation":"The rank-factor influence formula replaces the scalar Frobenius product of gradients \\(G=uv^\\top\\) by an ordinary product of two cross-Gram matrices. The scalar identity is \\(\\langle G_{te},G_{tr}\\rangle_F=\\operatorname{tr}[(u_{te}^\\top u_{tr})(v_{te}^\\top v_{tr})^\\top]\\).","primaryUrl":"https://arxiv.org/html/2601.21929","noteUrl":"/review?paper=review-neurips-2026-149043","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149045","title":"Stay Fair! Ensuring Group Fairness in Diffusion Models Across Guidance Scales","authors":["Myeongsoo Kim","Eunji Kim","Minwoo Chae","Sangwoo Mo"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Stay Fair! Ensuring Group Fairness in Diffusion Models Across Guidance Scales","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28036","noteUrl":"/review?paper=review-neurips-2026-149045","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149047","title":"SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning","authors":["Seokju Cho","Ryo Hachiuma","Abhishek Badki","Hang Su","Byung-Kwan Lee","Chan Hee Song","Sifei Liu","Subhashree Radhakrishnan","Seungryong Kim","Frank Wang","Min-Hung Chen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.13673","noteUrl":"/review?paper=review-neurips-2026-149047","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149048","title":"Negation Neglect: When models fail to learn negations in training","authors":["Harry Mayne","Lev McKinney","Jan Dubiński","Adam Karvonen","James Chua","Owain Evans"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Negation Neglect: When models fail to learn negations in training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13829","noteUrl":"/review?paper=review-neurips-2026-149048","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149049","title":"Population-Aligned Persona Generation for LLM-based Social Simulation","authors":["Zhengyu Hu","Jianxun Lian","Zheyuan Xiao","Max Xiong","Tianfu Wang","Teng Xiao","Yuxuan Lei","Fengqing Jiang","Kaize Ding","Ziang Xiao","Nicholas Jing Yuan","Xing Xie","Radha Poovendran"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The alignment guarantees do not cover the stated finite sampling procedure","plainExplanation":"Theorem 1 omits finite-pool discrepancy and uses incorrect Wasserstein rates. The transport-plan cost guarantee does not transfer to the final row-cost resampling rule, and top-weight truncation differs from the multinomial sampling law analyzed in the proof.","primaryUrl":"https://arxiv.org/html/2509.10127","noteUrl":"/review?paper=review-neurips-2026-149049","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","formula_error","incorrect_bound","invalid_inference","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149050","title":"Toward Cognitive Supersensing in Multimodal Large Language Models","authors":["Boyi Li","Yifan Shen","Yuanzhe Liu","Yifan Xu","Jiateng Liu","Xinzhuo Li","Zhengyuan Li","Jingyuan Zhu","Yunhan Zhong","Lan Fangzhou","Jianguo Cao","James Rehg","Heng Ji","Ismini Lourentzou","Xu Cao"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Toward Cognitive Supersensing in Multimodal Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.01541","noteUrl":"/review?paper=review-neurips-2026-149050","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149052","title":"Improving Conditional Modeling via Inter-Class Likelihood-Ratio Maximization and Unifying Classifier-Free Guidance with Alignment Objectives","authors":["Xiang Li","Yixuan Jia","Xiao Li","Jeffrey Fessler","Rongrong Wang","Qing Qu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Improving Conditional Modeling via Inter-Class Likelihood-Ratio Maximization and Unifying Classifier-Free Guidance with Alignment Objectives","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.22364","noteUrl":"/review?paper=review-neurips-2026-149052","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149055","title":"Fast Rates for Offline Contextual Bandits with Forward-KL Regularization under Single-Policy Concentrability","authors":["Qingyue Zhao","Kaixuan Ji","Heyang Zhao","Quanquan Gu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The upper-bound proof must retain the pessimism penalty","plainExplanation":"Subtracting a nonconstant confidence bonus changes the optimizing policy even when empirical rewards are exact. The raw regression error can vanish while the pessimistic estimate and policy still differ from their true counterparts.","primaryUrl":"https://arxiv.org/html/2605.09214","noteUrl":"/review?paper=review-neurips-2026-149055","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149057","title":"Prudent-Banker: No Extra Fees for Baseline Safety in Adversarial Bandits With and Without Delays","authors":["Ting Hu","Luanda Cai","Emmanouil-Vasileios Vlatakis-Gkaragkounis"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The restart test uses delays unavailable at the decision time","plainExplanation":"The restart test uses delays unavailable at the decision time. The simplified delay term does not retain the stated margin dependence. A fixed-comparator expectation bound does not cover a data-selected comparator.","primaryUrl":"https://arxiv.org/html/2605.23351","noteUrl":"/review?paper=review-neurips-2026-149057","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149059","title":"PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop","authors":["Xinge Peng","Yiting Lu","tianwu zhi","Wen Wen","Jianzhao Liu","Xin Li","Zhibo Chen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.00559","noteUrl":"/review?paper=review-neurips-2026-149059","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149060","title":"Intelligence per Watt: Measuring Intelligence Efficiency of Local AI","authors":["Jon Saad-Falcon","Avanika Narayan","Hakki Akengin","J. W Griffin","Herumb Shandilya","Adrian G Lafuente","Medhya Goel","Rebecca Joseph","Shlok Natarajan","Etash Guha","Shang Zhu","Ben Athiwaratkun","Azalia Mirhoseini","Christopher Ré"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Intelligence per Watt: Measuring Intelligence Efficiency of Local AI","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2511.07885","noteUrl":"/review?paper=review-neurips-2026-149060","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149061","title":"MC$^2$Mark: Distortion-Free Multi-Bit Watermarking for Long Messages","authors":["Xuehao Cui","Ruibo Chen","Yihan Wu","Weidong Cai","Heng Huang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The watermark redistribution needs boundary and zero-residual cases","plainExplanation":"The watermark redistribution needs boundary and zero-residual cases.","primaryUrl":"https://arxiv.org/html/2602.14030","noteUrl":"/review?paper=review-neurips-2026-149061","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149062","title":"High dimensional theory of two-phase optimizers","authors":["Atish Agarwala"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Worker-dependent noise does not strictly increase the entire spectrum","plainExplanation":"The rate example claims strict improvement beyond \\(\\ell\\) when the inner transfer is \\(F=a e^{-\\ell S}\\) and outer Nesterov is set to its critical value. At \\(\\nu=1/(1-\\beta)\\), the update matrix has trace \\((1+\\beta)F\\) and determinant \\(\\beta F\\).","primaryUrl":"https://arxiv.org/html/2603.26954","noteUrl":"/review?paper=review-neurips-2026-149062","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149065","title":"Neural Structural Reasoner: A Brain-inspired Architecture for Reasoning over Structured Knowledge","authors":["Zixing Jia","Yuhang Pan","Ni Ji"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The online Oja update is not the offline running mean","plainExplanation":"The online Oja update is not the offline running mean.","primaryUrl":"https://arxiv.org/html/2609.36620","noteUrl":"/review?paper=review-neurips-2026-149065","collection":"neurips-2026","severity":"repairable","issueTypes":["mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149066","title":"Spectral Graph Sparsification Preserves Representation Geometry in Graph Neural Networks","authors":["Sanjukta Krishnagopal"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Spectral Graph Sparsification Preserves Representation Geometry in Graph Neural Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.01136","noteUrl":"/review?paper=review-neurips-2026-149066","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149067","title":"Unveiling Fine-Grained Visual Traces: Evaluating MultiModal Interleaved Reasoning Chains in Multimodal STEM Tasks","authors":["Jing Jin","Hao Liu","Yan Bai","Yihang Lou","Zhenke Wang","Tianrun Yuan","Yongkang Zhu","Juntong Chen","Fanhu Zeng","Xuanyu Zhu","Tao Feng","Yige Xu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Unveiling Fine-Grained Visual Traces: Evaluating MultiModal Interleaved Reasoning Chains in Multimodal STEM Tasks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.19697","noteUrl":"/review?paper=review-neurips-2026-149067","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149068","title":"Controlling Transient Amplification Improves Long-horizon Rollouts","authors":["Adeel Pervez","Francesco Locatello"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Nonnormality need not make spectral radius strictly smaller than norm","plainExplanation":"Remark A.3 claims strict inequality between spectral radius and spectral norm for every nonnormal contractive matrix. Take the direct sum of scalar \\(0.8\\) and \\(\\begin{pmatrix}0&0.4\\\\0&0\\end{pmatrix}\\).","primaryUrl":"https://arxiv.org/html/2605.08856","noteUrl":"/review?paper=review-neurips-2026-149068","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149069","title":"All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow","authors":["Wei Dong","Han Zhou","Terry Ji","Guanhua Zhao","Shahab Asoodeh","Yulun Zhang","Guangtao Zhai","Jun Chen","Xiaohong Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14045","noteUrl":"/review?paper=review-neurips-2026-149069","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149072","title":"Tracing Persona Vectors Through LLM Pretraining","authors":["Viktor Moskvoretskii","Dominik Glandorf","Jorge Medina Moreira","Tanja Käser","Robert West"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Tracing Persona Vectors Through LLM Pretraining","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13329","noteUrl":"/review?paper=review-neurips-2026-149072","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149076","title":"ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning","authors":["Di Zhu","Ziheng Yan","Fang Wan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34982","noteUrl":"/review?paper=review-neurips-2026-149076","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149077","title":"SS3D: End2End Self-Supervised 3D from Web Videos","authors":["Marwane Hariat","Gianni Franchi","David Filliat","Antoine Manzanera"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"SS3D: End2End Self-Supervised 3D from Web Videos","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.22686","noteUrl":"/review?paper=review-neurips-2026-149077","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149078","title":"Debiased Counterfactual Generation via Flow Matching from Observations","authors":["Hugh Dance","Johnny Xi","Peter Orbanz","Benjamin Bloem-Reddy"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A monotone endpoint flow need not minimize the intervening path energy","plainExplanation":"A monotone endpoint flow need not minimize the intervening path energy. Arm-restricted minibatches omit the population nuisance average. A product-measure kernel norm does not control a V-statistic diagonal. The relative empirical-process rate omits an additive oracle term.","primaryUrl":"https://arxiv.org/html/2605.07665","noteUrl":"/review?paper=review-neurips-2026-149078","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","lemma_error","mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149080","title":"Simulating Students or Sycophantic Problem Solving? On Misconception Faithfulness of LLM Simulators","authors":["Heejin Do","Shashank Sonkar","Mrinmaya Sachan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Simulating Students or Sycophantic Problem Solving? On Misconception Faithfulness of LLM Simulators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12748","noteUrl":"/review?paper=review-neurips-2026-149080","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149083","title":"Rethinking Forward Processes for Score-Based Nonlinear Data Assimilation in High Dimensions","authors":["EUNBEE YOUN","Donghan Kim","Dae W Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rethinking Forward Processes for Score-Based Nonlinear Data Assimilation in High Dimensions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.02889","noteUrl":"/review?paper=review-neurips-2026-149083","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149085","title":"PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation","authors":["Yidong Huang","Zun Wang","Han Lin","Dong-Ki Kim","Shayegan Omidshafiei","Jaehong Yoon","Jaemin Cho","Yue Zhang","Mohit Bansal"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Cartesian acceleration times rotational inertia does not have torque units","plainExplanation":"Cartesian acceleration times rotational inertia does not have torque units.","primaryUrl":"https://arxiv.org/html/2605.14269","noteUrl":"/review?paper=review-neurips-2026-149085","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149087","title":"Convergence Analysis of Newton's Method for Neural Networks in the Overparameterized Limit","authors":["Konstantin Riedl","Justin Sirignano","Konstantinos Spiliopoulos"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Convergence Analysis of Newton's Method for Neural Networks in the Overparameterized Limit","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08352","noteUrl":"/review?paper=review-neurips-2026-149087","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149088","title":"Effective Multi-sensor Conditioning for Street-view Novel-view Synthesis","authors":["Zhengfei Kuang","Adam Sun","Liyuan Zhu","Tong Wu","Shengqu Cai","Jonathan Tremblay","Iro Armeni","Ehsan Adeli","Lior Yariv","Gordon Wetzstein"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The value transform is not undone by a homogeneous-transform transpose","plainExplanation":"Relative-camera attention inverse-transforms values and applies \\(D^\\top\\) outside the weighted sum, although \\(D\\) is a homogeneous pose transform. For a single token the attention weight is one, and the output is \\(D^\\top D^{-1}v\\).","primaryUrl":"https://arxiv.org/html/2606.01590","noteUrl":"/review?paper=review-neurips-2026-149088","collection":"neurips-2026","severity":"repairable","issueTypes":["implementation_mismatch"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149090","title":"On the Complexity of Offline Reinforcement Learning with Q*-Approximation and Partial Coverage","authors":["Haolin Liu","Braham Snyder","Chen-Yu Wei"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Pseudoinverse coverage does not control off-span Bellman error","plainExplanation":"The coverage argument uses a pseudoinverse quadratic form as if it controlled all target feature directions without a range condition. For \\(\\Sigma=\\operatorname{diag}(0,1)\\), \\(X_*=(1,0)\\) and \\(W=(-1,0)\\), the squared target inner product is one while both quadratic factors in the proposed generalized Cauchy–Schwarz bound are zero.","primaryUrl":"https://arxiv.org/html/2602.12107","noteUrl":"/review?paper=review-neurips-2026-149090","collection":"neurips-2026","severity":"repairable","issueTypes":["mathematical_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149092","title":"CORVUS: Context Optimization and Reduction Via Underlying Synchronization for LLM Coding Agents","authors":["Mingwei Zheng","David OBrien","Siwei Cui","Pardis Pashakhanloo","RAJDEEP MUKHERJEE","Myeongsoo Kim","Sachit Kuhar"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"CORVUS: Context Optimization and Reduction Via Underlying Synchronization for LLM Coding Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.22711","noteUrl":"/review?paper=review-neurips-2026-149092","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149094","title":"Capturing In-Context Learning Dynamics with Task Operators","authors":["Guangzhi Xiong","Zhenghao He","Bohan Liu","Sanchit Sinha","Wenqian Ye","Aidong Zhang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Capturing In-Context Learning Dynamics with Task Operators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01054","noteUrl":"/review?paper=review-neurips-2026-149094","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149097","title":"Beyond Parameter Arithmetic: Sparse Complementary Fusion for Distribution-Aware Model Merging","authors":["Weihong Lin","Lin Sun","Qilong Shi","Aomufei Yuan","Yuxuan Tian","Zhengyang Wang","Guangxiang Zhao","Xiangzheng Zhang","Tong Yang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Beyond Parameter Arithmetic: Sparse Complementary Fusion for Distribution-Aware Model Merging","plainExplanation":"The fusion rule can concentrate a harmless common logit shift on one coordinate, creating a much larger probability change. The report separates that failure from valid conditional perturbation bounds and states the extra energy, spectral-gap, or geometric conditions needed for the other guarantees.","primaryUrl":"https://arxiv.org/html/2602.11717","noteUrl":"/review?paper=review-neurips-2026-149097","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","invalid_inference","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149098","title":"Binding Multiple Modalities via Multimodal Wasserstein Barycenter","authors":["Xiaole Tang","Jiayi Xu","Xiang Gu","Yan Yang","Jian Sun"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Binding Multiple Modalities via Multimodal Wasserstein Barycenter","plainExplanation":"Independent source and barycenter samples do not form an optimal transport coupling. Canceling the potentials loses the coupling needed by the claimed duality.","primaryUrl":"https://arxiv.org/html/2609.33800","noteUrl":"/review?paper=review-neurips-2026-149098","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149101","title":"Information-Geometric Forward Policy Training in GFlowNets","authors":["Yordan Raykov","Rodrigo Veiga"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The expected-log-reward gap is not a divergence","plainExplanation":"The expected-log-reward gap is not a divergence.","primaryUrl":"https://arxiv.org/html/2608.03967","noteUrl":"/review?paper=review-neurips-2026-149101","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149102","title":"Spectral Alignment in Forward–Backward Representations via Temporal Abstraction","authors":["Seyed Mahdi Basiri Azad","Jasper Hoffmann","Iman Nematollahi","Hao Zhu","Abhinav Valada","Joschka Boedecker"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"A k-dependent condition number does not give a uniform fixed-base rate","plainExplanation":"The action-repetition eigenvalue bound is interpreted as decay at fixed base \\(|\\lambda_{d+1}(P_A)|^k\\), although its stated multiplier includes a condition number depending on \\(k\\). Use an action resetting to state two and another acting as \\(\\rho\\) times a swap of states one/two plus \\((1-\\rho)\\) times reset to state three, with \\(\\rho=1/4\\).","primaryUrl":"https://arxiv.org/html/2603.20103","noteUrl":"/review?paper=review-neurips-2026-149102","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149103","title":"Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata","authors":["Meihua Dang","Stefano Ermon"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Automaton-path exactness is not membership conditioning for an ambiguous NFA","plainExplanation":"Automaton-path exactness is not membership conditioning for an ambiguous NFA.","primaryUrl":"https://arxiv.org/html/2607.07026","noteUrl":"/review?paper=review-neurips-2026-149103","collection":"neurips-2026","severity":"repairable","issueTypes":["implementation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149104","title":"Training-Free Generative Sampling via Moment-Matched Score Smoothing","authors":["ZHENYU YAO","Daniel Paulin"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Training-Free Generative Sampling via Moment-Matched Score Smoothing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14276","noteUrl":"/review?paper=review-neurips-2026-149104","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149105","title":"Narrative Paradoxes in LLM Inference: Shaping Reasoning Trajectories toward Misalignment","authors":["Haoming Yang","Ke Ma","Yangbangyan Jiang","Ligong Zhang","Xiaojun Jia","Longtao Huang","Yingfei Sun","Qianqian Xu","Qingming Huang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Narrative Paradoxes in LLM Inference: Shaping Reasoning Trajectories toward Misalignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2610.16312v1","noteUrl":"/review?paper=review-neurips-2026-149105","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149106","title":"Muon Dynamics as a Spectral Wasserstein Flow","authors":["Gabriel Peyré"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Muon Dynamics as a Spectral Wasserstein Flow","plainExplanation":"No issue was found in the checked claims. The unrestricted symmetrized covariance formula has been replaced by the correct ordered expression. Its remaining use has the commuting hypothesis that the earlier counterexample violated.","primaryUrl":"https://arxiv.org/html/2604.04891","noteUrl":"/review?paper=review-neurips-2026-149106","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149107","title":"Multigroup Fairness and Omniprediction: Separations and Equivalences","authors":["Sílvia Casacuberta","Parikshit Gopalan","Varun Kanade","Omer Reingold","Konstantinos Stavropoulos","Pranay Tankala"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Multigroup Fairness and Omniprediction: Separations and Equivalences","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07374","noteUrl":"/review?paper=review-neurips-2026-149107","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149108","title":"Concept frustration: Aligning human concepts and machine representations","authors":["Christopher R. S. Banerji","Enrico Parisini","Christopher J Soelistyo","Ahab Isaac","Alessandro Barp"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Dependence through a multicomponent concept is not transitive","plainExplanation":"Dependence through a multicomponent concept is not transitive. The sampled depth mixture has positive marginal concept covariance.","primaryUrl":"https://arxiv.org/html/2603.29654","noteUrl":"/review?paper=review-neurips-2026-149108","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149109","title":"ExtraVAR: Stage-Aware RoPE Remapping for Resolution Extrapolation in Visual Autoregressive Models","authors":["Feihong Yan","Shaoyu Liu","Haixuan Wang","Shuai Lu","Weijie Ma","Linfeng Zhang","Huiqi Li","Xiangyang Ji"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ExtraVAR: Stage-Aware RoPE Remapping for Resolution Extrapolation in Visual Autoregressive Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10045","noteUrl":"/review?paper=review-neurips-2026-149109","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149112","title":"Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding","authors":["Zihan Zhou","Libo Zhu","Jue Gong","Zhiyi Zhou","Jiezhang Cao","Yong Guo","Yulun Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.29845","noteUrl":"/review?paper=review-neurips-2026-149112","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149114","title":"Stage-wise Attention-Guided Region Sequencing for Adversarial Attacks on Large Vision-Language Models","authors":["Jaehyun Kwak","Nam Cao","Boryeong Cho","Segyu Lee","Sumyeong Ahn","Se-Young Yun"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Stage-wise Attention-Guided Region Sequencing for Adversarial Attacks on Large Vision-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.04356","noteUrl":"/review?paper=review-neurips-2026-149114","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149115","title":"Prediction-Powered Inference Across Many Tasks for AI Evaluations and Social Science Research","authors":["Nicolas Emmenegger","Ellery Stahler","Chara Podimata"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Prediction-Powered Inference Across Many Tasks for AI Evaluations and Social Science Research","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29249","noteUrl":"/review?paper=review-neurips-2026-149115","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149116","title":"APPSolver: Adaptive Patch Partitioning for Point-Wise Ship Flow Prediction on Unstructured Meshes","authors":["Wenhua Huo","Fenglei Han","Wangyuan Zhao","Xiao Peng","Chunhui Wang","Jialin Wu","JiaYi Han"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"APPSolver: Adaptive Patch Partitioning for Point-Wise Ship Flow Prediction on Unstructured Meshes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.29355","noteUrl":"/review?paper=review-neurips-2026-149116","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149118","title":"MapPFN: Learning Causal Perturbation Maps in Context","authors":["Marvin Sextro","Weronika Kłos","Gabriel Dernbach"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"MapPFN: Learning Causal Perturbation Maps in Context","plainExplanation":"The displayed entropy-regularized transport objective can be negative before its square root. The stated debiasing step does not resolve that domain error.","primaryUrl":"https://arxiv.org/html/2601.21092","noteUrl":"/review?paper=review-neurips-2026-149118","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149121","title":"Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs","authors":["Elizabeth Mieczkowski","Alexander Ku","Tiwalayo Eisape","Dilip Arumugam","John Matters","Katie Collins","Ilia Sucholutsky","Tom Griffiths"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06320","noteUrl":"/review?paper=review-neurips-2026-149121","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149124","title":"FaithSieve: Fine-Grained Evaluation of Math Proofs with Faithful Formal Evidence","authors":["Ziyu Wang","Qiming Dai","Yishan Wu","Zaiwen Wen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FaithSieve: Fine-Grained Evaluation of Math Proofs with Faithful Formal Evidence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.26310","noteUrl":"/review?paper=review-neurips-2026-149124","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149126","title":"Tadpole: Autoencoders as Foundation Models for 3D PDEs with Online Learning","authors":["Qiang Liu","Felix Koehler","Benjamin Holzschuh","Nils Thuerey"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Tadpole: Autoencoders as Foundation Models for 3D PDEs with Online Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15284","noteUrl":"/review?paper=review-neurips-2026-149126","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149128","title":"AMPS: Adaptive Modality Preference Steering via Functional Entropy","authors":["Zihan Huang","Xintong Li","Rohan Surana","Tong Yu","Rui Wang","Julian McAuley","Jingbo Shang","Junda Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Gaussian log-Sobolev bound omits the perturbation scale","plainExplanation":"The Gaussian log-Sobolev bound omits the perturbation scale. The equation and algorithm reverse the cross-entropy arguments. The normalization sums over draws instead of modalities.","primaryUrl":"https://arxiv.org/html/2602.12533","noteUrl":"/review?paper=review-neurips-2026-149128","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149131","title":"Online Learning on Hidden-Convex Losses via Algorithmic Equivalence: Optimal Regret, Geometric Barrier, and Bandit Feedback","authors":["Anas Barakat","Andreas Kontogiannis","Vasilis Pollatos","Ioannis Panageas","Antonios Varvitsiotis"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The regret bound has the wrong gradient scale","plainExplanation":"The bound scales as $G_F^{3/2}$, although even first-round regret can scale linearly with $G_F$. The proof absorbs quadratic gradient terms into cubic ones without the needed restriction. Its final optimization also omits a factor of two. The explicit inequality fails under the permitted gradient scales. The corrected algebra retains a $\\sqrt T$ rate, conditional on the preceding geometric comparison proposition. This calculation does not establish the paper's separate lower-bound construction or bandit theorem.","primaryUrl":"https://arxiv.org/html/2605.26373","noteUrl":"/review?paper=review-neurips-2026-149131","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149134","title":"Winning Lottery Tickets in Neural Networks via a Quantum-Inspired Classical Algorithm","authors":["Natsuto Isogai","Hayata Yamasaki","Sho Sonoda","Mio Murao"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Winning Lottery Tickets in Neural Networks via a Quantum-Inspired Classical Algorithm","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13979","noteUrl":"/review?paper=review-neurips-2026-149134","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149137","title":"FLARE: Verifying MILP Reformulations with LLM-Based Formal Proof Synthesis","authors":["Henry Robbins","Connor Lawless","Madeleine Udell","Ellen Vitercik"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FLARE: Verifying MILP Reformulations with LLM-Based Formal Proof Synthesis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.25220v1","noteUrl":"/review?paper=review-neurips-2026-149137","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149138","title":"The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations","authors":["Rania Elbadry","Ahmed Heakl","Fan Zhang","Dani Bouch","Yuxia Wang","Preslav Nakov","Zhuohan Xie"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Orthogonal probe axes do not imply impossibility for every nuisance-score detector","plainExplanation":"Orthogonal probe axes do not imply impossibility for every nuisance-score detector.","primaryUrl":"https://arxiv.org/html/2605.09195","noteUrl":"/review?paper=review-neurips-2026-149138","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149141","title":"PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion Language Models","authors":["Xiaocheng Lu","Shuhan Guo","Ziyue Ma","Jie ZHANG","Jian Liu","Jingcai Guo","Haoxuan Che","Song Guo"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Cursor consistency and finite oracle horizons","plainExplanation":"A boundary position can exceed the ordinary threshold while failing the stricter commit test. Equation (5) then advances past a position that remains masked, breaking the stated prefix invariant.","primaryUrl":"https://arxiv.org/html/2609.26249","noteUrl":"/review?paper=review-neurips-2026-149141","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","definition_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149142","title":"LARK: Learnability-Grounded Trajectory Selection for Efficient Reasoning Distillation","authors":["Tianrun Yu","Kaixiang Zhao","Chih-Chun Chen","Amanda Hughes","Taylor Killian","Fenglong Ma","Weitong Zhang","Porter Jenkins"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The main proxy mixes summed residuals with an average loss","plainExplanation":"The main-text proxy sums token residual squares but divides by average token loss. Relative to the intended estimator in the algorithm and appendix, this introduces an unwanted trajectory-length factor. The algorithm and appendix already specify the corrected proxy. This local normalization typo does not establish a defective implemented selector.","primaryUrl":"https://arxiv.org/html/2605.30651","noteUrl":"/review?paper=review-neurips-2026-149142","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149145","title":"How I learned to stop worrying and love StopGrads: Stationarity, Convergence, and a case study on Flow Map Learning","authors":["Mark Goldstein","Max Shen","Zichu Wang","Aahlad Manas Puli","Rajesh Ranganath"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A compactly supported cutoff cannot have an everywhere nonnegative derivative","plainExplanation":"The cutoff’s claimed globally nonnegative derivative is impossible: it drops from value \\(R\\) to zero over \\([R,2R]\\), so its derivative must be negative somewhere. The false sign restriction is unnecessary for the next limit argument. The positive limiting variation and main result are not refuted by this local error.","primaryUrl":"https://arxiv.org/html/2609.16222","noteUrl":"/review?paper=review-neurips-2026-149145","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149146","title":"Point Tracking Improves World Action Models","authors":["Jiarui Guan","Wenshuai Zhao","Yue Pei","Ziliang Chen","Arno Solin","Juho Kannala"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Point Tracking Improves World Action Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23856","noteUrl":"/review?paper=review-neurips-2026-149146","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149148","title":"RelationVGGT : Visual Geometry Transformers for 3D Spatial Relation Segmentation","authors":["Minsu Kim","Jaesung Choe","Jiwoo Lee","Frank Wang","Seon Joo Kim"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"RelationVGGT : Visual Geometry Transformers for 3D Spatial Relation Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.00970","noteUrl":"/review?paper=review-neurips-2026-149148","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149150","title":"Nearly Optimal Robust Covariance and Scatter Matrix Estimation Beyond Gaussians","authors":["Gleb Novikov"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The spherical reference tensor is normalized at the wrong radius","plainExplanation":"For \\(i\\ne j\\), \\(\\mathbb E[U_i^2U_j^2]=1/[d(d+2)]\\), giving that entry in the printed reference tensor. A radius-\\(\\sqrt d\\) sphere instead gives \\(d/(d+2)\\), so the difference tends to one. Also \\(\\langle I/\\sqrt d,UU^\\top-I\\rangle^2=(d-1)^2/d\\), whereas the correctly scaled radius-\\(\\sqrt d\\) reference has zero trace fluctuation. Negligible clipping tails cannot supply the missing dimension factors. The literal moment comparison fails. This scaling correction removes this counterexample but does not validate every remaining tensor estimate.","primaryUrl":"https://arxiv.org/html/2502.06564","noteUrl":"/review?paper=review-neurips-2026-149150","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149152","title":"AT-SKM-Net: An Accelerated Trainable Sampling Kaczmarz-Motzkin Framework for Linear Hard-Constraint Feasibility on Dynamic Graphs","authors":["Xiaochen Zhang","Haoyu Zhu","Yao Zhang","Qingchun Hou"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"AT-SKM-Net: An Accelerated Trainable Sampling Kaczmarz-Motzkin Framework for Linear Hard-Constraint Feasibility on Dynamic Graphs","plainExplanation":"Active-set stability does not make every active residual positive. The claimed one-step contraction can be zero while half the squared error remains.","primaryUrl":"https://arxiv.org/html/2609.30088","noteUrl":"/review?paper=review-neurips-2026-149152","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference"],"classifications":["major"]},{"slug":"review-neurips-2026-149153","title":"Argument Graph Uncertainty: Quantifying Uncertainty from the Logical Structure of Reasoning Chains","authors":["Simon Baur","Sumire Honda","Wojciech Samek","Jackie Ma"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Rank-space residualization removes one linear projection, not all monotone dependence","plainExplanation":"Rank-space residualization removes one linear projection, not all monotone dependence.","primaryUrl":"https://arxiv.org/pdf/2610.20328v1","noteUrl":"/review?paper=review-neurips-2026-149153","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149154","title":"LARGO: Low-Rank Hypernetwork for Handling Missing Modalities","authors":["Niels Vyncke","Pooya Ashtari","Aleksandra Pizurica"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LARGO: Low-Rank Hypernetwork for Handling Missing Modalities","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06086","noteUrl":"/review?paper=review-neurips-2026-149154","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149157","title":"Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads","authors":["Aryo Gema","Beatrice Alex","Pasquale Minervini"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Increasing the correct-token logit need not reduce cross-entropy","plainExplanation":"Cross-entropy depends on competing logits as well as the correct-token logit. A head can increase the correct logit while increasing another logit more. The direct-path assumption does not remove that competing contribution. This invalidates the universal loss-gradient and ablation implications of Corollary 3. It does not invalidate the definition of the score or establish an error in the empirical ranking of retrieval heads.","primaryUrl":"https://arxiv.org/html/2607.01002","noteUrl":"/review?paper=review-neurips-2026-149157","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","interpretation_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149160","title":"Less Language, More Latents: Annotation-Efficient VLAs for Driving","authors":["Alexey Zakharov","Kemal Oksuz","Puneet Dokania"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Less Language, More Latents: Annotation-Efficient VLAs for Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.27747","noteUrl":"/review?paper=review-neurips-2026-149160","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149163","title":"Can Large Language Models Develop Gambling Addiction?","authors":["Seungpil Lee","Donghyeon Shin","Yunjeong Lee","Sundong Kim"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Can Large Language Models Develop Gambling Addiction?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2509.22818","noteUrl":"/review?paper=review-neurips-2026-149163","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149164","title":"HYDRA: Representation Harmonized Tokenization for Multimodal Generation and Understanding","authors":["Xuerui Qiu","Yutao Cui","Guozhen Zhang","Junzhe Li","Yaqi Zhao","Xiao Zhang","Yang Li","Songtao Liu","Miles Yang","Yu Shi","Zhao Zhong","Liefeng Bo"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"HYDRA: Representation Harmonized Tokenization for Multimodal Generation and Understanding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.15228","noteUrl":"/review?paper=review-neurips-2026-149164","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149165","title":"Reason in the Words You Speak: Idiolectal Paraphrasing Off-Policy Traces for Reasoning Distillation in VideoLLMs","authors":["Ji Soo Lee","Jinyoung Park","Seohyun Lee","Jongha Kim","Joonmyung Choi","Jinsung Yoon","Hyunwoo J. Kim"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"PPO clipping is one-sided for each advantage sign","plainExplanation":"For \\(\\min(rA,\\operatorname{clip}(r,1-\\varepsilon,1+\\varepsilon)A)\\), positive \\(A\\) clips only the upper side, and negative \\(A\\) clips only the lower side. With positive advantage, \\(r=0.5\\) and \\(\\varepsilon=0.2\\), the objective remains \\(rA\\) and its ratio derivative is \\(A\\), despite lying outside the interval. Small absolute token probability also need not imply a small ratio when both policies agree. The universal gradient-clipping description is false. Measured clipping rates and empirical mismatch reduction can remain valid.","primaryUrl":"https://arxiv.org/html/2608.26684","noteUrl":"/review?paper=review-neurips-2026-149165","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149174","title":"MJ: Multi-turn LLM Jailbreaking via Decomposed Credit Assignment","authors":["Junyoung Park","Namgyu Park","Sechan Lee","Yoon-Chan Jhi","Jihoon Cho","Sangdon Park"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MJ: Multi-turn LLM Jailbreaking via Decomposed Credit Assignment","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.11070","noteUrl":"/review?paper=review-neurips-2026-149174","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149175","title":"Bidirectional Information Flow (BIF) - A Sample Efficient Hierarchical Gaussian Process for Bayesian Optimization","authors":["Juan D. Guerra","Thomas Garbay","Numa Dancause","Guillaume Lajoie","Marco Bonizzato"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Bidirectional Information Flow (BIF) - A Sample Efficient Hierarchical Gaussian Process for Bayesian Optimization","plainExplanation":"Posterior variance is prior variance minus a data-dependent quadratic term. The proof bounds the quadratic-term difference but drops the separate discrepancy $k_{t-1}(x,x)-k^*(x,x)$. With no observations, its proposed upper bound is therefore zero regardless of the prior mismatch. Its kernel-entry estimates also need to follow the stated diagonal bound $\\kappa^2$.","primaryUrl":"https://arxiv.org/html/2505.11294","noteUrl":"/review?paper=review-neurips-2026-149175","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149177","title":"Beyond Confidence: Rethinking Self-Assessments for Performance Prediction in LLMs","authors":["Sree Bhattacharyya","Samarth Khanna","Leona Chen","Lucas Craig","Tharun Dilliraj","James Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The binned Brier decomposition omits within-bin forecast variation","plainExplanation":"The binned Brier decomposition omits within-bin forecast variation.","primaryUrl":"https://arxiv.org/html/2605.07806","noteUrl":"/review?paper=review-neurips-2026-149177","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149178","title":"The Constitutional Coverage Trilemma in AI Governance","authors":["Natalija Mitic","Soona S O.","Mamadou Selly Ly","Moustapha Cisse"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Concavity does not uniformly increase the regret floor","plainExplanation":"Concavity controls how slopes change across the utility interval. It does not ensure that every utility difference becomes larger. The proposed extension fails even for a smooth, strictly increasing concave function normalized to have $f(0)=0$ and $f(1)=1$. The counterexample concerns the concave-utility robustness claim. It does not refute the original linear-welfare theorem. Concavity alone also gives no general conclusion that a difference between group regrets must widen.","primaryUrl":"https://arxiv.org/html/2609.01275","noteUrl":"/review?paper=review-neurips-2026-149178","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149180","title":"Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models","authors":["Yanting Miao","Yutao Sun","Dexin Wang","Pascal Poupart","Mengyu Zhou","Lei Lv","Qi Zhao","Li Wang","Hao Li","xiaoxi jiang","Guanjun Jiang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12374","noteUrl":"/review?paper=review-neurips-2026-149180","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149182","title":"AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion","authors":["Gal Wertheizer","Rom Himelstein","Tomer Peretz","Avi Mendelson"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Two-prompt centered CKA has zero repulsion gradient","plainExplanation":"With two prompts, every nonzero centered Gram matrix has the same direction. Their CKA is identically one and supplies no repulsion gradient. The two-prompt boundary is affected; four-prompt experiments and auxiliary losses are not refuted.","primaryUrl":"https://arxiv.org/html/2609.32602","noteUrl":"/review?paper=review-neurips-2026-149182","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149185","title":"LiFT: Lifted Inter-slice Feature Trajectories for 3D Image Generation from 2D Generators","authors":["Xinhe Zhang","Yuyang Zhang","Pengfei Jin","Arnau Marin-Llobet","Na Li","Quanzheng Li"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"LiFT: Lifted Inter-slice Feature Trajectories for 3D Image Generation from 2D Generators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19060","noteUrl":"/review?paper=review-neurips-2026-149185","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149189","title":"Private Adaptive Covariance Estimation via Gaussian Graphical Models","authors":["Cecilia Ferrando","Miguel Fuentes","Brett Mullins","Cameron Musco","Daniel Sheldon"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Completing the Gaussian likelihood square leaves a factor of one half","plainExplanation":"The grouped likelihood equality omits the Gaussian quadratic’s factor \\(1/2\\). This is a nonconstant scaling difference, not an additive constant, although the unpenalized minimizer is unchanged. Positive global scaling preserves the unpenalized minimizer. The error is the claimed equality up to an additive constant, rather than the reconstruction argmin.","primaryUrl":"https://arxiv.org/html/2605.24295","noteUrl":"/review?paper=review-neurips-2026-149189","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149190","title":"CIG: Exploration via Conditional Information Gain","authors":["Tim Joseph","Marcus Fechner","Philipp Stegmaier","Karam Daaboul","Marius Zöllner"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Training residuals do not upper-bound transition noise variance","plainExplanation":"Training residuals do not upper-bound transition noise variance. The logarithmic kernel bound needs a real-domain condition.","primaryUrl":"https://arxiv.org/html/2605.20878","noteUrl":"/review?paper=review-neurips-2026-149190","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149193","title":"Hierarchical Semantic Tree Anchoring for CLIP-Based Class-Incremental Learning","authors":["Tao Hu","Lan Li","Zhenhao Wen","Da-Wei Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Hierarchical Semantic Tree Anchoring for CLIP-Based Class-Incremental Learning","plainExplanation":"No issue was found in the checked claims. The former counterexample changes the input features as well as the transformation. It does not refute the current, explicitly local statement about the transformation’s response to stored features.","primaryUrl":"https://arxiv.org/html/2511.15633","noteUrl":"/review?paper=review-neurips-2026-149193","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149197","title":"Delve into the Applicability of Advanced Optimizers for Multi-Task Learning","authors":["Zhipeng Zhou","Linxiao Cao","Yiming Cao","Pengcheng Wu","Peilin Zhao","Chunyan Miao"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Delve into the Applicability of Advanced Optimizers for Multi-Task Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.08939","noteUrl":"/review?paper=review-neurips-2026-149197","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149200","title":"PhGPO: Pheromone-Guided Policy Optimization for Long-Horizon Tool Planning","authors":["Yu Li","Guangfeng Cai","shengtian yang","Han Luo","Shuo Han","Xu He","Dong Li","Lei Feng"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Unvisited edges never execute the evaporation update","plainExplanation":"An edge never revisited after initialization never executes \\(\\tau_e\\leftarrow(1-\\rho)\\tau_e+\\alpha q\\), so its absolute pheromone stays fixed. Reinforcing other edges can reduce its relative normalized importance, but that is not evaporation of the untouched edge. This is an update-scope specification gap; actual implementation and relative visual fading remain undetermined.","primaryUrl":"https://arxiv.org/html/2602.13691","noteUrl":"/review?paper=review-neurips-2026-149200","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149201","title":"Integrated Imputation-Classification for Supervised Learning with Missing Data","authors":["Yue Liu","Ben Liang","Ali Tizghadam","Ilijc Albanese"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Integrated Imputation-Classification for Supervised Learning with Missing Data","plainExplanation":"Correct pooled imputation statistics do not determine the missing-feature posterior for each mask. A globally aligned generator can therefore make a wrong mask-specific prediction.","primaryUrl":"https://arxiv.org/html/2610.04273","noteUrl":"/review?paper=review-neurips-2026-149201","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-149202","title":"DiffScore: Text Evaluation Beyond Autoregressive Likelihood","authors":["Wen Lai","Yingli Shen","Dingnan Jin","Qing Cui","Jun Zhou","Maosong Sun","Alexander Fraser"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The default score divides a token mean by sequence length again","plainExplanation":"The default score divides a token mean by sequence length again.","primaryUrl":"https://arxiv.org/html/2605.11601","noteUrl":"/review?paper=review-neurips-2026-149202","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149206","title":"Rethinking Expressivity and Efficiency in Test-Time Training","authors":["Zeyun Zhong","Joya Chen","Manuel Martin","Frederik Diederichs","Jürgen Gall","Jürgen Beyerer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rethinking Expressivity and Efficiency in Test-Time Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.21308","noteUrl":"/review?paper=review-neurips-2026-149206","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149208","title":"Closing the Reflection Gap: A Free Calibration Bonus for Agentic RL","authors":["Yinglun Zhu"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The utility condition for informative flags","plainExplanation":"Positive association between flags and errors does not imply that flagged-answer correctness is below the chosen abstention credit. That utility threshold, not informativeness alone, determines the sign.","primaryUrl":"https://arxiv.org/html/2606.14211","noteUrl":"/review?paper=review-neurips-2026-149208","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149209","title":"Steering Away from Memorization: Reachability-Constrained Reinforcement Learning for Text-to-Image Diffusion","authors":["Sathwik Karnik","Juyeop Kim","Sanmi Koyejo","Jong-Seok Lee","Somil Bansal"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The discounted safety target is not a hard minimum-margin certificate","plainExplanation":"The discounted target can be positive before a future negative margin. Its sign therefore does not by itself certify that the whole remaining trajectory avoids the failure set.","primaryUrl":"https://arxiv.org/html/2603.00140","noteUrl":"/review?paper=review-neurips-2026-149209","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["invalid_inference"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149210","title":"Automated Kernel Discovery Towards Understanding High-dimensional Bayesian Optimization","authors":["Taeyoung Yun","Woocheol Shin","Inhyuck Song","Jaewoo Lee","Jinkyoo Park"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"An unrestricted linear-kernel coefficient can violate covariance positivity","plainExplanation":"An unrestricted linear-kernel coefficient can violate covariance positivity. The spherical covariance component exceeds its diagonal near zero.","primaryUrl":"https://arxiv.org/html/2605.20249","noteUrl":"/review?paper=review-neurips-2026-149210","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149211","title":"Tabular Foundation Model for Generative Modelling","authors":["Xiangjian Jiang","Mingxuan Liu","Nikola Simidjievski","Tassilo Klein","Mateja Jamnik"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Tabular Foundation Model for Generative Modelling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09424","noteUrl":"/review?paper=review-neurips-2026-149211","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149214","title":"Two-Level Softmax Sampling Done Right: Correcting Bias from Size Imbalance and Dispersion","authors":["Walid Bendada","Guillaume Salha-Galvan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Two-Level Softmax Sampling Done Right: Correcting Bias from Size Imbalance and Dispersion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10483","noteUrl":"/review?paper=review-neurips-2026-149214","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149215","title":"PAPO-VLA: Planning-Aware Policy Optimization for Vision-Language-Action Models","authors":["Peizheng Guo","Jingyao Wang","Changwen Zheng","Wenwen Qiang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"A constant positive trajectory reward cannot change within-trajectory TopK","plainExplanation":"A constant positive trajectory reward cannot change within-trajectory TopK. The harmonic planning score needs a zero-effect convention.","primaryUrl":"https://arxiv.org/html/2605.19580v1","noteUrl":"/review?paper=review-neurips-2026-149215","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149217","title":"Optimal Transport Reweighting for Robust Learning under Spurious Correlations and Label Noise","authors":["Sungho Jo","Seonghwi Kim","Wonsang Yun","Minwoo Chae"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Optimal Transport Reweighting for Robust Learning under Spurious Correlations and Label Noise","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01028","noteUrl":"/review?paper=review-neurips-2026-149217","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149219","title":"S2T-RLHF: Hierarchical Credit Assignment for Stable Preference-Based RLHF","authors":["Wei Chen","Guanghui Zhu","Yafei Li","Limin Wang","Yihua Huang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Discounting gives the opposite per-token variance direction in the zero-baseline case","plainExplanation":"Discounting gives the opposite per-token variance direction in the zero-baseline case. Bounded simplex modulation does not by itself attenuate global advantage dependence.","primaryUrl":"https://arxiv.org/html/2607.18258","noteUrl":"/review?paper=review-neurips-2026-149219","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149220","title":"Exposing the Illusion of Erasure in Knowledge Editing for LLMs","authors":["Advik Basani","Anshuman Chhabra"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stated alignment sign predicts a decrease rather than an increase","plainExplanation":"The lemma establishes only that the update vanishes on the perpendicular subspace. With \\(W=I\\) and \\(\\Delta=-e_1e_1^\\top\\), a rank-one edit annihilates the response to \\(e_1\\), which has no complementary projection. More strongly, setting \\(W_{11}=a\\) and \\(\\Delta_{11}=-a\\) produces identical edited weights for every \\(a\\), losing the original value. Unchanged dimensions need not encode that fact, and decomposition itself does not move information into them. The algebraic lemma and empirical extraction results survive. Low rank alone does not imply semantic preservation or displacement.","primaryUrl":"https://arxiv.org/html/2606.23276","noteUrl":"/review?paper=review-neurips-2026-149220","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149222","title":"LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models","authors":["Songtao Wei","Yi Li","Zhikai Li","Xu Hu","Yuede Ji","Guanpeng Li","Feng Chen","Carl Yang","Zhichun Guo","Bingzhe Li"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09806","noteUrl":"/review?paper=review-neurips-2026-149222","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149223","title":"LookWhen? Fast Video Recognition by Learning When, Where, and What to Compute","authors":["Ali Salamatian","Anthony Fuller","Pritam Sarkar","James Green","Leonid Sigal","Evan Shelhamer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LookWhen? Fast Video Recognition by Learning When, Where, and What to Compute","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06809","noteUrl":"/review?paper=review-neurips-2026-149223","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149224","title":"Knowing When to Ask: Segment-Level Credit Assignment for LLM Tool Use","authors":["Abhijit Kumar","Zoey WU","Mohit Suley"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Knowing When to Ask: Segment-Level Credit Assignment for LLM Tool Use","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.27788","noteUrl":"/review?paper=review-neurips-2026-149224","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149225","title":"Decision-Focused Learning in MDPs: An Occupancy Measure Approach","authors":["Zihao Zhao","Ashwath K Karunakaram","Ali Eshragh","Yuexing Li","Kai Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Decision-Focused Learning in MDPs: An Occupancy Measure Approach","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.08384","noteUrl":"/review?paper=review-neurips-2026-149225","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149226","title":"Emergence of a Shared Canonical Object Frame from In-the-Wild Videos","authors":["Tom Fischer","Martin Sundermeyer","Adam Kortylewski","Eddy Ilg"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The continuous symmetry test and discrete metric act on opposite sides","plainExplanation":"The continuous symmetry test and discrete metric act on opposite sides.","primaryUrl":"https://arxiv.org/html/2606.30058","noteUrl":"/review?paper=review-neurips-2026-149226","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149232","title":"Flexible Flows for Biological Sequence Design","authors":["Yogesh Verma","Dani Korpela","Harri Lähdesmäki","Vikas Garg"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The guided rate factors use incompatible exponent conventions","plainExplanation":"The guided rate factors use incompatible exponent conventions.","primaryUrl":"https://arxiv.org/html/2606.10543","noteUrl":"/review?paper=review-neurips-2026-149232","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149233","title":"Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport","authors":["Malte Franke","Stefan P. Schmid","Žarko Ivković","Kjell Jorner","Andreas Krause"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The small-step kernel omits the lost stay probability","plainExplanation":"The small-step kernel omits the lost stay probability. The sampler drift does not preserve the constant-variance matched path. The Euler hazard step can exceed a Bernoulli probability of one.","primaryUrl":"https://arxiv.org/html/2606.07239","noteUrl":"/review?paper=review-neurips-2026-149233","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149234","title":"Universal Adaptive Proximal Gradient Methods via Gradient Mapping Accumulation","authors":["Zimeng Wang","Alp Yurtsever"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Universal Adaptive Proximal Gradient Methods via Gradient Mapping Accumulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05944","noteUrl":"/review?paper=review-neurips-2026-149234","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149237","title":"FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization","authors":["Quanjian Song","Yefeng Shen","Mengting Chen","Hao Sun","Jinsong Lan","Xiaoyong Zhu","Bo Zheng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15824","noteUrl":"/review?paper=review-neurips-2026-149237","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149240","title":"AudioCALM: Continuous Autoregressive Language Modeling for Universal Audio Generation","authors":["Huadai Liu","Kaicheng Luo","Wen Wang","Qian Chen","Bin Ma","Xiangang Li","Wei Xue"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"AudioCALM: Continuous Autoregressive Language Modeling for Universal Audio Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.23080","noteUrl":"/review?paper=review-neurips-2026-149240","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149241","title":"Differentially Private Model Merging","authors":["Qichuan Yin","Manzil Zaheer","Tian Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Differentially Private Model Merging","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.20985","noteUrl":"/review?paper=review-neurips-2026-149241","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149244","title":"ReCoVer: Resilient LLM Pre-Training System via Fault-Tolerant Collective and Versatile Workload","authors":["Ziyue (Alvin) Liu","Zhengyang Wang","Ruijie (Ray) Zhang","Avinash Maurya","Hui Zhou","Paul Hovland","Sheng Di","Franck Cappello","Bogdan Nicolae","Zheng Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ReCoVer: Resilient LLM Pre-Training System via Fault-Tolerant Collective and Versatile Workload","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11215","noteUrl":"/review?paper=review-neurips-2026-149244","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149247","title":"VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon","authors":["Yi Pan","Miao Pan","Qi Lu","Jiaming Huang","Man Zhang","Siteng Huang","Xin Li","Jie Zhang","Yongliang Shen","Xuhong Zhang","Wenqi Zhang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.01804","noteUrl":"/review?paper=review-neurips-2026-149247","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149249","title":"Autonomous Scientific Discovery via Iterative Meta-Reflection","authors":["Bingchen Zhao","Sara Beery","Oisin Mac Aodha"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Held-out significance does not categorically prevent false discoveries","plainExplanation":"The acceptance cutoffs still admit false discoveries with positive probability, even for a prespecified null claim and a single held-out test.","primaryUrl":"https://arxiv.org/html/2607.01131","noteUrl":"/review?paper=review-neurips-2026-149249","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149250","title":"Practical Adversarial Attacks on Stochastic Bandits via Fake Data Injection","authors":["Qirun Zeng","Eric He","Richard Hoffmann","Xuchuang Wang","Jinhang Zuo"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The cooldown calculation squares before checking the sign","plainExplanation":"The comparison requires \\(3\\sigma\\sqrt{\\log t/N}\\leq L-\\bar r_{\\mathrm{post}}\\), so \\(L-\\bar r_{\\mathrm{post}}\\) must first be nonnegative. The eventual reward-floor assumption does not ensure this for each partial batch. With genuine counts \\((14,1)\\) and means \\((2,1)\\), the first batch leaves mean \\(1.2\\), above \\(L\\approx0.42127\\); the printed schedule can even give a negative cooldown. Squaring hides the impossible sign condition. The periodic-suppression lemma has a local feasibility gap. Suppression with enough initial injection remains possible.","primaryUrl":"https://arxiv.org/html/2505.21938","noteUrl":"/review?paper=review-neurips-2026-149250","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149251","title":"Sequential Membership Inference Attacks","authors":["Thomas Michel","Debabrota Basu","Emilie Kaufmann"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The batch-mean likelihood ratio omits a factor of batch size","plainExplanation":"Theorem 3.1 defines the unscaled residual \\(N=\\overline X-\\mu\\), with null covariance \\(\\Sigma/n\\) and member covariance \\((n-1)\\Sigma/n^2\\). Subtracting the Gaussian precisions gives \\(n\\Sigma^{-1}/(n-1)\\).","primaryUrl":"https://arxiv.org/html/2602.16596","noteUrl":"/review?paper=review-neurips-2026-149251","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","incorrect_bound","interpretation_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149252","title":"One Algorithm, Two Goals: Dual Scoring for Parameter and Data Selection in LLM Fine-Tuning","authors":["Xinrui Chen","Liu Yang","Ou Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"One Algorithm, Two Goals: Dual Scoring for Parameter and Data Selection in LLM Fine-Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06166","noteUrl":"/review?paper=review-neurips-2026-149252","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149254","title":"Sequential Solution Concepts in Cooperative Games with Generalized Characteristic Functions","authors":["Ashwin Goyal","Drashthi Doshi","Swaprava Nath"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The revised example is valid; the prefix notation still needs alignment","plainExplanation":"With the stated strict relation, a full-length sequence has no extension, so its carrier game is identically zero. Such games cannot form the asserted basis.","primaryUrl":"https://arxiv.org/abs/2510.11255v5","noteUrl":"/review?paper=review-neurips-2026-149254","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149255","title":"MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding","authors":["Zexuan Chen","Sichao Liu","Runhao Lu","Huichao Qi","Alexandra Woolgar","Xi V Wang","Lihui Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24523","noteUrl":"/review?paper=review-neurips-2026-149255","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149257","title":"DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images","authors":["Zeng Tao","Ying Jiang","Yunuo Chen","Tianyi Xie","Huamin Wang","Ying Nian Wu","Yin Yang","Chenfanfu Jiang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.16502","noteUrl":"/review?paper=review-neurips-2026-149257","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149260","title":"STILL: Selecting Tokens for Intra-Layer Hybrid Attention to Linearize LLMs","authors":["Weikang Meng","Liangyu Huo","Yadan Luo","Jiawen Guan","Jingyi Zhang","Yingjian Li","Zheng Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The local ratio does not determine expected global self-attention rank","plainExplanation":"Given self mass \\(S\\), local nonself mass \\(L\\), and iid outside mass \\(X\\), global expected self probability is \\(\\mathbb E[S/(S+L+X)]\\). Local ratio \\(S/(S+L)\\) does not determine it. With \\(X\\) equally likely \\(0.1\\) or \\(100\\), local masses \\((S,L)=(1,0.1)\\) outrank \\((900,100)\\), but their expected global probabilities have the reverse order. The unconditional strict rank-preservation statement fails. A suitable concentrated asymptotic result and empirical utility remain possible.","primaryUrl":"https://arxiv.org/html/2602.02180","noteUrl":"/review?paper=review-neurips-2026-149260","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149261","title":"Debiasing Random Oblique Projections for Subsampled OLS and Fast CUR in High Dimensions","authors":["Chengmei Niu","Sachin Garg","Michal Derezinski","Zhenyu Liao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Debiasing Random Oblique Projections for Subsampled OLS and Fast CUR in High Dimensions","plainExplanation":"The main variance concern is an equality obtained from upper bounds on cross terms. The other findings distinguish impossible sampling probabilities and an incorrect concentration scale from a repairable auxiliary matrix identity and two minor notation corrections.","primaryUrl":"https://arxiv.org/html/2605.24955","noteUrl":"/review?paper=review-neurips-2026-149261","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","false_statement","formula_error","incorrect_bound","missing_assumption","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149266","title":"The Surprising Effectiveness of Video Diffusion Models for Hand Motion Reconstruction","authors":["Yuxi Wang","Chengkai Jin","Yufei Liu","Wenqi Ouyang","Tianyi Wei","Zhiwei Zeng","Siyuan Huang","Zhiqi Shen","Xingang Pan"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"The Surprising Effectiveness of Video Diffusion Models for Hand Motion Reconstruction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30308","noteUrl":"/review?paper=review-neurips-2026-149266","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149267","title":"A Latent World-Action Model with Jointly Aligned Reasoning","authors":["Hao Luo","Wanpeng Zhang","Yicheng Feng","Sipeng Zheng","Haiweng Xu","Chaoyi Xu","Ziheng Xi","Yuhui Fu","Zongqing Lu"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"A Latent World-Action Model with Jointly Aligned Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.00078v1","noteUrl":"/review?paper=review-neurips-2026-149267","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149268","title":"DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models","authors":["Yuqiu Deng","Muyun Jiang","Zongpeng Zhang","Mingqing Xiao","Yansen Wang","Peiliang Gong","XINLIANG ZHOU","Yi Ding","Ziyu Jia","Cuntai Guan","Chenyu Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The within-subject class score is undefined for diagnosis-only subjects","plainExplanation":"The diagnostic dataset gives each subject one class, leaving the stated within-subject class-separation average undefined.","primaryUrl":"https://arxiv.org/html/2610.14497","noteUrl":"/review?paper=review-neurips-2026-149268","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149269","title":"Focus, Align, and Diffuse: Time-Series-Aware Keyframe Diffusion for Cardiac Dynamic Synthesis from Sparse Observations","authors":["Junkai Liu","Haofan Wu","Nay Aung","Joao A Lima","Steffen E Petersen","Le Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Focus, Align, and Diffuse: Time-Series-Aware Keyframe Diffusion for Cardiac Dynamic Synthesis from Sparse Observations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.19050v1","noteUrl":"/review?paper=review-neurips-2026-149269","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149270","title":"Training with (Swap) Regret Loss in a Single-Layer Self-Attention Model: A Case Study on the Probability Simplex","authors":["Chanwoo Park","Asuman Ozdaglar"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Training with (Swap) Regret Loss in a Single-Layer Self-Attention Model: A Case Study on the Probability Simplex","plainExplanation":"The centering operation scales vectors by an extra factor, while the stationarity proof removes a derivative that remains at a zero parameter. These are distinct from the row-versus-column convention error in the definition of swap regret; the missing sample derivative alone does not refute population stationarity.","primaryUrl":"https://arxiv.org/html/2607.23333","noteUrl":"/review?paper=review-neurips-2026-149270","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","false_statement","formula_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149272","title":"Inverse Reinforcement Learning with Just Classification and a Few Regressions","authors":["Lars van der Laan","Nathan Kallus","Aurelien Bibaut"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The likelihood objective exponentiates an already logged partition","plainExplanation":"The categorical conditional log likelihood is \\(\\mathbb E_\\pi[q-\\Xi(q)]\\). Replacing \\(\\Xi\\) by its exponential breaks invariance to a common logit shift: adding \\(c\\) changes the printed objective by \\(c-(e^c-1)\\sum_a e^{q_a}\\), although the policy is unchanged. Thus the printed optimization cannot have the claimed likelihood and shaping characterization. The printed objective is wrong; the intended log-partition likelihood theory is recoverable by the local correction.","primaryUrl":"https://arxiv.org/html/2509.21172","noteUrl":"/review?paper=review-neurips-2026-149272","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149273","title":"Unpaired Canonical Correlation Analysis","authors":["Nir Ben-Ari","Ronen Talmon","Uri Shaham"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Unpaired Canonical Correlation Analysis","plainExplanation":"The strict event depends on the data and can have probability zero even for centered, whitened samples. A valid concentration bound must retain the optimal signal size and dimension. These witnesses correct the supporting probability argument without refuting the separate conditional equivalence theorem.","primaryUrl":"https://arxiv.org/html/2610.09530","noteUrl":"/review?paper=review-neurips-2026-149273","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149274","title":"Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection","authors":["Manni Cui","Ziheng Qin","ZiAn Wang","Ruiqi Liu","Dianyuan Zou","Jianglan Wei","Han Zhou","Yu Liu","Xu Jingrui","Wenhao Wang","Zhenyu Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Channel covariance does not preserve the patch Gram matrix","plainExplanation":"Channel covariance loses patch identity and cannot reconstruct the claimed patch Gram matrix.","primaryUrl":"https://arxiv.org/html/2607.15321","noteUrl":"/review?paper=review-neurips-2026-149274","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149276","title":"Amortized Linear-time Exact Shapley Value for Product-Kernel Methods","authors":["Majid Mohammadi","Siu Lun (Alan) Chau","Krikamol Muandet"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Amortized Linear-time Exact Shapley Value for Product-Kernel Methods","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.16516","noteUrl":"/review?paper=review-neurips-2026-149276","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149277","title":"Behavioral Geometric Supervision Aligns Video Foundation Models with Human Social Perception","authors":["Kathy Garcia","Leyla Isik"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Behavioral Geometric Supervision Aligns Video Foundation Models with Human Social Perception","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.01502","noteUrl":"/review?paper=review-neurips-2026-149277","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149280","title":"Dynamic Model Merging Made Slim","authors":["Guodong DU","Wanyu Lin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Dynamic Model Merging Made Slim","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18904","noteUrl":"/review?paper=review-neurips-2026-149280","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149282","title":"Prompts to Proxies: Emulating Human Preferences via a Compact LLM Ensemble","authors":["Bingchen Wang","Zi-Yu Khoo","Jingtan Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Demographic matching does not prove a second distinct ensemble","plainExplanation":"Observation 2.5 says demographic learnability implies ensemble nonuniqueness. The following discussion seeks proxy agents with distinct preferences. Here distinct ensembles mean different probability measures on those preferences, after identifying reordered, duplicated or zero-weight list entries. Membership in a convex hull supplies a decomposition, but not a different decomposition. In particular, an extreme response can have only its trivial representation. The proof replaces one response by a convex decomposition without showing that the replacement changes the ensemble measure.","primaryUrl":"https://arxiv.org/html/2509.11311","noteUrl":"/review?paper=review-neurips-2026-149282","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149283","title":"PAGER: Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control","authors":["Jingxuan Wei","Xi Bai","Shan Liu","caijun jia","Zheng Sun","Xinglong Xu","Siyuan Li","Linzhuang Sun","Bihui Yu","Conghui He","Cheng Tan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"PAGER: Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15963","noteUrl":"/review?paper=review-neurips-2026-149283","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"lynx-early-exit-conditioning","title":"LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning","authors":["Ömer Faruk Akgül","Yusuf H Kalayci","Rajgopal Kannan","Willie Neiswanger","Viktor Prasanna"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Positive-only calibration does not bound early-exit error","plainExplanation":"Accepting most correct cues does not ensure that most accepted cues are correct. The printed rule also accepts more cues when its confidence parameter increases; the released code uses a different, both-label singleton rule.","primaryUrl":"https://arxiv.org/abs/2512.05325v1","collection":"neurips-2026","severity":"critical","issueTypes":["algorithm_specification","false_statement"],"classifications":["critical"]},{"slug":"review-neurips-2026-149285","title":"Neural Modal Decomposition: Architectural Priors from Observables","authors":["Juho Park","Kaushik Sengupta"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The anchor penalty is stronger than a gauge choice","plainExplanation":"Requiring both modal anchor factors to be positive real imposes restrictions that a gauge change cannot generally satisfy.","primaryUrl":"https://arxiv.org/html/2609.14402","noteUrl":"/review?paper=review-neurips-2026-149285","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149287","title":"Feedforward Novel View Synthesis for Heterogeneous Cameras","authors":["Meng Wei","Cheng Zhang","Boying Li","Yihang Chen","Jianmin Zheng","Hamid Rezatofighi","Jianfei Cai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Feedforward Novel View Synthesis for Heterogeneous Cameras","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03522","noteUrl":"/review?paper=review-neurips-2026-149287","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149288","title":"Causal Effects with Unobserved Unit Types in Interacting Human–AI Systems","authors":["William Overman","Mohamad Sadegh Shirani Faradonbeh","Mohsen Bayati"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The subgroup recursion and identification assumptions need correction","plainExplanation":"The subgroup recursion uses the population's receiver composition, which can make the fitted mean model wrong. Separately, the rank assumptions admit an unidentified regression, and unrestricted subgroup selection can invalidate the direct-treatment factorization. The initial-state assumption and human-effect definition need narrower wording.","primaryUrl":"https://arxiv.org/html/2603.01339","noteUrl":"/review?paper=review-neurips-2026-149288","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","missing_assumption","notation","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-149289","title":"Understanding the Curse of Unrolling","authors":["Sheheryar Mehmood","Florian Knoll","Peter Ochs"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Understanding the Curse of Unrolling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.19733","noteUrl":"/review?paper=review-neurips-2026-149289","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149290","title":"KL for a KL: On-Policy Distillation with Control Variate Baseline","authors":["Minjae Oh","Sangjun Song","Gyubin Choi","Yunho Choi","Yohan Jo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"KL for a KL: On-Policy Distillation with Control Variate Baseline","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07865","noteUrl":"/review?paper=review-neurips-2026-149290","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149292","title":"MIRROR: Manifold Ideal Reference ReconstructOR for AI-Generated Image Detection","authors":["Ruiqi Liu","Manni Cui","Ziheng Qin","Zhiyuan Yan","Ruoxin Chen","Yi Han","Zhiheng Li","Junkai Chen","ZhiJin Chen","Kaiqing Lin","Jialiang Shen","Lubin Weng","Yan Wang","Shu Wu"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"MIRROR: Manifold Ideal Reference ReconstructOR for AI-Generated Image Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.02222","noteUrl":"/review?paper=review-neurips-2026-149292","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149294","title":"Quantizing With Randomized Hadamard Transforms: Efficient Heuristic Now Proven","authors":["Ran Ben-Basat","William Kuszmaul","Michael Mitzenmacher","Amit Portnoy","Shay Vargaftik"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Quantizing With Randomized Hadamard Transforms: Efficient Heuristic Now Proven","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06014","noteUrl":"/review?paper=review-neurips-2026-149294","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149296","title":"Parameterized Stripe Attention for Efficient Video Generation","authors":["xingyu jia","Baole Ai","Ang Wang","Kang Zhao","Yong Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Dominant RoPE channels do not alone imply a fixed heatmap period","plainExplanation":"Dominance of one rotary channel does not make its content-dependent phase constant across keys, so it does not establish a fixed attention period.","primaryUrl":"https://arxiv.org/html/2609.37001","noteUrl":"/review?paper=review-neurips-2026-149296","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149298","title":"Simple KNN-Based Outlier Detection Achieves Robust Clustering","authors":["Tianle Jiang","Yufa Zhou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Simple KNN-Based Outlier Detection Achieves Robust Clustering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07130","noteUrl":"/review?paper=review-neurips-2026-149298","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149299","title":"LongMINT: Evaluating Memory under Multi-Target Interference in Long-Horizon Agent Systems","authors":["Hyunji Lee","Justin Chen","Joykirat Singh","Zaid Khan","Elias Stengel-Eskin","Mohit Bansal"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LongMINT: Evaluating Memory under Multi-Target Interference in Long-Horizon Agent Systems","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.18565","noteUrl":"/review?paper=review-neurips-2026-149299","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149300","title":"Why Heavy-Tailed Weights Predict Model Quality","authors":["Joseph Wilson","Chris van der Heide","Liam Hodgkinson","Zhichao Wang","Fred Roosta","Michael Mahoney"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Why Heavy-Tailed Weights Predict Model Quality","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.13123","noteUrl":"/review?paper=review-neurips-2026-149300","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149304","title":"Constrained Stochastic Spectral Preconditioning Converges for Nonconvex Objectives","authors":["Konstantinos Oikonomidis","Jan Quan","Kimon Antonakopoulos","Antonio Silveti-Falls","Volkan Cevher","Panagiotis Patrinos"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Constrained Stochastic Spectral Preconditioning Converges for Nonconvex Objectives","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11850","noteUrl":"/review?paper=review-neurips-2026-149304","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149306","title":"Change-Robust Online Topological Memory for Long-Term Relocalization and Semantic Navigation","authors":["Jiaming Wang","Diwen Liu","Chen Jizhuo","Atharva A Ghotavadekar","Da Jiaxuan","Linh Kästner","Harold Soh"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Change-Robust Online Topological Memory for Long-Term Relocalization and Semantic Navigation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.02227","noteUrl":"/review?paper=review-neurips-2026-149306","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149307","title":"Frame the adversary: a structure-aware attack methodology","authors":["Vicky Kouni","Stelios Perrakis","Francis Bach","Pascal Frossard","Yann Chevaleyre"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Frame the adversary: a structure-aware attack methodology","plainExplanation":"The correct projection solves a single coupled stationarity equation, rather than multiplying two one-dimensional resolvents. Exact projection ensures feasible iterates; it does not solve a generic nonconcave attack objective in a prescribed finite number of steps.","primaryUrl":"https://arxiv.org/html/2609.31128","noteUrl":"/review?paper=review-neurips-2026-149307","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-149311","title":"OPERA: An Agent for Image Restoration with End-to-End Joint Planning–Execution Optimization","authors":["Feng Zhu","Shuyang Xie","Zeng Yihan","Ming Liu","Wangmeng Zuo"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"OPERA: An Agent for Image Restoration with End-to-End Joint Planning–Execution Optimization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22104","noteUrl":"/review?paper=review-neurips-2026-149311","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149312","title":"Evolutionary System Prompt Learning for Reinforcement Learning in LLMs","authors":["Lunjun Zhang","Ryan Chen","Bradly Stadie"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Evolutionary System Prompt Learning for Reinforcement Learning in LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.14697","noteUrl":"/review?paper=review-neurips-2026-149312","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149316","title":"Extended Wasserstein-GAN Approach to Causal Distribution Learning: Density-Free Estimation and Minimax Optimality","authors":["Shu Tamano","Masaaki Imaizumi"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Extended Wasserstein-GAN Approach to Causal Distribution Learning: Density-Free Estimation and Minimax Optimality","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10206","noteUrl":"/review?paper=review-neurips-2026-149316","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149320","title":"Transforming Image Editors into Video Editors","authors":["Feng Wang","Zijie Li","Ceyuan Yang","Alan Yuille","Peng Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Transforming Image Editors into Video Editors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.11037","noteUrl":"/review?paper=review-neurips-2026-149320","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149324","title":"High-Probability Minimax Adaptive Estimation in Besov Spaces via Online-to-Batch","authors":["Paul Liautaud","Pierre Gaillard","Olivier Wintenberger"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Zero origin regret forces the parameter-free oracle to remain fixed","plainExplanation":"A zero regret allowance at the starting comparator forces every prediction to remain there, contradicting the advertised sublinear guarantee for other comparators. The implementable sublinear oracle claim is affected. Results conditional on an oracle satisfying their premises are not independently refuted.","primaryUrl":"https://arxiv.org/html/2602.11747","noteUrl":"/review?paper=review-neurips-2026-149324","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149326","title":"Lean Refactor: Multi-Objective Controllable Proof Optimization via Agentic Strategy Search","authors":["Jialin Lu","Soonho Kong","Rodrigo Stehling","Kaiyu Yang","Zhangyang &amp;quot;Atlas&amp;quot; Wang","Weiran Sun","Wuyang Chen"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Lean Refactor: Multi-Objective Controllable Proof Optimization via Agentic Strategy Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20244","noteUrl":"/review?paper=review-neurips-2026-149326","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149327","title":"Correlation-Aware Contextual Bandits with Surrogate Rewards for LLM Routing","authors":["Ajay Narayanan Sridhar","Ronak Singh","Mehrdad Mahdavi","Vijaykrishnan Narayanan"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Correlation-Aware Contextual Bandits with Surrogate Rewards for LLM Routing","plainExplanation":"The played mixture uses every exploration copy, so a comparator’s tuned parameter cannot replace the whole grid in its bias bound. The report keeps that unresolved regret-proof gap separate from the second-moment calculation, which follows under the stated assumptions by weighted Cauchy–Schwarz.","primaryUrl":"https://arxiv.org/html/2607.09015","noteUrl":"/review?paper=review-neurips-2026-149327","collection":"neurips-2026","severity":"major","issueTypes":["proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149330","title":"FAST-Brain: A Flow-Aligned Spatio-Temporal Surrogate Brain Model","authors":["Shucheng Liu","Chengchun Shi","Kai Zhang","Hongtu Zhu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FAST-Brain: A Flow-Aligned Spatio-Temporal Surrogate Brain Model","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34354","noteUrl":"/review?paper=review-neurips-2026-149330","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149331","title":"Direct Product Flow Matching: Decoupling Radial and Angular Dynamics for Few-Shot Adaptation","authors":["Hongxu Chen","Yanghao Wang","Bowei Zhu","Hongxiang Li","Ziqi Jiang","Zhen Wang","Lin Li","Rui Liu","Long Chen"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The angular split step is not the full manifold exponential","plainExplanation":"The positive-radius manifold has only a restricted exponential domain. Unrestricted predicted or guided radial velocities can request an undefined step.","primaryUrl":"https://arxiv.org/html/2605.05054","noteUrl":"/review?paper=review-neurips-2026-149331","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149332","title":"Skip the Hessian, Keep the Rates: Globalized Semismooth Newton with Lazy Hessian Updates","authors":["Amal Alphonse","Pavel Dvurechenskii","Clemens Sirotenko"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Skip the Hessian, Keep the Rates: Globalized Semismooth Newton with Lazy Hessian Updates","plainExplanation":"An accepted step with an indefinite Hessian can violate the claimed factor-two gradient bound. The line-search count must also include each successful solve.","primaryUrl":"https://arxiv.org/html/2602.08069","noteUrl":"/review?paper=review-neurips-2026-149332","collection":"neurips-2026","severity":"major","issueTypes":["complexity_error","false_statement","formula_error","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-149334","title":"DeformGen: Dynamics-Based Topology Augmentation for Deformable Manipulation Policy Learning","authors":["Zili Lin","Wenyao Zhang","Yuyang Zhang","Zekun Qi","Junyan Lin","Hanxin Zhu","Jiaolong Yang","Zhibo Chen","Yao Mu","Xiaokang Yang","Xin Jin","Wenjun Zeng"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"DeformGen: Dynamics-Based Topology Augmentation for Deformable Manipulation Policy Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.25939","noteUrl":"/review?paper=review-neurips-2026-149334","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149336","title":"Mitigating Retaliatory Algorithmic Collusion in Repeated Games","authors":["Karthik Sivachandran","Rohan Paleja"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The continuation values use different discount origins","plainExplanation":"The equilibrium characterization subtracts continuation values with different discount origins and can reject the unique equilibrium of a one-action game.","primaryUrl":"https://arxiv.org/html/2609.20548","noteUrl":"/review?paper=review-neurips-2026-149336","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149337","title":"Dense Structural Compression of Transformers via Gauge-Correct Channel Removal","authors":["Jed A Duersch","Naïm Es-sebbani","Nathanaël Haas","Zied Bouraoui"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Dense Structural Compression of Transformers via Gauge-Correct Channel Removal","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.07264","noteUrl":"/review?paper=review-neurips-2026-149337","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149338","title":"ESS-Flow: training-free guidance as Bayesian inference in source space","authors":["Adhithyan Kalaivanan","Zheng Zhao","Jens Sjölund","Fredrik Lindsten"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Multifidelity importance correction needs domination","plainExplanation":"Coarse-sample importance weighting cannot recover fine-target mass outside the coarse support; the general correction needs a domination assumption.","primaryUrl":"https://arxiv.org/html/2510.05849v1","noteUrl":"/review?paper=review-neurips-2026-149338","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149339","title":"ELSA3D: Elastic Semantic Anchoring for Unified 3D Understanding and Generation","authors":["Tianjiao (Joey) Yu","Xinzhuo Li","Yifan Shen","Onkar Susladkar","Yuanzhe Liu","Xiaona Zhou","Ismini Lourentzou"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"ELSA3D: Elastic Semantic Anchoring for Unified 3D Understanding and Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.06565","noteUrl":"/review?paper=review-neurips-2026-149339","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149340","title":"Scale-Sensitive Shattering: Learnability and Evaluability at Optimal Scale","authors":["Shashaank Aiyer","Yishay Mansour","Shay Moran","Han Shao","Tom Waknine"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The sufficient sample-size substitution drops dimension and an error power","plainExplanation":"The proof substitutes an entropy estimate into a uniform-convergence lemma and replaces its sufficient sample-size condition by one lacking dimension and one inverse power of the error. The direct substitution requires \\(n>(128R^2/\\varepsilon^2)[14d\\log(32R/\\varepsilon)\\log^2n+\\log(1/\\delta)]\\).","primaryUrl":"https://arxiv.org/html/2605.13684","noteUrl":"/review?paper=review-neurips-2026-149340","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149345","title":"Touch-R1: Reinforcing Touch Reasoning in MLLMs","authors":["Yingxin Lai","Yafei Zhou","Fucai Zhu","Siyu Zhu","Weihao Yuan"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Touch-R1: Reinforcing Touch Reasoning in MLLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27154","noteUrl":"/review?paper=review-neurips-2026-149345","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149346","title":"GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning","authors":["Shaoxiang Qin","Xiongye Xiao","Yucheng Zhao","Fuyuan Lyu","Di Zhou","Jiachen Yao","Steve Liu","Animashree Anandkumar","Liangzhu Leon Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36056","noteUrl":"/review?paper=review-neurips-2026-149346","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149347","title":"PENEX: AdaBoost-Inspired Neural Network Regularization","authors":["Klaus-Rudolf Kladny","Bernhard Schölkopf","Michael Muehlebach"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"PENEX: AdaBoost-Inspired Neural Network Regularization","plainExplanation":"The 24 October revision retains the finite-minimizer claim without positive conditional class probabilities, the potentially empty unit-direction feasible class, the unweighted class sum in the final KKT equation, and the reversed sign in the background SAMME update. The existing proofs and repairs apply to these same passages. Their scopes remain separate from the empirical classifier results.","primaryUrl":"https://arxiv.org/html/2510.02107","noteUrl":"/review?paper=review-neurips-2026-149347","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149348","title":"Unified Panoramic Geometry Estimation via Multi-View Foundation Models","authors":["Vukasin Bozic","Isidora Slavkovic","Dominik Narnhofer","Nando Metzger","Denis Rozumny","Konrad Schindler","Nikolai Kalischek"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Unified Panoramic Geometry Estimation via Multi-View Foundation Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26368","noteUrl":"/review?paper=review-neurips-2026-149348","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149349","title":"Robust Importance Sampling for Rare Events via Constrained Gaussian Mixtures","authors":["Pawel Lorek","Rafal Nowak","Rafał Topolnicki","Tomasz Trzcinski","Maciej Zieba"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Gaussian second moment omits a variance factor","plainExplanation":"The expansion omits the inverse proposal variance from the linear term in the exponent. Its quadratic divergence criterion remains correct. The integral expansion is affected. The infinite-variance conclusion for \\(\\sigma^2<1/2\\) is governed by the unchanged quadratic coefficient, and the boundary sign test also retains its direction.","primaryUrl":"https://arxiv.org/html/2610.07485","noteUrl":"/review?paper=review-neurips-2026-149349","collection":"neurips-2026","severity":"repairable","issueTypes":["numerical_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149350","title":"GeoSPRINT: Geometric Redundancy-Aware Step Pruning for Inference in Diffusion Trajectories","authors":["Arpita Joshi"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Retained-window thresholds need not support binary-search optimality","plainExplanation":"Pruning changes later reference windows, making threshold responses nonmonotone and defeating the stated generic binary-search maximality justification.","primaryUrl":"https://arxiv.org/html/2609.02160","noteUrl":"/review?paper=review-neurips-2026-149350","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149352","title":"Beyond Spatial and Temporal Priors: A Generalizable Approach for Dense Correspondence Matching","authors":["Luping Liu","Bingyi Kang","Yifan Wang","Dong Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond Spatial and Temporal Priors: A Generalizable Approach for Dense Correspondence Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.12421","noteUrl":"/review?paper=review-neurips-2026-149352","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149353","title":"Contrast encodes inductive bias: separating slow noise from dynamics in predictive representation learning","authors":["Paarth Gulati","Ilya Nemenman"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Within-trajectory negatives remove a shortcut but do not identify common coordinates","plainExplanation":"Removing a nuisance-only shortcut does not identify a nuisance-independent representation. A trajectory-dependent symmetry can preserve every comparison while changing the latent coordinates used across trajectories.","primaryUrl":"https://arxiv.org/html/2606.07770","noteUrl":"/review?paper=review-neurips-2026-149353","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149354","title":"GHOST: Geometry-Hierarchical Online Streaming Token Eviction for Efficient 3D Reconstruction","authors":["Leyang Chen","Junyi Wu","Zhiteng Li","Yulun Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A fixed score boost does not protect every special token globally","plainExplanation":"In the eviction branch, TopK keeps \\(B_\\ell\\) old entries, then the next step appends \\(N\\) current-frame pairs. For \\(N>0\\), the returned cache has \\(B_\\ell+N>B_\\ell\\). Calling \\(B_\\ell\\) an old-cache budget would change the explicit output specification; assuming current entries were already included instead makes the unconditional append duplicate them. The pseudocode’s returned-size guarantee fails. No actual memory-use defect is inferred.","primaryUrl":"https://arxiv.org/html/2605.15852","noteUrl":"/review?paper=review-neurips-2026-149354","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149355","title":"Simple Baselines are Competitive with Code Evolution","authors":["Yonatan Gideoni","Sebastian Risi","Yarin Gal"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The tie branch overallocates probability of dominance","plainExplanation":"If \\(L\\) methods share the maximum, the displayed denominator counts only competitors of the focal method, hence \\(L-1\\). Each tied method receives \\(1/(L-1)\\), totaling \\(L/(L-1)>1\\), rather than one. Reordering each method into the focal position does not add the missing self-count. This is an exact probability-normalization error. It does not establish the empirical size of its effect; continuous no-tie settings do not exercise the branch.","primaryUrl":"https://arxiv.org/html/2602.16805","noteUrl":"/review?paper=review-neurips-2026-149355","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149356","title":"MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM","authors":["Omin Kwon","Yeonjae Kim","Doyeon Kim","Minseo Kim","Yeonhong Park","Jae W. Lee"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.14209","noteUrl":"/review?paper=review-neurips-2026-149356","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149357","title":"Calibration Is Not Control: Intervention Advantage for LLM-Agent Oversight","authors":["Chubin Zhang","Zhenglin Wan","Xingrui Yu","jingxuan wu","Qi Wen","Pengfei Zhou","Wangbo Zhao","Ivor Tsang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Calibration Is Not Control: Intervention Advantage for LLM-Agent Oversight","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.21399","noteUrl":"/review?paper=review-neurips-2026-149357","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149358","title":"On the Robustness of Watermarking for Autoregressive Image Generation","authors":["Andreas Müller","Denis Lukovnikov","Shingo Kodama","Minh Pham","Anubhav Jain","Jonathan Petit","Niv Cohen","Asja Fischer"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"On the Robustness of Watermarking for Autoregressive Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.11720","noteUrl":"/review?paper=review-neurips-2026-149358","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149359","title":"GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game Theory","authors":["Pepijn Cobben","Xuanqiang A Huang","Thao Pham","Isabel Dahlgren","Bernhard Schölkopf","Terry J. C. Zhang","Zhijing Jin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Raw signed payoff products do not have the stated welfare ordering","plainExplanation":"The No Conflict verifier is described as checking that both players globally prefer the same cell, to establish a pair of strictly dominant strategies. Row action zero is strictly dominant only if \\(R_{00}>R_{10}\\) and \\(R_{01}>R_{11}\\).","primaryUrl":"https://arxiv.org/html/2602.12316","noteUrl":"/review?paper=review-neurips-2026-149359","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149362","title":"Achieving Directional-Stationarity from a Single Random Direction Step","authors":["Dan Greenstein","Nadav Hallak"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Achieving Directional-Stationarity from a Single Random Direction Step","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22045","noteUrl":"/review?paper=review-neurips-2026-149362","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149364","title":"Factorized Spectral Representations for Reinforcement Learning","authors":["Junyi Wu","Dan Li"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Feature dimension does not determine whether NCE is informative","plainExplanation":"Feature dimension does not determine contrastive-loss validity, and enough samples do not ensure a full gradient span. The actor expression also detaches its only parameter-dependent path.","primaryUrl":"https://arxiv.org/html/2607.13498","noteUrl":"/review?paper=review-neurips-2026-149364","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149365","title":"Extracting Training Data from Diffusion Language Models via Infilling","authors":["Yihan Wang","N. Asokan"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Extracting Training Data from Diffusion Language Models via Infilling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.24173","noteUrl":"/review?paper=review-neurips-2026-149365","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149366","title":"BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning","authors":["Shaokai Ye","Vasileios Saveris","Yihao Qian","Jiaming Hu","Elmira Amirloo","Peter Grasch"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Reward-coordinate variance does not ensure the stated noninvariance","plainExplanation":"Nonzero variance in individual reward coordinates does not ensure a defined aggregate normalization or the claimed unconditional distinction between normalization schemes.","primaryUrl":"https://arxiv.org/html/2605.07394","noteUrl":"/review?paper=review-neurips-2026-149366","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149367","title":"What does a Bayes-filtered transformer believe? A predictive Monte Carlo approach","authors":["Afiq Abdillah Effiezal Aswadi","Haotong Ma","Susan Wei"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"What does a Bayes-filtered transformer believe? A predictive Monte Carlo approach","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.17060","noteUrl":"/review?paper=review-neurips-2026-149367","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149374","title":"Aligning Flow Map Policies with Optimal $Q$-Guidance","authors":["Christos Ziakas","Alessandra Russo","Joey Bose"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The objective equivalence needs correction; the trust-region formulas need consistent units","plainExplanation":"The Eulerian residual contains the instantaneous velocity \\(v_r\\). Replacing it with \\(u_{r,t}\\) changes the objective and its parameter gradient.","primaryUrl":"https://arxiv.org/html/2605.12416","noteUrl":"/review?paper=review-neurips-2026-149374","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149375","title":"When Language Overrules: Revealing Text Dominance in Multimodal Large Language Models","authors":["Huyu Wu","Meng Tang","Xinhan Zheng","Duo Su","Haiyun Jiang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The threshold minimum may not exist","plainExplanation":"The real-threshold feasible set can be open, so the displayed minimum need not exist. The earlier top-count retention rule avoids that problem.","primaryUrl":"https://arxiv.org/html/2508.10552v1","noteUrl":"/review?paper=review-neurips-2026-149375","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149376","title":"Neural Field Thermal Tomography: A Differentiable Physics Framework for Non-Destructive Evaluation","authors":["Tao Zhong","Yixun Hu","Dongzhe Zheng","Aditya Sood","Christine Allen-Blanchette"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Neural Field Thermal Tomography: A Differentiable Physics Framework for Non-Destructive Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.11045","noteUrl":"/review?paper=review-neurips-2026-149376","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149377","title":"Quotient-Categorical Representations for Bellman-Compatible Average-Reward Distributional Reinforcement Learning","authors":["Ege C. Kaya","Aliasghar Pourghanitazehshahri","Vijay Gupta","Abolfazl Hashemi"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Quotient-Categorical Representations for Bellman-Compatible Average-Reward Distributional Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11289","noteUrl":"/review?paper=review-neurips-2026-149377","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149380","title":"Training with Harnesses: On-Policy Harness Self-Distillation for Complex Reasoning","authors":["Zhengyang Zhao","Lu Ma","Wentao Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Training with Harnesses: On-Policy Harness Self-Distillation for Complex Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08741","noteUrl":"/review?paper=review-neurips-2026-149380","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149381","title":"Strategic Decision Support for AI Agents","authors":["Shayan Kiyani","Sima Noorani","George J. Pappas","Hamed Hassani"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Strategic Decision Support for AI Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.12587","noteUrl":"/review?paper=review-neurips-2026-149381","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149382","title":"Towards more general control of diffusion models using Jeffrey Guidance","authors":["Raphaël Razafindralambo","Rémy Sun","Frederic Precioso","Jes Frellsen","Pierre-Alexandre Mattei"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The cumulative noise product must end at the current time","plainExplanation":"The density-ratio score display has the wrong sign, and the cumulative noise product ends at the terminal rather than current time.","primaryUrl":"https://arxiv.org/html/2606.13240","noteUrl":"/review?paper=review-neurips-2026-149382","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149385","title":"Reinforcement Learning for Code Optimization","authors":["Pierre Chambon","Kunhao Zheng","Juliette Decugis","Benoît Sagot","Gabriel Synnaeve"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Reinforcement Learning for Code Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.25970","noteUrl":"/review?paper=review-neurips-2026-149385","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149386","title":"VACE: Learning Geometrically Structured Representations for Time Series Anomaly Detection","authors":["Alberto D Cencillo","Leonardo Concepción","Isaac Triguero","Julián Luengo"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"VACE: Learning Geometrically Structured Representations for Time Series Anomaly Detection","plainExplanation":"After standardization, two negative component values can produce a large positive product. The printed score therefore lacks its stated conjunction property; an explicit intersection or positive-part construction supplies that property. The released min-max-normalized score is a different case.","primaryUrl":"https://arxiv.org/html/2605.23504","noteUrl":"/review?paper=review-neurips-2026-149386","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149390","title":"Learning Options for Compositional Motor Control with Adapter Banks","authors":["Sreejan Kumar","Marcelo G Mattar","Lea Duncker"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Options for Compositional Motor Control with Adapter Banks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.17042","noteUrl":"/review?paper=review-neurips-2026-149390","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149391","title":"PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics","authors":["Bardienus Duisterhof","Kaifeng Zhang","Adam Hung","Bowen Wen","Stan Birchfield","Yunzhu Li","Deva Ramanan","Jeffrey Ichnowski"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.19142","noteUrl":"/review?paper=review-neurips-2026-149391","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149393","title":"Occupancy-based Quantile Risk Control","authors":["Zihao Shi","Huajun Xi","Bingyi Jing","Hongxin Wei"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The limiting upper-risk formula uses lower endpoints","plainExplanation":"The limiting upper-risk formula uses lower endpoints and can miss the stated coverage. Separately, the finite Monte Carlo algorithm omits simulation uncertainty, and the extension to tied losses needs independent tie breakers.","primaryUrl":"https://arxiv.org/html/2609.03104","noteUrl":"/review?paper=review-neurips-2026-149393","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","incorrect_bound","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149394","title":"Learning When to Trust LLM Priors: A Validated Framework for Semantic Prior Integration","authors":["Erica Zhang","Naomi Sagan","Danny Tse","Fangzhao Zhang","Mert Pilanci","Jose Blanchet"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Laplace scale gives larger penalties to larger semantic scores","plainExplanation":"At zero tilt, the derivative of the tilted score mean is \\(\\operatorname{Var}(s)\\). Differentiating the target mean condition gives \\(\\eta^{\\prime}(0)=(s_{\\mathrm{target}}-\\bar s)/\\operatorname{Var}(s)\\), hence \\(\\rho_i^{\\prime}(0)=(s_{\\mathrm{target}}-\\bar s)(s_i-\\bar s)/\\operatorname{Var}(s)\\). The printed expansion omits the target shift. For scores \\(0,2\\), target \\(1.5\\) and mixing \\(0.1\\), the exact weights are \\(0.95,1.05\\), rather than the unscaled expansion’s values. This is a missing factor in the appendix Taylor expansion. The main affine heuristic can remain a rescaled approximation.","primaryUrl":"https://arxiv.org/html/2601.21410","noteUrl":"/review?paper=review-neurips-2026-149394","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149395","title":"NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models","authors":["Ziyue Zhu","Shangyang Wu","Shuai Zhao","Zhao ZhiQiu","Jian Zhang","Li Shengjie","Yi Wang","Anh Tuan Luu","XINLIANG ZHOU","Fang Li","Haoran Luo"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Potential-based invariance does not cover the full dense reward","plainExplanation":"The potential result covers \\(r_t+\\gamma\\Phi_{t+1}-\\Phi_t\\). Equation (8) additionally includes positively weighted subtask, smoothness and grasp rewards that are not shown to be potential differences; the unshaped task reward is not defined to include them. Such additions can change the ordering of policies, so the proven identity does not establish that the complete dense reward introduces no new optimal policies. The potential-shaping theorem and separate KL argument remain valid within their scope. Full task-optimality preservation requires a new proof.","primaryUrl":"https://arxiv.org/html/2603.09542","noteUrl":"/review?paper=review-neurips-2026-149395","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149396","title":"Verification-Aware Training for Speculative Decoding","authors":["Geonmo Gu","Byeongho Heo","HeeJae Jun","Yoohoon Kang","Sangmin Lee","Sangdoo Yun","Dongyoon Han"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Verification-Aware Training for Speculative Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.30135","noteUrl":"/review?paper=review-neurips-2026-149396","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149398","title":"Generalizing Test-time Compute-optimal Scaling as an Optimizable Graph","authors":["Fali Wang","Jihai Chen","Shuhua Yang","Runxue Bao","Zhiwei Zhang","Tianxiang Zhao","Hui Liu","Xianfeng Tang","Qi He","Suhang Wang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Generalizing Test-time Compute-optimal Scaling as an Optimizable Graph","plainExplanation":"The order of the degree blocks depends on the original vertex order, so Listing 1 overcounts isomorphism classes. Sorting the invariant block keys repairs the canonicalization argument. This finding concerns the printed enumeration, not the graph optimizer’s experimental performance.","primaryUrl":"https://arxiv.org/html/2511.00086","noteUrl":"/review?paper=review-neurips-2026-149398","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149402","title":"Causal Discovery via Transformed Low-Rank Quantile Surfaces","authors":["Ryo Kamimura","Thong Pham"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Causal Discovery via Transformed Low-Rank Quantile Surfaces","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.16931","noteUrl":"/review?paper=review-neurips-2026-149402","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149404","title":"ProCTI: Prototype-Refined Global Conditioning for Diffusion-Based Time Series Imputation","authors":["Fariza Rashid","Duc Van Le","Rahat Masood","Gustavo E Batista","Aruna Seneviratne","Suranga Seneviratne"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"ProCTI: Prototype-Refined Global Conditioning for Diffusion-Based Time Series Imputation","plainExplanation":"Information about a target’s uncertainty need not improve its best squared-error prediction. Strict improvement requires the conditional means to differ.","primaryUrl":"https://arxiv.org/html/2609.37632","noteUrl":"/review?paper=review-neurips-2026-149404","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149405","title":"Fail-Closed Alignment for Large Language Models","authors":["Zachary Coalson","Sanghyun Hong"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Fail-Closed Alignment for Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.16977","noteUrl":"/review?paper=review-neurips-2026-149405","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149406","title":"Implicit Bias of Mirror Flow in Homogeneous Neural Networks: Sparse and Dense Feature Learning","authors":["Tom Jacobs","Guido Montufar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strict convexity does not ensure an invertible mirror Hessian","plainExplanation":"The finite-time KKT estimate omits a horizon-gradient approximation remainder. A fully radial valid flow point can have positive stationarity residual despite its claimed zero bound.","primaryUrl":"https://arxiv.org/html/2605.19458","noteUrl":"/review?paper=review-neurips-2026-149406","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149407","title":"An Information-Theoretic Evaluation Framework for Benchmark and Model Diagnosis in Knowledge Tracing","authors":["Houru Jiang","Zixi Wang","Tengteng Cheng","Xueyi Li","Mingliang Hou","Jiaqi Zheng","Renqiang Luo","Teng Guo","Zitao Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"An Information-Theoretic Evaluation Framework for Benchmark and Model Diagnosis in Knowledge Tracing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06988","noteUrl":"/review?paper=review-neurips-2026-149407","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149409","title":"Memorize Theorems, Not Instances: Probing SFT Generalization through Mathematical Reasoning","authors":["Ruiying Peng","Mengyu Yang","Jing Lei","Xiao-Hui Li","Xueyu Wu","Xinlei Chen"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The theorem-composition examples need local geometric and notation corrections","plainExplanation":"The final angle relation is correct, but \\(AC\\) is tangent to the circle at \\(C\\); \\(\\angle ACD\\) is not an ordinary inscribed angle with both sides chords.","primaryUrl":"https://arxiv.org/html/2605.09270","noteUrl":"/review?paper=review-neurips-2026-149409","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149410","title":"Enabling approximate joint sampling in diffusion LMs","authors":["Parikshit Bansal","Sujay Sanghavi"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Enabling approximate joint sampling in diffusion LMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.22738","noteUrl":"/review?paper=review-neurips-2026-149410","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149411","title":"ZEBRA: Zero-shot Budgeted Resource Allocation for LLM Orchestration","authors":["May Hamri","Inbal Talgam-Cohen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ZEBRA: Zero-shot Budgeted Resource Allocation for LLM Orchestration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20485","noteUrl":"/review?paper=review-neurips-2026-149411","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149414","title":"MulCLIP: A Multi-level Alignment Framework for Enhancing Fine-grained Long-context CLIP","authors":["Chau Truong","Hieu Ta","Zhenzhen Liu","Dung Le"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MulCLIP: A Multi-level Alignment Framework for Enhancing Fine-grained Long-context CLIP","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.07128","noteUrl":"/review?paper=review-neurips-2026-149414","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149416","title":"Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers","authors":["Shuhong Zheng","Michael Oechsle","Erik Sandström","Marie-Julie Rakotosaona","Federico Tombari","Igor Gilitschenski"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23892","noteUrl":"/review?paper=review-neurips-2026-149416","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149419","title":"STDec: Spatio-Temporal Stability Guided Decoding for dLLMs","authors":["Yuzhe Chen","Jiale Cao","Xuyang Liu","Jin Xie","Aiping Yang","Yanwei Pang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"STDec: Spatio-Temporal Stability Guided Decoding for dLLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.06330","noteUrl":"/review?paper=review-neurips-2026-149419","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149421","title":"Vision Transformers Learn Gestalt-Like Figure-Ground Cues from Natural Images","authors":["Matthias Tangemann","Benjamin Lo","Zygmunt Pizlo","Kaleem Siddiqi","Dirk Bernhardt-Walther","Sven Dickinson"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Vision Transformers Learn Gestalt-Like Figure-Ground Cues from Natural Images","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.08932","noteUrl":"/review?paper=review-neurips-2026-149421","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149422","title":"Masked Diffusion Vision-Language Models for Temporal Action Localization","authors":["Fengshun Wang","Zhengbo Zhang","Zhigang Tu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Revealed positions are no longer editable under the specified decoder","plainExplanation":"The masked set shrinks monotonically. Once position \\(i\\) is revealed, later transitions leave its value fixed because only still-masked positions are updated. Delayed revelation of some positions does not imply continued editability of all positions. The universal decoder-capability statement is false. Delayed boundary commitment and reported temporal-localization results are not refuted.","primaryUrl":"https://arxiv.org/html/2605.29858","noteUrl":"/review?paper=review-neurips-2026-149422","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149423","title":"V-GIFT: Boosting Visual Instruction Tuning with Self-Supervised Guidance","authors":["Sophia Sirko-Galouchenko","Monika Wysoczańska","Andrei Bursuc","Nicolas THOME","Spyridon Gidaris"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"V-GIFT: Boosting Visual Instruction Tuning with Self-Supervised Guidance","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.12966","noteUrl":"/review?paper=review-neurips-2026-149423","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149426","title":"Truthful Calibration Errors for Multi-Class Prediction","authors":["Yuxuan Lu","Yifan Wu","Jason Hartline","Lunjia Hu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Truthful Calibration Errors for Multi-Class Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.06388","noteUrl":"/review?paper=review-neurips-2026-149426","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149427","title":"Scalable Variational Bayesian Fine-Tuning of LLMs via Orthogonalized Low-Rank Adapters","authors":["Haotian Xiang","Bingcong Li","Qin Lu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The mean update descends the ELBO that is meant to increase","plainExplanation":"The algorithm descends an ELBO that the derivation maximizes. Its covariance construction also confuses a block of an inverse with an inverse of a block.","primaryUrl":"https://arxiv.org/html/2604.03388","noteUrl":"/review?paper=review-neurips-2026-149427","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149433","title":"BoxTuning: Object-Aware Visual Prompting for  Multimodal Model Fine-Tuning","authors":["Zekun Qian","Ruize Han","Wei Feng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"BoxTuning: Object-Aware Visual Prompting for  Multimodal Model Fine-Tuning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.11136v1","noteUrl":"/review?paper=review-neurips-2026-149433","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149435","title":"Learning When to Denoise: Optimizing Asynchronous Schedules for Latent Diffusion","authors":["Bingshuo Qian","Xiang Cheng"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Learning When to Denoise: Optimizing Asynchronous Schedules for Latent Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19662","noteUrl":"/review?paper=review-neurips-2026-149435","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149436","title":"MaPPO: Maximum a Posteriori Preference Optimization with Prior Knowledge","authors":["Guangchen Lan","Sipeng Zhang","Tianle Wang","Yuwei Zhang","Xinpeng Wei","Daoan Zhang","Xiaoman Pan","Hongming Zhang","Dong-Jun Han","Christopher Brinton"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The prior term has the wrong sign and need not be a probability below one","plainExplanation":"Let \\(a,b\\) be winner and loser rewards, \\(\\delta\\in[0,1]\\), and \\(p=(e^a+e^{\\delta b})/(e^a+e^b)\\). Then \\(\\log p=\\log\\sigma(a-b)-\\log\\sigma(a-\\delta b)\\). With \\(L_p=-\\log p\\), the intended new loss equals \\(L_{\\mathrm{MLE}}-L_p\\), not \\(L_{\\mathrm{MLE}}+L_p\\). The two printed sides differ by \\(2L_p\\).\n\nFor \\(\\delta<1\\), \\(p>1\\) whenever \\(b<0\\). The reward substitution \\(b=\\beta\\log(\\pi/\\pi_{\\mathrm{ref}})\\) permits negative rewards, so the claimed probability range also fails. The central derivation and range sentence fail. The directly trained equation (8) and its experiments do not depend on that equality.","primaryUrl":"https://arxiv.org/html/2507.21183","noteUrl":"/review?paper=review-neurips-2026-149436","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149437","title":"World Tracing: Pixel-Aligned Geometry Beyond the Visible","authors":["Hao Zhang","Mohamed El Banani","Jen-Hao Cheng","Paul Zhang","Yi Hua","Ben Mildenhall","Christoph Lassner","Narendra Ahuja","Gengshan Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"World Tracing: Pixel-Aligned Geometry Beyond the Visible","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.13652","noteUrl":"/review?paper=review-neurips-2026-149437","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149439","title":"Absolute State-wise Constrained Policy Optimization: High-Probability State-wise Constraints Satisfaction","authors":["Weiye Zhao","Feihan Li"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Return variance has no positive uniform minimum over all policies","plainExplanation":"The policy class has no positive uniform return-variance floor, and the finite-horizon safety correction has an incorrect sign and limiting argument.","primaryUrl":"https://arxiv.org/html/2410.01212","noteUrl":"/review?paper=review-neurips-2026-149439","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149440","title":"Unlearning That Lasts: Utility-Preserving, Robust, and Almost Irreversible Forgetting in LLMs","authors":["Naman Deep Singh","Maximilian Mueller","Amit Peleg","Francesco Croce","Matthias Hein"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Unlearning That Lasts: Utility-Preserving, Robust, and Almost Irreversible Forgetting in LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.02820","noteUrl":"/review?paper=review-neurips-2026-149440","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149442","title":"FASTER: Value-Guided Sampling for Fast RL","authors":["Perry Dong","Alexander Swerdlow","Dorsa Sadigh","Chelsea Finn"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FASTER: Value-Guided Sampling for Fast RL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.19730","noteUrl":"/review?paper=review-neurips-2026-149442","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149443","title":"TStruct: Learning Shared Temporal Structures for Long-Term Time-Series Forecasting","authors":["Yangze Li","Zhengyang Zhou","Qihe Huang","Pengkun Wang","Binwu Wang","Yang Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TStruct: Learning Shared Temporal Structures for Long-Term Time-Series Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.17864","noteUrl":"/review?paper=review-neurips-2026-149443","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149444","title":"Patch Hierarchical Attention Transformers for Efficient Particle Jet Tagging","authors":["Zihan Zhao","Aaron Wang","Alan Xia","Chang Sun","Javier Duarte","Abhijith Gandrakota","Jennifer Ngadiuba","Richard Cavanaugh"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Small patches do not by themselves give near-linear attention cost","plainExplanation":"Dense global patch attention can remain quadratic for fixed patch size; a patch size small relative to input size does not establish near-linear cost.","primaryUrl":"https://arxiv.org/html/2605.21789","noteUrl":"/review?paper=review-neurips-2026-149444","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149446","title":"Think, then Score: Decoupled Reasoning and Scoring for Video Reward Modeling","authors":["Yuan Wang","Ouxiang Li","Yulong Xu","Borui Liao","Jiajun Liang","Jinghan Li","Meng Wang","Xintao Wang","Pengfei Wan","Kuien Liu","Xiang Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Token averaging changes the response-length variance scaling","plainExplanation":"Under the appendix’s martingale and independence approximation, \\(S=\\sum_{t=1}^T s_t\\) has variance at least \\(TG_0\\). But equation (7) gives per-chain gradient \\(AS/(GT)\\), not \\(AS\\). With unit advantage second moment and per-step variance at most \\(G_{\\max}\\), the same calculation gives variance at most \\(G_{\\max}/(G^2T)\\). The omitted proportionality factor depends on length and reverses this scaling. The unnormalized conditional calculation can be correct. Its response-length conclusion does not establish the claimed scaling for the described token-averaged objective or explain the observed fluctuations.","primaryUrl":"https://arxiv.org/html/2605.05922","noteUrl":"/review?paper=review-neurips-2026-149446","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149448","title":"Generating the Unheard: Phylogeny-Guided Latent Generation for Ancestral Sound Reconstruction","authors":["Tianyi Xu","Shrinaath Narasimhan","Evan Gorstein","Santiago Perea","Yunyi Shen","Claudia Solis-Lemus"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Row orthogonality does not identify a unique projection subspace","plainExplanation":"The constraint fixes row lengths and orthogonality, but it does not identify which row space maximizes the objective. Different subspaces can give exactly the same correlation. This concerns the identifiability assertion, not the row-orthogonality calculation, the Brownian posterior conditional on a chosen representation, or the reported reconstruction results.","primaryUrl":"https://arxiv.org/html/2609.15240","noteUrl":"/review?paper=review-neurips-2026-149448","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149449","title":"Validating Causal Abstraction Metrics on Simulated Complex Systems","authors":["Maxime Méloux","Tiago Pimentel","François Portet","Maxime Peyrard"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The rare-trigger constant is too small for ninety-five percent detection","plainExplanation":"The stated rare-trigger probability requires a larger analytic constant for ninety-five percent detection. The linear dependence on path length remains.","primaryUrl":"https://arxiv.org/html/2607.00267","noteUrl":"/review?paper=review-neurips-2026-149449","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149451","title":"Impossibility of Distribution-Free Predictive Inference for Individual Treatment Effects","authors":["Chongguang Tao","Zheng Zhou","Yuhong Yang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Impossibility of Distribution-Free Predictive Inference for Individual Treatment Effects","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05051","noteUrl":"/review?paper=review-neurips-2026-149451","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149457","title":"Scaling Full Conformal Image Classifiers","authors":["Julio Silva-Rodríguez","Ender Konukoglu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Scaling Full Conformal Image Classifiers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.37298","noteUrl":"/review?paper=review-neurips-2026-149457","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149459","title":"VSearcher: Long-Horizon Multimodal Search Agent via Reinforcement Learning","authors":["RUIYANG ZHANG","Qianguo Sun","Chao Song","Yiyan Qi","Zhedong Zheng"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"VSearcher: Long-Horizon Multimodal Search Agent via Reinforcement Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.02795","noteUrl":"/review?paper=review-neurips-2026-149459","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149468","title":"Drift-Resistant Navigation World Model with Anchored Epipolar Guidance","authors":["Po-Chien Luan","Zimin Xia","WUYANG LI","Yang Gao","Alexandre Alahi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Drift-Resistant Navigation World Model with Anchored Epipolar Guidance","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24761","noteUrl":"/review?paper=review-neurips-2026-149468","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149471","title":"CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models","authors":["Xiao Zhu","Xinyu Zhou","Boyu Zhu","Hanxu Hu","Mingzhe Du","Haotian Zhang","Huiming Wang","Zhijiang Guo"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.17684","noteUrl":"/review?paper=review-neurips-2026-149471","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149475","title":"Strong Stochastic Flow Maps","authors":["Sam McCallum","Zander Blasingame","Timothy Herschell","Niklas Rindtorff","Alexander Tong","James Foster"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The same stochastic flow has distinct coefficient decompositions","plainExplanation":"Different path-dependent coefficient pairs can produce exactly the same map. Diagonal coefficient matching and the semigroup loss do not distinguish those decompositions. This refutes coefficient identifiability in the stated path-dependent parametrization. It does not refute uniqueness of the strong SDE solution or the characterization of the induced Itô map.","primaryUrl":"https://arxiv.org/html/2606.01086","noteUrl":"/review?paper=review-neurips-2026-149475","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149477","title":"The Horizon Threshold in Cooperative Multi-Agent Reward-Free Exploration","authors":["Idan Barnea","Orin Levy","Yishay Mansour"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The revealed reward identifies the supposedly hidden final action","plainExplanation":"The lower-bound learner receives a reward that reveals the supposedly hidden final action. Two further proof constants need correction: the multinomial confidence radius is too small, and the final failure calculation uses a success bound.","primaryUrl":"https://arxiv.org/html/2602.01453","noteUrl":"/review?paper=review-neurips-2026-149477","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149479","title":"Program-as-Weights: A Programming Paradigm for Fuzzy Functions","authors":["Wentao Zhang","Liliana Hotsko","Woojeong Kim","Pengyu Nie","Stuart Shieber","Yuntian Deng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Program-as-Weights: A Programming Paradigm for Fuzzy Functions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.02512","noteUrl":"/review?paper=review-neurips-2026-149479","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149483","title":"A Theoretical Analysis of Test-Driven Code Generation","authors":["Nicolas Menet","Michael Hersche","Andreas Krause","Abbas Rahimi"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A Theoretical Analysis of Test-Driven Code Generation","plainExplanation":"Distinct programs and a nondegenerate test-suite distribution can still give a constant agreement fraction, making the estimator unbiased. Positive variance of that fraction gives the strict Jensen conclusion; weak upward bias and consistency remain valid.","primaryUrl":"https://arxiv.org/html/2602.06098","noteUrl":"/review?paper=review-neurips-2026-149483","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149486","title":"RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping","authors":["Yuzong Chen","Xilai Dai","Jake Hyun","Chi-Chih Chang","Wonsuk Jang","Yuheng Wu","Thierry Tambe","Jae-sun Seo","Mohamed Abdelfattah"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2501.04052","noteUrl":"/review?paper=review-neurips-2026-149486","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149487","title":"Ensemble Modeling for Time Series Forecasting: an Adaptive Robust Optimization Approach","authors":["Leonard Boussioux","Henry Mao","Dimitris Bertsimas"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The stated Gaussian scale can be negative","plainExplanation":"The synthetic Gaussian rule permits negative standard deviations, the test step uses the wrong split design, and the displayed exponential weights are not normalized.","primaryUrl":"https://arxiv.org/html/2304.04308","noteUrl":"/review?paper=review-neurips-2026-149487","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149488","title":"Continuous Open-ended Discovery and Evolution of Skills as Hierarchical Reward Programs","authors":["Richard Bornemann","Pierluigi V Amadori","Antoine Cully"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Continuous Open-ended Discovery and Evolution of Skills as Hierarchical Reward Programs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.10085","noteUrl":"/review?paper=review-neurips-2026-149488","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149489","title":"ODEWorld: A Continuous Predictive Architecture via Physical-Time Flow","authors":["Dongxiu Liu","Haoyi Niu","Peng Cheng","Yuan Gao","Xirui Kang","Sangli Teng","Koushil Sreenath","Xianyuan Zhan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Resetting time and conditioning does not invert the learned flow","plainExplanation":"An inverse path ending at \\(T\\) needs \\(dy/du=-v(y,T-u;z_0,c)\\), preserving original conditioning. The procedure instead evaluates \\(-v(y,0;y,c)\\) at each reset. For spatially constant \\(v(t)=t^2(1-t)\\), its primitive is \\(F(t)=t^3/3-t^4/4\\). The reset procedure is stationary because \\(v(0)=0\\), while the reversed solution is \\(F(1-u)\\). This example also has zero terminal velocity. The inverse-flow justification fails; step refinement does not fix the time reset. The displayed generated videos are not disputed.","primaryUrl":"https://arxiv.org/html/2607.27924","noteUrl":"/review?paper=review-neurips-2026-149489","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149490","title":"A Differentiable 3D Scene Graph Metric via Contextual Hellinger Triplet Geometry","authors":["U-Chae Jun","Yoonhyuk Choi","Jiwoo Kang"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"A Differentiable 3D Scene Graph Metric via Contextual Hellinger Triplet Geometry","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2610.17705v1","noteUrl":"/review?paper=review-neurips-2026-149490","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149491","title":"PAAC: Privacy-Aware Agentic Device-Cloud Collaboration","authors":["Liangqi Yuan","Wenzhi Fang","Shiqiang Wang","Christopher Brinton"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Reversibility needs a reserved placeholder namespace","plainExplanation":"Injectivity of the value-to-token registry does not imply injectivity of text substitution. A literal token in the original text becomes indistinguishable from an inserted token; the fallback does not recheck the completed round trip.","primaryUrl":"https://arxiv.org/html/2605.08646","noteUrl":"/review?paper=review-neurips-2026-149491","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149494","title":"Learning to Continually Learn via Meta-learning Agentic Memory Designs","authors":["Yiming Xiong","Shengran Hu","Jeff Clune"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning to Continually Learn via Meta-learning Agentic Memory Designs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.07755","noteUrl":"/review?paper=review-neurips-2026-149494","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149496","title":"To Align or Not To Align: Check Your COMPASS Before You Train","authors":["Fabian Gröger","Petar Damjanović","Maria Brbic"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"To Align or Not To Align: Check Your COMPASS Before You Train","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.18638","noteUrl":"/review?paper=review-neurips-2026-149496","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149497","title":"Attention as In-Context Empirical Bayes: A Two-Stage View via Particle Dynamics","authors":["Matthew Smart","Soumya Ganguly","Nilava Metya","Alexandre V Morozov","Anirvan Sengupta"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The assumed Sobolev backward-heat bound cannot hold","plainExplanation":"The backward-heat estimate assumes a uniform Sobolev bound that high frequencies violate. A separate one-shot attention claim retains smoothing bias even with infinitely many tokens. Two local sign and time-index errors also need correction, while the source uses the correct score direction elsewhere.","primaryUrl":"https://arxiv.org/html/2605.29351","noteUrl":"/review?paper=review-neurips-2026-149497","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","missing_assumption","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-149498","title":"SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework","authors":["Tianshu Wu","Xiangqi Kong","Yue Chen","Qize Yu","Hang Ye","Jia Li","Yizhou Wang","Hao Dong"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.20373","noteUrl":"/review?paper=review-neurips-2026-149498","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149499","title":"Agentic Multi-Turn Reasoning: A Fairness Approach","authors":["Thanh-Dat Truong","Sankalp Pandey","Hugh Churchill","Jackson Cothren","Marios Savvides","Khoa Luu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"An unrestricted trajectory tilt need not be realizable by the planner","plainExplanation":"A Gibbs tilt over unrestricted trajectories need not be realizable by a planner with fixed stochastic tools. The deterministic-tool experimental setting avoids that obstruction.","primaryUrl":"https://arxiv.org/html/2609.33323","noteUrl":"/review?paper=review-neurips-2026-149499","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149504","title":"TopoCurve: Geometry-Aware Topology Reasoning via Bézier Curves in Autonomous Driving","authors":["Mihai Bogdan Deaconu","Laura Diosan"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Uniform curve parameter does not imply curvature-aware sampling","plainExplanation":"Uniform Bézier parameter sampling does not preferentially sample high curvature; its arc-length density depends on curve speed instead.","primaryUrl":"https://arxiv.org/html/2610.09118","noteUrl":"/review?paper=review-neurips-2026-149504","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149506","title":"UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence","authors":["Hongwei Zhang","qiqiang zhong","Jiangxia Cao","Junfeng Shu","Yiyang Lv","Huanjie Wang","Liwei Guan","Jing Yao","Yiyu Wang","Liu Zhaojie","Han Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09040","noteUrl":"/review?paper=review-neurips-2026-149506","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149508","title":"PhyProbe: Rethinking Physical Consistency Evaluation in Generated Videos","authors":["Max KU","Jiaojiao Fan","Zekun Hao","Francesco Ferroni","Heng Wang","Wenhu Chen","Ming-Yu Liu","Prithvijit Chattopadhyay"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"PhyProbe: Rethinking Physical Consistency Evaluation in Generated Videos","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.38377","noteUrl":"/review?paper=review-neurips-2026-149508","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149509","title":"Scalable Adaptation of 3D Geometric Foundation Models via Weak Supervision from Internet Video","authors":["Zihui Gao","Ke Liu","Jiawang Bian","Donny Y. Chen","Duochao Shi","Guosheng Lin","Hao Chen","Chunhua Shen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Scalable Adaptation of 3D Geometric Foundation Models via Weak Supervision from Internet Video","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.07891","noteUrl":"/review?paper=review-neurips-2026-149509","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149513","title":"DT-PBO: an Interpretable Tree-based Surrogate Model for Preferential Bayesian Optimization","authors":["Thomas Quadt","Nick Leenders","Roy Lindelauf","Herman Monsuur","Mark Voskuijl","Joost van Oijen","Boris Cule"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A fixed comparison sign does not make a noisy label uninformative","plainExplanation":"The gap fixes the sign of \\(2\\gamma+h(x)-h(x^{\\prime})\\). Under the stated probit likelihood, however, the label probability is \\(\\Phi((2\\gamma+h(x)-h(x^{\\prime}))/(\\sqrt2\\sigma))\\), which varies strictly with the residual at finite \\(\\sigma\\). Such comparisons can therefore inform within-child utilities and splits even when every noiseless ordering is unchanged. The recursive-exclusion justification is unsupported for the noisy model. Empirical retention tests and final leaf fitting do not by themselves establish that corollary.","primaryUrl":"https://arxiv.org/html/2512.14263","noteUrl":"/review?paper=review-neurips-2026-149513","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149514","title":"Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring","authors":["Yuxin Wang","Dennis Frauen","Jonas Schweisthal","Maresa Schröder","Stefan Feuerriegel"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The final score-bias regrouping reverses selected terms","plainExplanation":"Grouping its preceding three braces gives coefficient \\(\\nu_0-\\nu_1-\\gamma\\) on \\(\\hat\\xi-\\xi\\), and \\(\\nu_1-\\hat\\nu_1\\) on \\(\\hat\\xi\\). The final display instead uses \\(\\nu_1-\\gamma-\\nu_0\\) and \\(\\hat\\nu_1-\\nu_1\\). Directly, expected score minus oracle is \\((\\pi/\\hat\\pi-1)(m-\\hat m)\\), with \\(m=(1-\\xi)\\nu_0+\\xi(\\nu_1+\\gamma)\\). Squaring a sum does not erase selective sign changes among its terms. The local algebra is inconsistent. Double robustness and a product-rate bound can survive the corrected expression.","primaryUrl":"https://arxiv.org/html/2510.13397","noteUrl":"/review?paper=review-neurips-2026-149514","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149521","title":"Function graph transformers universally approximate operators between  function spaces","authors":["Takashi Furuya","S D Mis","Ivan Dokmanić","Maarten V. de Hoop","Matti Lassas"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Function graph transformers universally approximate operators between  function spaces","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17968","noteUrl":"/review?paper=review-neurips-2026-149521","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149524","title":"APS: Bias-Controlled Adaptive Prototype Simulation for Population-Scale LLM Agents","authors":["Quan ZHENG","Yan Gao","shaobin he","Haoxiang Guan","Yuanhe Tian","Jie Feng","Ming Wang","Shuxin Zheng","Zhen Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"APS: Bias-Controlled Adaptive Prototype Simulation for Population-Scale LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.27419","noteUrl":"/review?paper=review-neurips-2026-149524","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149526","title":"Enhancing Tabular Learners with Context-Aware Semantic Embeddings","authors":["Günther Schindler","Maximilian Schambach","Johannes Höhne"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Enhancing Tabular Learners with Context-Aware Semantic Embeddings","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.03565","noteUrl":"/review?paper=review-neurips-2026-149526","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149527","title":"Strengthening LLMs for Tabular Prediction with Structural Priors","authors":["Pengxiang Cai","Zihao Gao","Wanchen Lian","Guocong Li","Jintai Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strengthening LLMs for Tabular Prediction with Structural Priors","plainExplanation":"The variance, unbiasedness and convergence guarantees need separate assumptions. Group centering alone establishes none of them for the configured normalized update.","primaryUrl":"https://arxiv.org/html/2510.17385","noteUrl":"/review?paper=review-neurips-2026-149527","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149529","title":"Controllable Molecular Generative Foundation Models","authors":["Yihan Zhu","Yuhan Liu","Weijiang Li","Tengfei Luo","Meng Jiang"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Aggregate compression cannot be substituted into a per-molecule bound","plainExplanation":"Equation (17) uses each molecule’s own compression. Two equally weighted \\(60\\)-atom molecules with motif counts \\(1\\) and \\(20\\) have aggregate compression \\(120/21>5.5\\), but aggregate motif-to-atom factor ratio \\((2+1180)/(2\\cdot3660)\\approx0.161475\\). Connected chains admit both partitions. The valid per-molecule inequality cannot be made uniform by substituting the aggregate ratio into its nonlinear expression. The average-to-uniform implication fails. The per-molecule theorem remains valid, and the actual tokenizer’s aggregate ratio is not shown to exceed the reported figure.","primaryUrl":"https://arxiv.org/html/2605.15354","noteUrl":"/review?paper=review-neurips-2026-149529","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149530","title":"LLM Active Alignment: A Nash Equilibrium Perspective","authors":["Tonghan Wang","Yuqi Pan","Xinyi Yang","Xinrui Song","Yanchen Jiang","Milind Tambe","David Parkes"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LLM Active Alignment: A Nash Equilibrium Perspective","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.06836","noteUrl":"/review?paper=review-neurips-2026-149530","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149532","title":"Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration","authors":["Haisen He","XIANGYU ZOU","SongLin Dong","Heng Li","Yihong Gong","Zhiheng Ma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06127","noteUrl":"/review?paper=review-neurips-2026-149532","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149533","title":"CineOrchestra: Unified Entity-Centric Conditioning for Cinematic Video Generation","authors":["Sharath Girish","Tsai-Shien Chen","Zhikang Dong","Mukesh Singhal","Hao Chen","Sergey Tulyakov","Aliaksandr Siarohin"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"CineOrchestra: Unified Entity-Centric Conditioning for Cinematic Video Generation","plainExplanation":"The appendix proves an upper bound whose value is independent of duration, but that bound need not be attained at every duration. A two-frequency example changes the peak while preserving the normalized Frobenius norm. Exact or uniform approximate peak stability needs additional assumptions.","primaryUrl":"https://arxiv.org/html/2606.13768","noteUrl":"/review?paper=review-neurips-2026-149533","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149534","title":"SapiensID 2.0: Aligning Human Recognition Foundation Models with Human Perception","authors":["Yiyang Su","Jie Zhu","Feng Liu","Anil Jain","Xiaoming Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SapiensID 2.0: Aligning Human Recognition Foundation Models with Human Perception","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.10497","noteUrl":"/review?paper=review-neurips-2026-149534","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149535","title":"Revisiting Value Iteration: Unified Analysis of Discounted and Average-Reward Cases","authors":["Arsenii Mustafin","Xinyi Sheng","Dominik Baumann"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The own-state action coefficient subtracts the discount instead of one","plainExplanation":"Take the deterministic cycle \\(1\\to2\\to3\\to1\\), one action per state and constant reward \\(1/2\\). Its sole policy is uniquely optimal and unichain, but periodic. Centered rewards and bias vanish. Starting from \\(x_0=v_0/C=(2,-1,-1)\\), the anchor vanishes and \\(x_t=\\gamma^tP^tx_0\\). Since \\(P^9=I\\), \\(\\operatorname{sp}(x_9)=\\gamma^9\\operatorname{sp}(x_0)=3\\gamma^9\\), contradicting every extra factor below one. At \\(\\gamma=1\\), values cycle.\n\nA valid restricted argument needs a horizon \\(L\\) and common column \\((P^{*L})_{ij}\\geq b>0\\) for every starting state. If every nonoptimal action also has normalized Bellman gap at most \\(-\\delta<0\\), positive initial span permits optimal-kernel weight \\(\\alpha=\\min\\{1,\\delta/\\operatorname{sp}(x_0)\\}\\). Each interpolated \\(L\\)-step product then has common-column mass at least \\(\\alpha^Lb\\), giving contraction factor \\(\\gamma^L(1-\\alpha^Lb)\\). The cycle refutes the extra strict span factor and average-reward value convergence. Its only policy is optimal, so it does not refute the policy-optimality corollary or ordinary discounted contraction. The restricted repair requires its stated hypotheses and proof.","primaryUrl":"https://arxiv.org/html/2510.23914","noteUrl":"/review?paper=review-neurips-2026-149535","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149536","title":"GAIA: Geometry-Adaptive Operator Learning for Forward and Inverse Problems","authors":["Meenakshi Krishnan","Pranav Pulijala","Ke Chen","Haizhao Yang","Ramani Duraiswami"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"GAIA: Geometry-Adaptive Operator Learning for Forward and Inverse Problems","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.01128","noteUrl":"/review?paper=review-neurips-2026-149536","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149537","title":"Mult-DPO: Multinomial Direct Preference Optimization for Recommender Systems","authors":["Yaochen Zhu","Harald Steck","James McInerney","Aditya Sinha","Yinhan He","Nathan Kallus","Jundong Li"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Mult-DPO: Multinomial Direct Preference Optimization for Recommender Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.10078","noteUrl":"/review?paper=review-neurips-2026-149537","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149538","title":"STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation","authors":["Won June Cho","Daeky Jeong","Hyeongyeol Lim","Hongjun Yoon"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation","plainExplanation":"Observation noise reweights the endpoint law without removing its singular support. The revised fixed-target loss comparison is valid and remains separate from this density error.","primaryUrl":"https://arxiv.org/html/2606.07036","noteUrl":"/review?paper=review-neurips-2026-149538","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149539","title":"DriveDreamer-Policy: A Geometry-Grounded World-Action Model for Unified Generation and Planning","authors":["Yang Zhou","Xiaofeng Wang","Hao Shao","Letian Wang","Guosheng Zhao","Jiangnan Shao","Jiagang Zhu","Tingdong Yu","Zheng Zhu","Guan Huang","Steven Waslander"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"DriveDreamer-Policy: A Geometry-Grounded World-Action Model for Unified Generation and Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.01765","noteUrl":"/review?paper=review-neurips-2026-149539","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149540","title":"What Fits (Into Few Tokens) Doesn't Overfit:  Compression and Generalization in ML Research Agents","authors":["Martin Bertran","Aaron Roth","Steven Wu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"What Fits (Into Few Tokens) Doesn't Overfit:  Compression and Generalization in ML Research Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11045","noteUrl":"/review?paper=review-neurips-2026-149540","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149541","title":"Nearly Optimal Attention Coresets","authors":["Alexandr Andoni","Eldar Kleiner","Edo Liberty"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Nearly Optimal Attention Coresets","plainExplanation":"The centering reduction loses a factor in the exponent for uncentered caches. Separate local corrections retain the series’ logarithmic factor and make the degree-zero and degree-one constraints well defined.","primaryUrl":"https://arxiv.org/html/2605.05602","noteUrl":"/review?paper=review-neurips-2026-149541","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149542","title":"M$^3$: Reframing Training Measures for Discretized Physical Simulations","authors":["Yuan Mei","Xingyu Song","Xiaowen Song","Naoya Takeishi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"M$^3$: Reframing Training Measures for Discretized Physical Simulations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08843","noteUrl":"/review?paper=review-neurips-2026-149542","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149547","title":"Discretizing Continuous Time Series for Imputation with Masked Diffusion Training","authors":["Dongbin Kim","Seungyun Lee","Geonwoo Shin","Jaewook Lee"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The masked-diffusion cross-entropy weight has the wrong sign","plainExplanation":"The masked-diffusion NELBO display gives positive cross-entropy a negative weight. The appendix supplies the correctly signed schedule factor.","primaryUrl":"https://arxiv.org/html/2608.19119","noteUrl":"/review?paper=review-neurips-2026-149547","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149548","title":"On the Recoverability of Causal Relations from Bulk Gene Expression Data","authors":["Gongxu Luo","Boyang Sun","Kun Zhang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"On the Recoverability of Causal Relations from Bulk Gene Expression Data","plainExplanation":"An invertible nonlinear transformation of cell means preserves both properties in the affine Gaussian subclass. The necessity claim for linear aggregation is therefore broader than the general consistency definition supports. Theorems that explicitly assume linear aggregation remain separate, and Appendix D.2 already identifies nonlinear aggregation as an open case.","primaryUrl":"https://arxiv.org/html/2606.00568","noteUrl":"/review?paper=review-neurips-2026-149548","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149549","title":"Geometry-Centered 3D Latent World Models for Growing Surfaces","authors":["Xiaoyi Liu","Hao Tang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Geometry-Centered 3D Latent World Models for Growing Surfaces","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2506.03173","noteUrl":"/review?paper=review-neurips-2026-149549","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149550","title":"Soft Token Alignment for Cross-Lingual Reasoning","authors":["Jiayi He","Jungsoo Park","Wei \"Coco\" Xu","Alan Ritter"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Soft Token Alignment for Cross-Lingual Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.26466","noteUrl":"/review?paper=review-neurips-2026-149550","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149551","title":"Information-Theoretic Generalization Bounds for Sequential Decision Making","authors":["Futoshi Futami","Masahiro Fujisawa"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Information-Theoretic Generalization Bounds for Sequential Decision Making","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12190","noteUrl":"/review?paper=review-neurips-2026-149551","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149552","title":"Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models","authors":["Xiangwen Wang","Ananth Balashankar","Varun Chandrasekaran"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.11149","noteUrl":"/review?paper=review-neurips-2026-149552","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149553","title":"Coherent Hierarchical Multi-Label  Learning to Defer for Medical Imaging","authors":["Joshua Strong","Pramit Saha","Emma Sun","Helen Higham","Alison Noble"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Coherent Hierarchical Multi-Label  Learning to Defer for Medical Imaging","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.02734","noteUrl":"/review?paper=review-neurips-2026-149553","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149554","title":"FA-LAM: Focus-Aware Large Avatar Model for One-Shot 4D Animatable Gaussian Head","authors":["Yingdong Hu","Yisheng He","Yiming Jiang","Zehong Lin","Steven Hoi","Jun Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FA-LAM: Focus-Aware Large Avatar Model for One-Shot 4D Animatable Gaussian Head","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.20922","noteUrl":"/review?paper=review-neurips-2026-149554","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149555","title":"Agentic Video Editing from Underspecified Requests","authors":["Yongsheng Yu","Ziyun Zeng","Zhiyuan Xiao","Zhenghong Zhou","Hang Hua","Wei Xiong","Jiebo Luo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Agentic Video Editing from Underspecified Requests","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18748","noteUrl":"/review?paper=review-neurips-2026-149555","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149556","title":"OxyGen: Unified KV Cache Management for VLA Inference under Multi-Task Parallelism","authors":["Xiangyu Li","Huaizhi Tang","Xin Ding","Weijun Wang","Ting Cao","Yunxin Liu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"OxyGen: Unified KV Cache Management for VLA Inference under Multi-Task Parallelism","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.14371","noteUrl":"/review?paper=review-neurips-2026-149556","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149558","title":"Reformulating Neural Operators in $d+1$ Dimensions for Embedding Evolution","authors":["Haoze Song","Zhihao Li","Xiaobo Zhang","Zecheng Gan","Zhilu Lai","Wei Wang"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"An additional kernel dimension does not make the operator self-adjoint","plainExplanation":"The real adjoint kernel is \\(\\kappa(y,x,p^{\\prime},p)\\), swapping both input/output coordinate pairs. Equality to the forward kernel requires that symmetry, which equation (4) does not impose. A bounded two-cell kernel realizes \\(\\begin{pmatrix}0&1\\\\0&0\\end{pmatrix}\\), whose transpose differs and whose forward reuse violates the defining inner-product identity. The general forward-equals-adjoint claim is false. An optionally symmetric architecture can share kernels, and the local transpose correction does not refute the empirical method.","primaryUrl":"https://arxiv.org/html/2505.11766","noteUrl":"/review?paper=review-neurips-2026-149558","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149562","title":"ProCARE: Real-World Study Automation via Profile-Grounded Evidence Contracts","authors":["Zhaoyu Fan","Bowen Han","Yanwei Ren","Jincheng Ou","Siyi Cao","Zehua He","Kaiyu Huang","Pengcheng He","Shanshan Ni","Chengchen Gong","Qingjiang Shi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ProCARE: Real-World Study Automation via Profile-Grounded Evidence Contracts","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.16175","noteUrl":"/review?paper=review-neurips-2026-149562","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149563","title":"Overcoming Kernel Redundancy for Scaling Logic Gate Networks","authors":["Sejin Park","Hongjae Lee","Changwoo Han","Seung-Won Jung"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Overcoming Kernel Redundancy for Scaling Logic Gate Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01069","noteUrl":"/review?paper=review-neurips-2026-149563","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149570","title":"When to Align, When to Predict: A Phase Diagram for Multimodal Learning","authors":["Ilay Kamai","Hugues Van Assel","Aviv Regev","Hagai B Perets","Randall Balestriero"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"When to Align, When to Predict: A Phase Diagram for Multimodal Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11190","noteUrl":"/review?paper=review-neurips-2026-149570","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149571","title":"Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift Correction","authors":["Jiahe Guo","Xiangran Guo","Jiaxuan Chen","Weixiang Zhao","Yanyan Zhao","Yutai Hou","Qianchao Wang","Dandan Tu","Bing Qin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift Correction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18104","noteUrl":"/review?paper=review-neurips-2026-149571","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149573","title":"Active Probabilistic Reasoning in Humans and LLMs","authors":["Gonçalo Guiomar","Elia Torre","Pehuen Moure","Victoria Shavina","Mario Giulianelli","Shih-Chii Liu","Valerio Mante"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Observed red fraction is not the unequal-sample posterior ranking","plainExplanation":"Observed red fractions and posterior biased-arm probabilities can select different arms when sample counts differ.","primaryUrl":"https://arxiv.org/html/2602.08693","noteUrl":"/review?paper=review-neurips-2026-149573","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149574","title":"Hide to See: Reasoning-prefix Masking for Visual-anchored Thinking in VLM Distillation","authors":["Seonghoon Yu","Dongjun Nam","Byung-Kwan Lee","Jeany Son"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Hide to See: Reasoning-prefix Masking for Visual-anchored Thinking in VLM Distillation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11651","noteUrl":"/review?paper=review-neurips-2026-149574","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149575","title":"Selective Off-Policy Reference Tuning with Plan Guidance","authors":["Duc A Le","Tien-Phat Nguyen","Thien H Nguyen","Linh Ngo","Trung Le"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Selective Off-Policy Reference Tuning with Plan Guidance","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11505","noteUrl":"/review?paper=review-neurips-2026-149575","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149576","title":"SAMoR: Motion Modelling for Articulated Objects of Any Skeleton and Topology","authors":["Yuhao Zhang","Gerard Pons-Moll","Tolga Birdal"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"SAMoR: Motion Modelling for Articulated Objects of Any Skeleton and Topology","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.02148","noteUrl":"/review?paper=review-neurips-2026-149576","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149577","title":"ECHO: Continuous Hierarchical Memory for Vision-Language-Action Models","authors":["Yanbin Hu","Jin Cui","Jiayi Lu","Ruixuan Yang","Jun Ye","Boran Zhao","Xingyu Chen","Xuguang Lan","Pengju Ren"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The origin-pointing tangent selects the inward cone","plainExplanation":"The cone angle points toward the origin and can reject an outward descendant on the same hyperbolic ray.","primaryUrl":"https://arxiv.org/html/2605.10993","noteUrl":"/review?paper=review-neurips-2026-149577","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149578","title":"Query-Limited Community Recovery in Stochastic Block Models","authors":["Sabyasachi Basu","MANUJ MUKHERJEE","Lutz Oettershagen","Suhas Thejaswi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Query-Limited Community Recovery in Stochastic Block Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02055","noteUrl":"/review?paper=review-neurips-2026-149578","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149580","title":"Parameter Exploration for RLVR via Variational Learning","authors":["Vatsal Venkatkrishna","Nico Daheim","Iryna Gurevych"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Parameter Exploration for RLVR via Variational Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.09805","noteUrl":"/review?paper=review-neurips-2026-149580","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149582","title":"Aligning LLMs with Biomedical Knowledge using Balanced Fine-Tuning","authors":["Zhenchao Tang","Fang Wang","Haohuai He","Jiale Zhou","Tianxu Lv","Jun Zhu","Shou Z Chen","Minghao Yang","Yu Wang","Jiayang Wu","Yidong Song","Yaokun Li","Jiehui Huang","Jun Zhou","Bing He","Jianhua Yao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Aligning LLMs with Biomedical Knowledge using Balanced Fine-Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.21075","noteUrl":"/review?paper=review-neurips-2026-149582","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149583","title":"When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors","authors":["Adarsh Jamadandi","Nicolas Keriven","Aline Roumy"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.13765","noteUrl":"/review?paper=review-neurips-2026-149583","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149585","title":"Robust Conditional Conformal Prediction via Branched Normalizing Flow","authors":["Rui Xu","Xingyuan Chen","Wenxing Huang","Minxuan Huang","Weiyan Chen","Sihong Xie","Hui Xiong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Robust Conditional Conformal Prediction via Branched Normalizing Flow","plainExplanation":"An optimal joint transport plan can fail to couple the full conditional label laws, invalidating Theorem 3.3’s joint-to-conditional bound. Separately, the empirical Wasserstein concentration bound omits the scale of the ground metric.","primaryUrl":"https://arxiv.org/html/2605.01868","noteUrl":"/review?paper=review-neurips-2026-149585","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","invalid_inference","missing_assumption"],"classifications":["critical"]},{"slug":"review-neurips-2026-149587","title":"FM-ChangeNet: Learning Change through Pathwise Feature Transport","authors":["Roie Kazoom","George Leifman","Genady Beryozkin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Endpoint reconstruction uses the wrong interpolation time","plainExplanation":"The endpoint reconstruction uses the final path state instead of the current interpolated state, so it fails even for the exact transport field.","primaryUrl":"https://arxiv.org/html/2607.04750","noteUrl":"/review?paper=review-neurips-2026-149587","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149588","title":"Memorize When Needed: Decoupled Memory Control for Spatially Consistent Long-Horizon Video Generation","authors":["Yanjun Guo","Zhengqiang ZHANG","Pengfei Wang","Xinyue Liang","Zhiyuan Ma","Lei Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Memorize When Needed: Decoupled Memory Control for Spatially Consistent Long-Horizon Video Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.18215","noteUrl":"/review?paper=review-neurips-2026-149588","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149589","title":"Trie-Aware Transformers for Generative Recommendation","authors":["Zhenxiang Xu","Sirui Chen","Yong He","Jieyu Yang","Chuan Yuan","Ke Ding","Jing Cao","Can Wang","Jiawei Chen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Trie-Aware Transformers for Generative Recommendation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.21677","noteUrl":"/review?paper=review-neurips-2026-149589","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149593","title":"Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design","authors":["Lingxiao Li","Haobo Zhang","Bin Chen","Jiayu Zhou"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.11894","noteUrl":"/review?paper=review-neurips-2026-149593","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"coupled-reward-machines-stochastic-ordering","title":"Reinforcement Learning for Long-Horizon Unordered Tasks: From Boolean to Coupled Reward Machines","authors":["Kristina Levina","Nikolaos Pappas","Athanasios Karapantelakis","Aneta Vulgarakis Feljan","Jendrik Seipp"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Stochastic reward shaping can reverse policy preference","plainExplanation":"A stochastic optimal policy need not finish at its expected duration. The transformation penalizes its short and long realizations. In a finite example, the originally better action becomes worse under the shaped reward even when the expected durations are known exactly.","primaryUrl":"https://arxiv.org/abs/2510.27329v2","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-149596","title":"Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift","authors":["Zhihua Wei","Qiang Li","Jian Ruan","Zhenxin Qin","蕾蕾 温","Ruiyang Qin","Qingzhuo Wang","Dongrui Liu","Wen Shen"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.17372","noteUrl":"/review?paper=review-neurips-2026-149596","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149598","title":"Mildly Overparameterized ReLU Networks on Orthogonal Data: Incremental Learning and Implicit Bias","authors":["James Town","Etienne Boursier","Ben Lewis","Matthias Englert","Ranko Lazic"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The differential upper bound strengthens the damping incorrectly","plainExplanation":"The proof changes an error allowance into additional damping. The subsequent differential bound needs the opposite sign and revised constants.","primaryUrl":"https://arxiv.org/html/2605.27097","noteUrl":"/review?paper=review-neurips-2026-149598","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149601","title":"Euclidean Embedding of Data Using Local Distances","authors":["Dimitrios Arampatzis"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Euclidean Embedding of Data Using Local Distances","plainExplanation":"From the prescribed initialization, the alternating embedding updates can remain at a positive-error fixed point even when a zero-error embedding is feasible. Exact block minimization guarantees nonincreasing objective values, rather than convergence to a global minimum.","primaryUrl":"https://arxiv.org/html/2605.19243","noteUrl":"/review?paper=review-neurips-2026-149601","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-149606","title":"A Geometric Perspective on Reward Function Updates in Inverse Reinforcement Learning","authors":["Anish Abhijit Diwan","Jan Peters","Oleg Arenz"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"A Geometric Perspective on Reward Function Updates in Inverse Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.14041","noteUrl":"/review?paper=review-neurips-2026-149606","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149611","title":"Dyna-Style Safety Augmented Reinforcement Learning: Staying Safe in the Face of Uncertainty","authors":["Artur Eisele","Bernd Frauenknecht","Friedrich Solowjow","Sebastian Trimpe"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Dyna-Style Safety Augmented Reinforcement Learning: Staying Safe in the Face of Uncertainty","plainExplanation":"The hyperplane map is bijective only on its actual half-open offset range. Separately, the critic determines a discounted moment of failure time; its logarithmic inverse is a certainty equivalent, not the mean. The report gives the corrected range and the valid Jensen relation.","primaryUrl":"https://arxiv.org/html/2604.25508","noteUrl":"/review?paper=review-neurips-2026-149611","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149612","title":"FRInGe: Distribution-Space Integrated Gradients with Fisher–Rao Geometry","authors":["Gabriele Martino","Sebastian Tschiatschek"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"FRInGe: Distribution-Space Integrated Gradients with Fisher–Rao Geometry","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06404","noteUrl":"/review?paper=review-neurips-2026-149612","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149616","title":"MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments","authors":["Yicheng Gao","Xiaolin Zhou","Yahan Li","Yue Zhao","Ruishan Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07058","noteUrl":"/review?paper=review-neurips-2026-149616","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149618","title":"Quantile-Coupled Flow Matching for Distributional Reinforcement Learning","authors":["Michael Groom","Victor-Alexandru Darvariu","Lars Kunze","James Wilson","Nick Hawes"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Quantile-Coupled Flow Matching for Distributional Reinforcement Learning","plainExplanation":"The learned flow can leave the paths used for training, so Proposition 1 needs a flow-stability factor. A shared quantile input also need not make an unrestricted learned map monotone, leaving shortcut loss as a coupling upper bound rather than an identity. The Bellman-contraction result remains conditional on an actual Wasserstein projection-error bound.","primaryUrl":"https://arxiv.org/html/2605.08515","noteUrl":"/review?paper=review-neurips-2026-149618","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149619","title":"Agent-Native Research Artifacts","authors":["Jiachen Liu","Jiaxin Pei","Jintao Huang","Chenglei Si","Ao Qu","Robert Tang","Runyu Lu","Lichang Chen","Xiaoyan Bai","Haizhong Zheng","Shirui Chen","Zhiyang Chen","Haojie Ye","Yujuan Fu","Zexue He","Zijian Jin","Zhenyu Zhang","Shangquan Sun","Maestro Harmon","Dianzhuo Wang","Qian-Ze Zhu","Jiachen Sun","Mingyuan Wu","Baoyu Zhou","Chenyu You","Shijian Lu","Yiming Qiu","Fan Lai","Yuan Yuan","Yao Li","Junyuan Hong","Ruihao Zhu","Beidi Chen","Alex Pentland","Ang Chen","Mosharaf Chowdhury","Zechen Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Agent-Native Research Artifacts","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.24658","noteUrl":"/review?paper=review-neurips-2026-149619","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149620","title":"M3-AD: Reflection-aware Multi-modal, Multi-category, and Multi-dimensional Benchmark and Framework for Industrial Anomaly Detection","authors":["Chao Huang","Yanhui Li","Hongxi Huang","Yunkang Cao","Wei Wang","Jie Wen","Zhihua Wang","Wenqi Ren","XIAOCHUN CAO"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"M3-AD: Reflection-aware Multi-modal, Multi-category, and Multi-dimensional Benchmark and Framework for Industrial Anomaly Detection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.00055","noteUrl":"/review?paper=review-neurips-2026-149620","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149621","title":"Mirror Descent Beyond Euclidean Stability:  An Exponential Separation in Initialization Sensitivity","authors":["Shira Vansover-Hager","Matan Schliserman","Ofir Schlisselberg","Tomer Koren"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mirror Descent Beyond Euclidean Stability:  An Exponential Separation in Initialization Sensitivity","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11431","noteUrl":"/review?paper=review-neurips-2026-149621","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149623","title":"CAPER: Clause-Aligned Process Supervision for Text-to-SQL","authors":["Lujie Ban","Jiasheng Shi","Jinyang Li","Xiaolin Han","Tsz Nam Chan","Chenhao Ma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CAPER: Clause-Aligned Process Supervision for Text-to-SQL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03327","noteUrl":"/review?paper=review-neurips-2026-149623","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149624","title":"Mechanisms of Misgeneralization in Physical Sequence Modeling","authors":["Kento Nishi","Raphael Tang","Karun Kumar","Core Francisco Park","Hidenori Tanaka"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Mechanisms of Misgeneralization in Physical Sequence Modeling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.20299","noteUrl":"/review?paper=review-neurips-2026-149624","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149625","title":"Adaptively Incorporating Directional Hints into Zeroth-Order Optimization","authors":["Alexander Ryabchenko","Jian Qian","Wenlong Mou"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Adaptively Incorporating Directional Hints into Zeroth-Order Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.08277","noteUrl":"/review?paper=review-neurips-2026-149625","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149626","title":"Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts","authors":["Seungmin Oh","Seunghun Kang","Jongbin Ryu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06147","noteUrl":"/review?paper=review-neurips-2026-149626","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149627","title":"Sparse Koopman Autoencoders Identify Local Dynamical Regimes in Multibasin Systems","authors":["Aidan Li","Uday Kiran Reddy Tadipatri","Mahan Fathi","Sarath Chandar","Ross Goroshin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Sparse Koopman Autoencoders Identify Local Dynamical Regimes in Multibasin Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.29057","noteUrl":"/review?paper=review-neurips-2026-149627","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149628","title":"Towards Identifiable Latent Additive Noise Models","authors":["Yuhang Liu","Zhen Zhang","Dong Gong","Erdun Gao","Biwei Huang","Mingming Gong","Anton van den Hengel","Kun Zhang","Prof Javen Qinfeng Shi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Environment subtraction evaluates sufficient statistics at different noises","plainExplanation":"At one observed \\(x\\), equation (35) has \\(n_u=(h^u)^{-1}(f^{-1}(x))\\). Subtraction gives \\(T(n_\\ell)^\\top\\eta_\\ell-T(n_0)^\\top\\eta_0\\), not \\(T(n)^\\top(\\eta_\\ell-\\eta_0)\\). A unit Jacobian determinant does not identify these noise arguments. For \\(z_1=n_1,z_2=a(u)n_1+n_2\\), the same observed \\(z\\) implies different noises as \\(a(u)\\) changes. A reversed Gaussian triangular representation can be a permutation at \\(a=0\\) and a mixture at \\(a=1\\), contradicting the claimed fixed monomial noise relation. The decoder relation being environment independent does not establish an environment-independent exogenous-noise relation. This is a Step I proof gap, not a disproof of the final causal-variable conclusion.","primaryUrl":"https://arxiv.org/html/2403.15711","noteUrl":"/review?paper=review-neurips-2026-149628","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149629","title":"Multi-Marginal Couplings for Metropolis--Hastings","authors":["Truong Buu Phan","Gergely Flamich","Ashish Khisti","Shahab Asoodeh"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Multi-Marginal Couplings for Metropolis--Hastings","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12807","noteUrl":"/review?paper=review-neurips-2026-149629","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149630","title":"Optimal scaling laws in learning hierarchical multi-index models","authors":["Leonardo Defilippis","Florent Krzakala","Bruno Loureiro","Antoine Maillard"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Optimal scaling laws in learning hierarchical multi-index models","plainExplanation":"The written preprocessing assumptions allow an estimator that ignores the labels entirely. Recovery needs stronger information and sign conditions than those assumptions provide.","primaryUrl":"https://arxiv.org/html/2602.05846","noteUrl":"/review?paper=review-neurips-2026-149630","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-149632","title":"Stitched Value Model for Diffusion Alignment","authors":["Hyojun Go","Hyungjin Chung","Prune Truong","Goutam Bhat","Li Mi","Zhaochong An","Zixiang Zhao","Dominik Narnhofer","Serge Belongie","Federico Tombari","Konrad Schindler"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Stitched Value Model for Diffusion Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19804","noteUrl":"/review?paper=review-neurips-2026-149632","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149634","title":"Publishing Below-Threshold Triangle Counts under Local Weight Differential Privacy","authors":["Kevin Pfisterer","Quentin Hillebrand","Vorapong Suppakitpaisarn"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Publishing Below-Threshold Triangle Counts under Local Weight Differential Privacy","plainExplanation":"The local assignment and variance calculations have explicit corrections. The lower-bound theorem additionally needs nontrivial graph families and a privacy regime that respects finite-range saturation.","primaryUrl":"https://arxiv.org/html/2601.01710","noteUrl":"/review?paper=review-neurips-2026-149634","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","lemma_error","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-149635","title":"Structured Sparse Memory for Recurrent Reasoning","authors":["Zixuan Zhao","Sam Wheeler","Neil Getty","Xiaotian Duan","Rick Stevens","Fangfang Xia"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Structured Sparse Memory for Recurrent Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33270","noteUrl":"/review?paper=review-neurips-2026-149635","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149637","title":"MentalHospital: A Virtual Environment for Evaluating Psychiatric Clinical Encounters","authors":["Yuming Yang","Xiao Sun","Yuanwei Zou","Zhengxiao Wu","Jiang Zhong","Haoyang Zeng","Jingwang Huang","Kaiwen Wei"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MentalHospital: A Virtual Environment for Evaluating Psychiatric Clinical Encounters","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.08257","noteUrl":"/review?paper=review-neurips-2026-149637","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149638","title":"Early Prediction of Future Behavioral Strategy from Process Traces","authors":["Robert Kasumba","Dennis Barbour","Chien-Ju Ho"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Early Prediction of Future Behavioral Strategy from Process Traces","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.30550","noteUrl":"/review?paper=review-neurips-2026-149638","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149643","title":"Near-Optimal Decentralized Stochastic Convex Optimization over Networks","authors":["Amit Attia","Nitai Kluger","Tomer Koren"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The dimension scaling needs an additive constant","plainExplanation":"The all-parameter dimension scaling can tend to zero while the constructed dimension remains a positive integer. It needs an additive constant or an explicit restriction to the later asymptotic regime. The objective lower bound and the later large-communication regime are not refuted. This correction concerns the uniform dimension scaling outside that regime.","primaryUrl":"https://arxiv.org/html/2606.04757","noteUrl":"/review?paper=review-neurips-2026-149643","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149646","title":"AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers","authors":["Ruibin Min","Yexin Liu","Aimin PAN","Changsheng Lu","Jiafei Wu","kelu Yao","Xiaogang Xu","Harry Yang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Gradient consistency does not establish constructive alignment","plainExplanation":"A large gradient signal-to-noise ratio measures consistency, not agreement with the denoising objective; an equally large ratio can accompany opposing gradients.","primaryUrl":"https://arxiv.org/html/2605.03317","noteUrl":"/review?paper=review-neurips-2026-149646","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149647","title":"Transfer Learning of Linear Regression with Multiple Pretrained Models: Benefiting from More Pretrained Models via Overparameterization Debiasing","authors":["Daniel Boharon","Yehuda Dar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Transfer Learning of Linear Regression with Multiple Pretrained Models: Benefiting from More Pretrained Models via Overparameterization Debiasing","plainExplanation":"The paper’s own bias-aware criterion permits beneficial transfer above the stated noise-only cutoff when there are multiple sources. For a general supplied task map, the vanishing-penalty estimate is a constrained weighted projection; the simpler Euclidean formula requires an isotropic penalty. The report keeps these general and specialized regimes distinct.","primaryUrl":"https://arxiv.org/html/2602.16531","noteUrl":"/review?paper=review-neurips-2026-149647","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149648","title":"Tokens-per-Parameter Coverage Is Critical for Robust LLM Scaling Law Extrapolation","authors":["Joshua Shay Kricheli","Alexander L. Reid","Soumajyoti Sarkar","Venkata Gandikota","Paulo Shakarian"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Tokens-per-Parameter Coverage Is Critical for Robust LLM Scaling Law Extrapolation","plainExplanation":"An upper bound on column separation does not give a matching eigenvalue order. The two-by-two condition-number approximation also cannot serve as an exact biconditional, and small Gaussian noise does not justify interchanging expectation and square root. The corrected bounds preserve the stated conclusions only under their explicit quantitative conditions.","primaryUrl":"https://arxiv.org/html/2605.08541","noteUrl":"/review?paper=review-neurips-2026-149648","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149650","title":"Instance-Optimal Estimation with Multiple LLM Judges on a Budget","authors":["Junghyun Lee","Sanghwa Kim","Yassir Jedra","Alexandre Proutiere","Se-Young Yun"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Instance-Optimal Estimation with Multiple LLM Judges on a Budget","plainExplanation":"Some admitted parameter boundaries leave no alternative means or no final samples. Confidence lower bounds also need strictly separated success events; the local display corrections do not resolve those broader conditions.","primaryUrl":"https://arxiv.org/html/2605.23362","noteUrl":"/review?paper=review-neurips-2026-149650","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","formula_error","incorrect_bound","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149651","title":"Wasserstein residuals: Learning Gradient Flows from Population Dynamics","authors":["Markus Heinonen","Yair Shenfeld","Ricardo Baptista","Daniel Waxman","Dmitry Batenkov","Tim Cooijmans","Elias Bingham"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The symmetric interaction energy differentiates twice","plainExplanation":"The interaction force omits the factor from differentiating both density arguments. The residual identity also needs a minimal continuity velocity to equal Wasserstein speed.","primaryUrl":"https://arxiv.org/html/2607.04738","noteUrl":"/review?paper=review-neurips-2026-149651","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149653","title":"Training for the Model You Return: Improving Optimization for Iterate-Averaged Language Models","authors":["Kwok C Au","Adam Block"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The corollary denominator differs from the control Gramian","plainExplanation":"Let \\(x\\) be the previous iterate, \\(a\\) its average and \\(b\\) the optimizer output. The printed update is \\(b+\\lambda(a-x)\\), whereas a convex combination is \\((1-\\lambda)b+\\lambda a\\). At \\(\\lambda=1\\), the former is \\(b+a-x\\), equal to \\(a\\) only when \\(b=x\\). The explicitly previous-iterate index prevents a post-step interpretation from reconciling them. This is a manuscript update/guarantee mismatch, not a verified implementation defect.","primaryUrl":"https://arxiv.org/html/2606.25086","noteUrl":"/review?paper=review-neurips-2026-149653","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149656","title":"Expanding the Role of Diffusion Models for Robust Classifier Training","authors":["Pin-Han Huang","Shang-Tse Chen","Hsuan-Tien (Tien) Lin"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Expanding the Role of Diffusion Models for Robust Classifier Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.19931","noteUrl":"/review?paper=review-neurips-2026-149656","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149658","title":"On the Complexity of Discounted Robust MDPs with $L_p$ Uncertainty Sets","authors":["Ali Asadi","Krishnendu Chatterjee","Alipasha Montaseri","Ali Shafiee"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"On the Complexity of Discounted Robust MDPs with $L_p$ Uncertainty Sets","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07459","noteUrl":"/review?paper=review-neurips-2026-149658","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149660","title":"Continuity Laws for Sequential Models","authors":["Annan Yu","Dongwei Lyu","N. Benjamin Erichson"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Continuity Laws for Sequential Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08539","noteUrl":"/review?paper=review-neurips-2026-149660","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149661","title":"NeuroRVQ: Multi-Scale Biosignal Tokenization for Generative Foundation Models","authors":["Konstantinos Barmpas","Na Lee","Dimitrios Chalatsis","William Raftery","Yannis Panagakis","Dimitrios Adamos","N Laskaris","Alexandros Koliousis","Dario Farina","Stefanos Zafeiriou"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The cosine loss has zero angular gradient at an antipodal prediction","plainExplanation":"For unit target \\(u\\), nonzero prediction \\(v\\), and \\(r=\\|v\\|\\), the gradient of \\(1-u^\\top v/r\\) is \\(-u/r+(u^\\top v)v/r^3\\). At \\(v=-ru\\), the terms cancel exactly. At angle \\(\\pi-\\varepsilon\\), its norm is \\(|\\sin\\varepsilon|/r\\), tending to zero. A tanh Jacobian cannot turn this zero input gradient into a nonzero one. A radial penalty can affect norm without providing angular descent at the antipode. The full temporal training method and empirical success are not refuted.","primaryUrl":"https://arxiv.org/html/2510.13068","noteUrl":"/review?paper=review-neurips-2026-149661","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149663","title":"Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory","authors":["Juno Kim","Eshaan Nichani","Denny Wu","Alberto Bietti","Jason Lee"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory","plainExplanation":"Sparse tails require rare-count corrections, covariance bounds must retain the largest rank-one contribution, and reused embeddings produce cross-time variance terms. The singular-value lemma needs its population qualifier. Four local matrix-calculus identities also need corrections; the accepted arguments distinguish these defects from an unproved claim that the main capacity exponents fail.","primaryUrl":"https://arxiv.org/html/2603.26554","noteUrl":"/review?paper=review-neurips-2026-149663","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149665","title":"StepCAD: Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search","authors":["Ghadi Nehme","Faez Ahmed"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"StepCAD: Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2610.03799v1","noteUrl":"/review?paper=review-neurips-2026-149665","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149666","title":"Fixed Universal Transformers","authors":["Jingwen Liu","Alexandr Andoni","Daniel Hsu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Fixed Universal Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31423","noteUrl":"/review?paper=review-neurips-2026-149666","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149669","title":"Vision to Geometry: 3D Spatial Memory for Sequential Embodied MLLM Reasoning and Exploration","authors":["Zhongyi Cai","Yi Du","Chen Wang","Yu Kong"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Vision to Geometry: 3D Spatial Memory for Sequential Embodied MLLM Reasoning and Exploration","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.02458","noteUrl":"/review?paper=review-neurips-2026-149669","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149670","title":"PACE: Two-Timescale Self-Evolution for Small Language Model Agents","authors":["Chen Ling","Pei Chen","Xiangchen Guan","Jiaming Qu","Shayan Ali Akbar","Madhu Gopinathan","Erwin Cornejo"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"PACE: Two-Timescale Self-Evolution for Small Language Model Agents","plainExplanation":"A truly worse candidate can pass a fresh finite validation comparison with positive probability. The gate therefore guarantees improvement only in its measured score; population guarantees need a false-accept bound or a stronger decision rule. Fresh sampling alone does not give exact non-regression.","primaryUrl":"https://arxiv.org/html/2605.23019","noteUrl":"/review?paper=review-neurips-2026-149670","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149672","title":"Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails","authors":["Gert Lek","Abele Mălan","Chaoyi Zhu","Pin-Yu Chen","Robert Birke","Lydia Chen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33634","noteUrl":"/review?paper=review-neurips-2026-149672","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149673","title":"Specificity-Aware Diffusion Steering via Variance-Reduced Sequential Monte Carlo","authors":["Luran Wang","Linrui Ma","Hannes Stark","Regina Barzilay"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Specificity-Aware Diffusion Steering via Variance-Reduced Sequential Monte Carlo","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.00395","noteUrl":"/review?paper=review-neurips-2026-149673","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149674","title":"Drag as Evidence: Motion-Grounded Latent Recomposition for Drag-Based Editing","authors":["Xinyu Pu","Hongsong Wang","Jie Gui","Pan Zhou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Drag as Evidence: Motion-Grounded Latent Recomposition for Drag-Based Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36755","noteUrl":"/review?paper=review-neurips-2026-149674","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149675","title":"LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling","authors":["Tong Zheng","Haolin Liu","Chengsong Huang","Huiwen Bao","Sheng Zhang","Rui Liu","Runpeng(Leo) Dai","Ruibo Chen","chenxi liu","Tianyi Xiong","Xidong Wu","Hongming Zhang","Heng Huang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The trajectory initialization supplies too few state components","plainExplanation":"The trajectory’s initial tuple has five components, but the defined state has six. One empty collection is missing from the later initialization. This is a tuple-arity typo. The earlier definition resolves the intended initial state; it does not demonstrate a controller failure.","primaryUrl":"https://arxiv.org/html/2605.08083","noteUrl":"/review?paper=review-neurips-2026-149675","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149676","title":"LAFP: Preserving Latent Action Structure in Latent Policy Learning via Flow Matching","authors":["Jiexi Lyu","Xi zhou Bu","Qingqiu Huang","Chufeng Tang","Xiaoshuai Hao","Hongbo Wang","Wei Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"LAFP: Preserving Latent Action Structure in Latent Policy Learning via Flow Matching","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.10517","noteUrl":"/review?paper=review-neurips-2026-149676","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149678","title":"Looped Transformers with Layer Normalization Provably Learn the Power Method","authors":["Lyumin Wu","Chenyang Zhang","Yuan Cao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"An arbitrary learning rate can make the claimed invariant negative","plainExplanation":"An arbitrary positive learning rate can make the claimed nonnegative invariant negative after the first update.","primaryUrl":"https://arxiv.org/html/2606.00605","noteUrl":"/review?paper=review-neurips-2026-149678","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149680","title":"Steering Optimisation Trajectories in Diffusion Representation Learning","authors":["Rajat Rasal","Tian Xia","Avinash Kori","Ben Glocker"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Steering Optimisation Trajectories in Diffusion Representation Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.05319","noteUrl":"/review?paper=review-neurips-2026-149680","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149681","title":"Minimax Optimal Early-Stopped Gradient Descent for Gaussian Mixture Classification","authors":["Alex Buna","Xiaoqi Liu","Patrick Rebeschini"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The induced inner product should not have a square root","plainExplanation":"The displayed square root is not an induced inner product: it is not bilinear and is not even real for opposite nonzero vectors. This is a local notation error. It does not supply a counterexample to the classification theorem.","primaryUrl":"https://arxiv.org/html/2608.06250","noteUrl":"/review?paper=review-neurips-2026-149681","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149683","title":"LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations","authors":["Mohan Li","Chengyu Yu","Francesco Sovrano","Marc Langheinrich","Martin Gjoreski"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.32717","noteUrl":"/review?paper=review-neurips-2026-149683","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149684","title":"Asking the Right Questions: Improving Reasoning with Generated Stepping Stones","authors":["Hengyuan Hu","Tingchen Fu","Minqi Jiang","Alexander Miller","Yoram Bachrach","Jakob Foerster"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Asking the Right Questions: Improving Reasoning with Generated Stepping Stones","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.19069","noteUrl":"/review?paper=review-neurips-2026-149684","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149685","title":"KroQuant: Kronecker-Structured Block Transforms for Efficient Post-Training Quantization of Diffusion Transformers","authors":["Yann Bouquet","Alireza Khodamoradi","Kristof Denolf","Mathieu Salzmann"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The invariance identity needs an inverse transpose","plainExplanation":"The stated transform-invariance identity uses an inverse where the weight convention requires an inverse transpose. The printed cancellation fails for admissible nonsymmetric transforms. This is a local transpose typo. Symmetric special cases do not establish the general identity, but the correct formulas elsewhere identify the intended transform; no code fault is inferred.","primaryUrl":"https://arxiv.org/html/2607.21446","noteUrl":"/review?paper=review-neurips-2026-149685","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149687","title":"Learning to Persuade a Biased Receiver","authors":["Yuqi Pan","Sadie Zhao","Milind Tambe","Yiling Chen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning to Persuade a Biased Receiver","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15331","noteUrl":"/review?paper=review-neurips-2026-149687","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149690","title":"OTROPE: Optimal Transport-based Robust Off-policy Evaluation for Large Language Models","authors":["Liner Xiang","Wenbo Zhang","Hengrui Cai"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Simplex transport weights must not be divided by sample count again","plainExplanation":"The transport weights already sum to one. Dividing their weighted score by sample count again misnormalizes the OT-only estimator.","primaryUrl":"https://arxiv.org/html/2609.36264","noteUrl":"/review?paper=review-neurips-2026-149690","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149694","title":"Scenes as Objects, Not Primitives : Instance-Structured 3D Tokenization from Unposed Views","authors":["Mijin Yoo","In Cho","Subin Jeon","Jiwoo Lee","Eunbyung Park","Seon Joo Kim"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Scenes as Objects, Not Primitives : Instance-Structured 3D Tokenization from Unposed Views","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.29513","noteUrl":"/review?paper=review-neurips-2026-149694","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149695","title":"The Score Kalman Filter","authors":["Kaito Iwasaki","Anthony Bloch","Taeyoung Lee","Maani Ghaffari"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Score Kalman Filter","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16644","noteUrl":"/review?paper=review-neurips-2026-149695","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149696","title":"Laplacian Heads Improve Transformers by Smoothing Token Representations","authors":["Yuchong Zhang","Vardan Papyan"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The positive Laplacian residual can amplify rather than diffuse","plainExplanation":"The stated Euler heat step is \\(X-(I-P)X\\). Set value and output weights to one and query/key weights to zero, giving uniform \\(P\\) on scalar tokens \\((1,-1)\\). Then \\((I-P)X=X\\), so the printed positive residual gives \\((2,-2)\\), raising variance from \\(1\\) to \\(4\\); the heat step gives zero. A learned negative projection can implement diffusion, but that is an additional weight property. The unconditional diffusion interpretation fails. Learned smoothing and empirical benefits remain possible.","primaryUrl":"https://arxiv.org/html/2602.09297","noteUrl":"/review?paper=review-neurips-2026-149696","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149698","title":"DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images","authors":["Pengfei Wang","Shihao Wang","liyi chen","Zhiyuan Ma","Guowen Zhang","Lei Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The displayed mask loss is squared error rather than binary cross-entropy","plainExplanation":"The displayed mask objective is squared error, despite being called binary cross-entropy. The two are different functions of the predicted probability and have different gradients. Both losses are valid objectives. This is a mathematical name/formula mismatch and does not show which objective was executed.","primaryUrl":"https://arxiv.org/html/2606.12368","noteUrl":"/review?paper=review-neurips-2026-149698","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149701","title":"Visual-Advantage On-Policy Distillation for Vision-Language Models","authors":["Ruiqi Liu","Xiaolei Lv","Gengsheng Li","Ximo Zhu","Zhiheng Wang","Zhengbo Zhang","Junkai Chen","Zhiheng Li","Bo Li","Jun Gao","Shu Wu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Teacher-scored token changes do not identify student visual sensitivity","plainExplanation":"A student can improve teacher-scored visual advantage through its output distribution while remaining completely insensitive to the image.","primaryUrl":"https://arxiv.org/html/2605.21924v1","noteUrl":"/review?paper=review-neurips-2026-149701","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149703","title":"PhotoFlow: Agentic 3D Virtual Photography Missions","authors":["Jiarui Guo","Haojia Wei","Yiming Zhang","Yifei Liu","Yuning Gong","Hongjie Zhang","Xue Yang","Zhihang Zhong"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"PhotoFlow: Agentic 3D Virtual Photography Missions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23771","noteUrl":"/review?paper=review-neurips-2026-149703","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149704","title":"Model Spec Midtraining: Improving How Alignment Training Generalizes","authors":["Chloe Li","Sara Price","Samuel Marks","Jonathan Kutasov"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Model Spec Midtraining: Improving How Alignment Training Generalizes","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.02087","noteUrl":"/review?paper=review-neurips-2026-149704","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149705","title":"BROS: Bias-Corrected Randomized Subspaces for Memory-Efficient Single-Loop Bilevel Optimization","authors":["Hengrui Zhang","Boao Kong","Engao Zhang","Kun Yuan"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"BROS: Bias-Corrected Randomized Subspaces for Memory-Efficient Single-Loop Bilevel Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10288","noteUrl":"/review?paper=review-neurips-2026-149705","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149706","title":"Learning to Bid in Repeated Second-Price Auctions with  Dynamic Values and  Aggregated Feedback","authors":["Benjamin Heymann","Otmane Sakhi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The ridge confidence radius omits deterministic bias","plainExplanation":"The ridge concentration radius omits deterministic regularization bias. Separate proof steps also discard Gram cross terms and select a quadratic branch without justification.","primaryUrl":"https://arxiv.org/html/2605.28133","noteUrl":"/review?paper=review-neurips-2026-149706","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149709","title":"Bug or Feature$^2$: Weight Drift, Activation Sparsity, and Spikes","authors":["Egor Shvetsov","Aleksandr Serkov","Shokorov Viacheslav","Redko Dmitry","Vladislav Goloshchapov","Evgeny Burnaev"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Bug or Feature$^2$: Weight Drift, Activation Sparsity, and Spikes","plainExplanation":"The allowed initialization class contains a counterexample to the column-correlation step, although that example does not refute every final gradient-sign conclusion. A direct Gaussian calculation gives a narrower repair. Separately, the B-SiLU derivative changes sign and has a zero, so its claimed everywhere-positive backward signal is false.","primaryUrl":"https://arxiv.org/html/2605.17659","noteUrl":"/review?paper=review-neurips-2026-149709","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149710","title":"Random Neural Network Expressivity for Non-Linear Partial Differential Equations","authors":["Muhammed Ali Mehmood","Lukas Gonon"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Random Neural Network Expressivity for Non-Linear Partial Differential Equations","plainExplanation":"At derivative order zero, the printed coefficient would require exact representation of a smooth bump by a finite analytic feature span. The proof instead supports a positive derivative-sum constant. Restoring that constant retains the Monte Carlo approximation rate under the stated moment conditions.","primaryUrl":"https://arxiv.org/html/2605.25057","noteUrl":"/review?paper=review-neurips-2026-149710","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149711","title":"Trust Region Q Adjoint Matching","authors":["Yonghoon Dong","Kyungmin Lee","Changyeon Kim","Jaehyuk Kim","Jinwoo Shin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The zero-start Euler grid evaluates singular coefficients","plainExplanation":"At \\(t=0\\), neither coefficient is defined for the specified Gaussian initial state. The diffusion energy from \\(\\delta\\) to \\(1\\) is \\(-2-2\\log\\delta+2\\delta\\), which diverges as \\(\\delta\\to0\\). Repeating the same zero-start kernels in an appendix is not an entrance-step definition. The printed Euler endpoint is undefined. The experiments may use a guard that is not stated; their behavior is not inferred.","primaryUrl":"https://arxiv.org/html/2605.27079","noteUrl":"/review?paper=review-neurips-2026-149711","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149712","title":"MIRAGE: Adaptive Multimodal Gating for Whole-Brain fMRI Encoding","authors":["Abdulkadir Gokce","Badr AlKhamissi","Martin Schrimpf"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MIRAGE: Adaptive Multimodal Gating for Whole-Brain fMRI Encoding","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.29850","noteUrl":"/review?paper=review-neurips-2026-149712","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149714","title":"The State-Prediction Separation Hypothesis","authors":["Giovanni Monea","Nathan Godey","Kianté Brantley","Yoav Artzi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The State-Prediction Separation Hypothesis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.01218","noteUrl":"/review?paper=review-neurips-2026-149714","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149716","title":"InstructVVT: Instruction-Driven Video Virtual Try-On without Auxiliary Spatial Priors","authors":["Dingbao Shao","Song Wu","Xinyu Chen","Qian Wang","Jiahang Li","Kuai Jiang","Jiang Lin","Yuhang Liu","Chen Ziyu","Duo Li","Hu Jiaxin","Shengrong Gu","Ziheng Tang","Liu Rongrong","Yanlun Peng","Liang Li","Junlan Feng","Lujia Jin","Ting Zhang","Jian Yang","Zili Yi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"InstructVVT: Instruction-Driven Video Virtual Try-On without Auxiliary Spatial Priors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.14070","noteUrl":"/review?paper=review-neurips-2026-149716","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149717","title":"Discovering Mechanistic Models of Neural Activity: System Identification in an in Silico Zebrafish","authors":["Jan-Matthis Lueckmann","Viren Jain","Michal Januszewski"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Discovering Mechanistic Models of Neural Activity: System Identification in an in Silico Zebrafish","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04492","noteUrl":"/review?paper=review-neurips-2026-149717","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149719","title":"Beyond LLM-Based Reasoning: Lightweight GNNs for Agent Failure Attribution","authors":["Ting-Wei Li","Yuanchen Bei","Xiao Lin","Hanghang Tong"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Beyond LLM-Based Reasoning: Lightweight GNNs for Agent Failure Attribution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.18575","noteUrl":"/review?paper=review-neurips-2026-149719","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149720","title":"PDE-SSM: A Spectral State Space Approach to Spatial Mixing in Diffusion Transformers","authors":["Eshed Gal","Moshe Eliasof","Eldad Haber"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"PDE-SSM: A Spectral State Space Approach to Spatial Mixing in Diffusion Transformers","plainExplanation":"A positive reaction amplifies the constant mode even with strictly positive diffusion. The scalar contraction claim needs a nonpositive reaction term; without it, the valid bound permits exponential growth at the reaction rate.","primaryUrl":"https://arxiv.org/html/2603.13663","noteUrl":"/review?paper=review-neurips-2026-149720","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149722","title":"Preference Conditioned Multi-Objective Reinforcement Learning: Decomposed, Diversity-Driven Policy Optimization","authors":["Tanmay Ambadkar","Sourav Panda","Shreyash Kale","Jonathan Dodge","Abhinav Verma"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The hierarchy analyzes a different order of weighting and normalization","plainExplanation":"For unit-variance Gaussian policies with means \\(\\theta\\) and zero, KL is \\(\\theta^2/2\\). Preference vertices have distance \\(2\\), so at the stated scaling the penalty is \\((\\theta^2/2-2)^2\\), with derivative \\(2(\\theta^2/2-2)\\theta\\). It is positive but has derivative zero at collapse \\(\\theta=0\\). Joint PPO or perturbations may still move the policies. The claimed diversity-gradient implication alone is false.","primaryUrl":"https://arxiv.org/html/2602.07764","noteUrl":"/review?paper=review-neurips-2026-149722","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149723","title":"Optimizing Social Utility in Sequential Experiments","authors":["Ander Artola Velasco","Stratis Tsirtsis","Manuel Rodriguez"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Optimizing Social Utility in Sequential Experiments","plainExplanation":"At a participation tie, the supplied solver can choose a different policy from the one used to evaluate the returned subsidy, and an optimum may not be attained. Averaging over a continuum of prior beliefs can also destroy finite piecewise linearity. Convexity belongs to the follower’s value, not automatically to the principal’s utility; one final monotonicity inequality additionally has a local sign error.","primaryUrl":"https://arxiv.org/html/2605.06520","noteUrl":"/review?paper=review-neurips-2026-149723","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement","formula_error","interpretation_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149724","title":"Efficient Reasoning via Constrained Optimization in Latent Space","authors":["Zhinan Hou","Xingchen Li","Keyou You"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The intercept construction proves feasibility rather than strict feasibility","plainExplanation":"Equation (4) places the Direct mean on or below every hyperplane, allowing equality. Take Direct states all zero and Vanilla clusters at opposing coordinate directions. The adjusted intercepts can all be zero, imposing both \\(u_j\\leq0\\) and \\(-u_j\\leq0\\). The feasible set is nonempty but no point satisfies both strictly. Nevertheless a squared-distance objective on a nonempty closed convex set is coercive and strictly convex, so its projection exists and is unique. Only strict feasibility is false; projection existence, uniqueness and the method remain valid.","primaryUrl":"https://arxiv.org/html/2609.34181","noteUrl":"/review?paper=review-neurips-2026-149724","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149725","title":"It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs","authors":["Sangwoo Park","Woongyeong Yeo","Yumin Choi","Hyomin Lee","Kangsan Kim","Seanie Lee","Jinheon Baek","Seong Joon Oh","Sung Ju Hwang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20258","noteUrl":"/review?paper=review-neurips-2026-149725","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149726","title":"SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation","authors":["Tianjiao (Joey) Yu","Xinzhuo Li","Yifan Shen","Ying Shen","Kiet Nguyen","Adheesh S Juvekar","Ismini Lourentzou"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02201","noteUrl":"/review?paper=review-neurips-2026-149726","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149727","title":"A Theoretical Analysis of Why Masked Diffusion Models Mitigate the Reversal Curse","authors":["Moongyu Jeon","Sangwoo Shin","Bumjun Kim","Kyelim Lee","Albert No"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"A Theoretical Analysis of Why Masked Diffusion Models Mitigate the Reversal Curse","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.02133","noteUrl":"/review?paper=review-neurips-2026-149727","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149728","title":"Hyper Hawkes Processes: Interpretable Models of Marked Temporal Point Processes","authors":["Alex Boyd","andrew warrington","Taha Kass-Hout","Parminder Bhatia","Cao (Danica) Xiao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Hyper Hawkes Processes: Interpretable Models of Marked Temporal Point Processes","plainExplanation":"The compensator depends on the realized event history and can take different values on positive-probability paths. Its expectation equals the expected count under the stated integrability conditions. Future conditional counts likewise require an expectation over future histories.","primaryUrl":"https://arxiv.org/html/2511.01096","noteUrl":"/review?paper=review-neurips-2026-149728","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149730","title":"Resilient Byzantine Agreement with Predictions","authors":["Julien Dallot","Darya Melnyk","Tijana Milentijević","Stefan Schmid","Patrik Welters"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Resilient Byzantine Agreement with Predictions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19452","noteUrl":"/review?paper=review-neurips-2026-149730","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149731","title":"PACZero: PAC-Private Fine-Tuning of Language Models via Sign Quantization","authors":["Murat Bilgehan Ertan","Xiaochen Zhu","Ha Nguyen","Marten van Dijk","Srinivas Devadas"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"PACZero: PAC-Private Fine-Tuning of Language Models via Sign Quantization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06505","noteUrl":"/review?paper=review-neurips-2026-149731","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149736","title":"Beyond Kemeny Medians: Consensus Ranking Distributions. Definition, Properties and Statistical Learning","authors":["Stephan Clémençon","Ekhine Irurozki"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Beyond Kemeny Medians: Consensus Ranking Distributions. Definition, Properties and Statistical Learning","plainExplanation":"The local estimator targets twice the stated variability, the proposed high-probability cell count exceeds the entire sample, and the proposed ranking masses need not sum to one. The report gives valid local replacements while keeping later learning-rate and computation claims separate.","primaryUrl":"https://arxiv.org/html/2602.10640","noteUrl":"/review?paper=review-neurips-2026-149736","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","formula_error","incorrect_bound","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149740","title":"Improved Baselines with Representation Autoencoders","authors":["Jaskirat Singh","Boyang Zheng","Zongze Wu","Richard Zhang","Eli Shechtman","Saining Xie"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Improved Baselines with Representation Autoencoders","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18324","noteUrl":"/review?paper=review-neurips-2026-149740","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149741","title":"Does This Gradient Spark Joy?","authors":["Ian Osband"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Does This Gradient Spark Joy?","plainExplanation":"A quantile price does not calibrate finite-temperature Bernoulli retention. The allowed expected-reward baseline also defeats the universal quadratic backward-cost lower bound, while the additive-score threshold confuses ranking with sign separation. A separate intermediate inner-product typo does not affect its already correct final expression.","primaryUrl":"https://arxiv.org/html/2603.20526","noteUrl":"/review?paper=review-neurips-2026-149741","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","calibration_error","complexity_error","false_statement","formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149743","title":"StreamPhy: Streaming Inference of High-Dimensional Physical Dynamics via  State Space Models","authors":["Panqi Chen","Yifan Sun","Shikai Fang","Xiao Fu","Lei Cheng"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"StreamPhy: Streaming Inference of High-Dimensional Physical Dynamics via  State Space Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07384","noteUrl":"/review?paper=review-neurips-2026-149743","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149744","title":"Beyond Thinking: Imagining in 360$^\\circ$ for Humanoid Visual Search","authors":["Jingdong Zhang","Yizhou Wang","Zhengzhong Tu","Xin Li","Wenping Wang","Xiaohang Zhan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Unsigned wrapping breaks the symmetric field-of-view test","plainExplanation":"Unsigned azimuth wrapping maps a small leftward displacement near a full turn, breaking the symmetric field-of-view test.","primaryUrl":"https://arxiv.org/html/2605.09146","noteUrl":"/review?paper=review-neurips-2026-149744","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149745","title":"Express Language Modeling","authors":["Albert Gong","Annabelle M Carrell","Raaz Dwivedi","Lester Mackey"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Express Language Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.10944","noteUrl":"/review?paper=review-neurips-2026-149745","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149746","title":"C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models","authors":["Xin Zhang","Liang Bai","Xian Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2502.17920","noteUrl":"/review?paper=review-neurips-2026-149746","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149747","title":"COLLAR: Cascaded Object-Level Latent Refinement for High-Fidelity Conditional Generation","authors":["Xinlong Zhang","Jia Wei","Xiaoyu Zhang","Teng Zhou","Chengyu Lin","Yongchuan Tang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Masked feature averaging needs a background branch","plainExplanation":"The object-feature quotient is undefined outside every object mask and needs an explicit background branch.","primaryUrl":"https://arxiv.org/html/2606.00954","noteUrl":"/review?paper=review-neurips-2026-149747","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149748","title":"Improving Flexible Image Tokenizers for Autoregressive Image Generation","authors":["Zixuan Fu","Lanqing Guo","Chong Wang","Binbin SONG","Ding Liu","Bihan Wen"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Improving Flexible Image Tokenizers for Autoregressive Image Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.01535","noteUrl":"/review?paper=review-neurips-2026-149748","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149751","title":"Beyond Sparse Captions: Aligning Slide-Level Text and Patch-Level Vision in Pathology","authors":["Daniel Shao","Dongmin Bang","Luca Weishaupt","Nic G Reitsam","Sophia J. Wagner","Ming Yang Lu","Long P Le","Richard Chen","Faisal Mahmood"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond Sparse Captions: Aligning Slide-Level Text and Patch-Level Vision in Pathology","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.17233","noteUrl":"/review?paper=review-neurips-2026-149751","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149753","title":"HyCO: A Hybrid Neural Solver for Combinatorial Optimization","authors":["Yuheng Li","Di Yang","Haipeng Chen","Yanhai Xiong"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"HyCO: A Hybrid Neural Solver for Combinatorial Optimization","plainExplanation":"An admissible scalar regret example has three optimal handovers despite the required endpoint inequalities and growth behavior. Strict increase of the marginal-regret difference gives a sufficient uniqueness condition. Without it, the original continuity and endpoint signs still guarantee an interior minimizer better than either endpoint.","primaryUrl":"https://arxiv.org/html/2609.07990","noteUrl":"/review?paper=review-neurips-2026-149753","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149754","title":"Training ML Models with Predictable Failures","authors":["Will Schwarzer","Scott Niekum"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Training ML Models with Predictable Failures","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15134","noteUrl":"/review?paper=review-neurips-2026-149754","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149757","title":"Causal Inference for Sequential Settings under Interference and Latent Confounding","authors":["Phevos Paschalidis","Constantinos Daskalakis","Devavrat Shah"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Causal Inference for Sequential Settings under Interference and Latent Confounding","plainExplanation":"The ordered-pair energy disagrees with the printed local field, and marginal treatment overlap does not ensure separation from low-rank confounding. An auxiliary self-normalized tail lemma also lacks scale control. The Brier formula uses the wrong outcome encoding, while transition normalizers create dependencies omitted by the proposed validation separator. The report supplies scoped repairs rather than attributing every defect to the intended half-energy model.","primaryUrl":"https://arxiv.org/html/2607.14940","noteUrl":"/review?paper=review-neurips-2026-149757","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","false_statement","formula_error","lemma_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149758","title":"Quantile Geometry Regularization for Distributional Reinforcement Learning","authors":["Zhaofan ZHANG","Minghao Yang","Rufeng Chen","Sihong Xie","Hui Xiong"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Quantile Geometry Regularization for Distributional Reinforcement Learning","plainExplanation":"No issue was found in the checked claims. The residual uses the sign required for expansion of the effective prediction. The stated effective-prediction identity is valid.","primaryUrl":"https://arxiv.org/html/2605.08182v1","noteUrl":"/review?paper=review-neurips-2026-149758","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149759","title":"Boosting Inference with Guided Reasoning: Stochastic Exploration for Recursive Models","authors":["Andrew Corbett","Archit Sood","Anna Tzatzopoulou","Sai-Aakash Ramesh","Tim Dodwell"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The tilt inequality mixes population mass with empirical spread","plainExplanation":"Integrating the tilt derivative and using the indicator standard deviation bound gives \\(|p_n(\\beta)-p_n(0)|\\leq(\\beta/2)\\sup_t\\operatorname{Std}_{\\pi_{n+1,t}}(q)\\). This uses population spread. A finite cloud can underestimate it with positive probability, so substitution does not give a deterministic population bound. If both mass and standard deviation instead refer to the same weighted empirical cloud, the analogous identity is valid. The mixed population/empirical bound is unsupported. The coherent empirical version and the other tilt propositions survive; failure probabilities depend on the sampling interpretation.","primaryUrl":"https://arxiv.org/html/2605.25230","noteUrl":"/review?paper=review-neurips-2026-149759","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149762","title":"BEAST3D: animal behavioral analysis and neural encoding from multi-view video via Gaussian splatting","authors":["Yanchen Wang","Lenny Aharon","Wangshu Zhu","Kyle Daruwalla","Linghua Zhang","Jiaru Zou","Selmaan N Chettih","X Helen Hou","Liam Paninski","Matthew Whiteway"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BEAST3D: animal behavioral analysis and neural encoding from multi-view video via Gaussian splatting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02937","noteUrl":"/review?paper=review-neurips-2026-149762","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149764","title":"Modeling Whole-Slide Images as Dynamic Tumor Microenvironment Fields","authors":["Lei Wu","Jiashuai Liu","Di Zhang","Zhangpeng Gong","Yingkang Zhan","Yi Niu","Jiusong Ge","Chunze Yang","Kai Yi","Mireia Crispin-Ortuzar","Chen Li","Zeyu Gao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Maximum-response nodes have no positive-difference flux support","plainExplanation":"Maximum-response nodes have zero positive-difference support, making the normalized flux undefined. The claimed directed energy derivative also omits incoming-edge terms.","primaryUrl":"https://arxiv.org/html/2609.34451","noteUrl":"/review?paper=review-neurips-2026-149764","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149766","title":"AffectGPT-RL: Revealing Roles of Reinforcement Learning in Open-Vocabulary Emotion Recognition","authors":["Zheng Lian","Fan Zhang","Lan Chen","Yazhou Zhang","Rui Liu","Jinyang Wu","Haoyu Chen","Xiaobai Li","Xiaojiang Peng","Bin He","Jianhua Tao"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"AffectGPT-RL: Revealing Roles of Reinforcement Learning in Open-Vocabulary Emotion Recognition","plainExplanation":"For one or two annotations, the logarithmic multiplier can impose a stronger penalty than the ratio multiplier. The stated ordering holds for integer counts of at least three; the formulas remain usable outside that range without the universal ordering claim.","primaryUrl":"https://arxiv.org/html/2605.06126","noteUrl":"/review?paper=review-neurips-2026-149766","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["false_statement"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149768","title":"World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories","authors":["Jiahui Lei","Qianqian Wang","Trevor Darrell","Angjoo Kanazawa"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01742","noteUrl":"/review?paper=review-neurips-2026-149768","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149769","title":"CAPO: A Primal-Dual Framework for Constraint-Aware Prompt Optimization","authors":["Victor Ye Dong","Reid Pryzant","Yi Liu","Jian Jiao"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"CAPO: A Primal-Dual Framework for Constraint-Aware Prompt Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.16068v1","noteUrl":"/review?paper=review-neurips-2026-149769","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149774","title":"Asymmetric Scaling Laws from Sparse Features","authors":["John Sous","Michael Winer"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Asymmetric Scaling Laws from Sparse Features","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23591","noteUrl":"/review?paper=review-neurips-2026-149774","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149775","title":"Learning Human-Intention Priors from Large-Scale Human Demonstrations for Robotic Manipulation","authors":["Yifan Xie","YuAn Wang","Guangyu Chen","Jinkun Liu","Yu Sun","Wenbo Ding"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Learning Human-Intention Priors from Large-Scale Human Demonstrations for Robotic Manipulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.24681","noteUrl":"/review?paper=review-neurips-2026-149775","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149778","title":"Active Learning for Conditional Generative Compressed Sensing","authors":["Alexander DeLise","Nick Dexter"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Active Learning for Conditional Generative Compressed Sensing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05435","noteUrl":"/review?paper=review-neurips-2026-149778","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149779","title":"Closed Loop Dynamic Driving Data Mixture for Real-Synthetic Co-Training","authors":["Hongzhi Ruan","Pei Liu","Weiliang Ma","Zhengning Li","Xueyang Zhang","Jun Ma","Dan Xu","Kun Zhan"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Zero predicted gains leave the mixture step undefined","plainExplanation":"Constant cluster scores yield zero predicted gains and an undefined gain-normalized mixture step.","primaryUrl":"https://arxiv.org/html/2605.21372","noteUrl":"/review?paper=review-neurips-2026-149779","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149780","title":"Hermes: A Multi-Scale Spatial-Temporal Hypergraph Network for Stock Time Series Forecasting","authors":["Xiangfei Qiu","Liu Yang","Xiangyu Xu","Hanyin Cheng","Xingjian Wu","Rongjia Wu","Zhang Zhigang","Tu ding","Chenjuan Guo","Bin Yang","Christian S. Jensen","Jilin Hu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"PSD distance can vanish between off-diagonal embeddings","plainExplanation":"Positive semidefiniteness permits \\(A=0\\), giving distance zero between distinct embeddings. The reciprocal then is undefined because only the diagonal case is guarded. Even positive definiteness would not rule out coincident embeddings, which also have zero distance. The real-valued adjacency definition lacks a boundary convention. Softmax on infinities does not itself supply it; no code failure is claimed.","primaryUrl":"https://arxiv.org/html/2509.23668","noteUrl":"/review?paper=review-neurips-2026-149780","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149784","title":"Searching Videos as Trees: Self-Correcting Agents for Grounded Long Video QA","authors":["Ce Zhang","Ziyang Wang","Yulu Pan","Oluwatumininu Oguntola","Pranav Wagh","Qiyu Wu","Hiromi Wakaki","Mohit Bansal","Gedas Bertasius"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Searching Videos as Trees: Self-Correcting Agents for Grounded Long Video QA","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.16189","noteUrl":"/review?paper=review-neurips-2026-149784","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149787","title":"FlowSteer: Prompt-Only Workflow Steering Exposes Planning-Time Vulnerabilities in Multi-Agent LLM Systems","authors":["Fanxiao Li","Jiaying Wu","Tingchao Fu","Natasha Jaques","Wei Zhou","Min-Yen Kan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FlowSteer: Prompt-Only Workflow Steering Exposes Planning-Time Vulnerabilities in Multi-Agent LLM Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11514","noteUrl":"/review?paper=review-neurips-2026-149787","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149789","title":"Embedded-Arena: Building Hardware-in-the-Loop Coding Agents to Run AI on Microcontrollers","authors":["Zhihan Zhang","Alexander Le Metzger","Jiuyang Lyu","Chun-Cheng Chang","Jiayi Shao","Yujia Liu","Emmanuel A Mensah","Edward Wang","Kurtis Heimerl","Gregory D Abowd","Natasha Jaques","Shwetak Patel","Vikram Iyer"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Embedded-Arena: Building Hardware-in-the-Loop Coding Agents to Run AI on Microcontrollers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.16190v1","noteUrl":"/review?paper=review-neurips-2026-149789","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149790","title":"AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent","authors":["Yinyi Luo","Yiqiao Jin","Weichen Yu","Mengqi Zhang","Srijan Kumar","Xiaoxiao Li","Weijie Xu","Xin chen","Jindong Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The supervised answer term has the wrong minimization sign","plainExplanation":"The minimized supervised loss gives the correct-answer log-probability a positive sign. Taken literally, that term pushes correct-answer probability downward rather than maximizing likelihood. The maximum-likelihood prose supplies the intended correction. This local sign/grouping typo does not establish a training-code or experimental failure.","primaryUrl":"https://arxiv.org/html/2602.03955","noteUrl":"/review?paper=review-neurips-2026-149790","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149792","title":"Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory","authors":["Runxi Cheng","Yuchen Guan","Yongxian Wei","Qianpu Sun","Qixiu Li","Sinan Du","Feng Xiong","Chun Yuan","Yan Lu","Yeyun Gong"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20948","noteUrl":"/review?paper=review-neurips-2026-149792","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149795","title":"An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models","authors":["Mingzhong Sun","Teresa Yeo","Armando Solar-Lezama","Tan Zhi-Xuan"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.01462","noteUrl":"/review?paper=review-neurips-2026-149795","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149798","title":"Reward Inflation: A Healthy Stimulus for Reinforcement Learning","authors":["Ganghun Lee","Minji Kim","Minsu Lee","Byoung-Tak Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Reward scaling does not imply Jacobian scaling","plainExplanation":"Scaling fitted values with rewards does not force their parameter Jacobians to scale, so it does not establish quadratic loss-gradient growth.","primaryUrl":"https://arxiv.org/html/2610.02545","noteUrl":"/review?paper=review-neurips-2026-149798","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149799","title":"Sampling-Based Safe Reinforcement Learning","authors":["Luca Vignola","Bruce D Lee","Manish Prajapat","Manuel Wendl","Melanie Zeilinger","Andreas Krause","Yarden As"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Sampling-Based Safe Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19469","noteUrl":"/review?paper=review-neurips-2026-149799","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149800","title":"Efficient Memory Crystallization for Graph Learning under Non-Stationary Distribution Shifts","authors":["Yue Hou","Ruomei Liu","Yingke Su","Wu Junran","Ke Xu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Efficient Memory Crystallization for Graph Learning under Non-Stationary Distribution Shifts","plainExplanation":"Marginal KL identities do not characterize independence for two arbitrary laws. Transferring source error to the target distribution also requires retaining the density ratio or an appropriate discrepancy bound.","primaryUrl":"https://arxiv.org/html/2610.02795","noteUrl":"/review?paper=review-neurips-2026-149800","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","invalid_inference"],"classifications":["major"]},{"slug":"review-neurips-2026-149803","title":"Stochastic Grouping Conformal Prediction for Effective Subgroup Reliability","authors":["Meihui Zhong","Wenxin Tai","Ting Zhong","Fan Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Stochastic Grouping Conformal Prediction for Effective Subgroup Reliability","plainExplanation":"The calibration cutoff can depend on the test point’s subgroup even when subgroup and marginal score distributions agree. Independent calibration and test observations, conditional on the fitted transform, give a sufficient repair. The counterexample preserves marginal conformal validity.","primaryUrl":"https://arxiv.org/html/2610.11957","noteUrl":"/review?paper=review-neurips-2026-149803","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149805","title":"Gaussian Mixture Models in Hilbert Spaces via Kernel Methods","authors":["Daniel López Montero","Antonio Álvarez-López","Marcos Matabuena"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Gaussian Mixture Models in Hilbert Spaces via Kernel Methods","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05996","noteUrl":"/review?paper=review-neurips-2026-149805","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149807","title":"Intrinsic Selection and Particle Resampling for Inference-Time Scaling Beyond Domain Verifiability","authors":["Giorgio Giannone","Mustafa Eyceoz","Shabana Baig","Shivchander Sudalairaj","Anna C Doris","Faez Ahmed","Akash Srivastava","Kai Xu"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The tail formula treats a length as a start index","plainExplanation":"The tail formula treats its parameter as a start index, while the embedded code treats it as a suffix length. The entropy-weighting display also has the opposite preference sign.","primaryUrl":"https://arxiv.org/html/2606.08850","noteUrl":"/review?paper=review-neurips-2026-149807","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149808","title":"Persona Generators: Generating Diverse Synthetic Personas for Arbitrary Contexts","authors":["Davide Paglieri","Logan Cross","William Cunningham","Joel Leibo","Sasha Vezhnevets"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Persona Generators: Generating Diverse Synthetic Personas for Arbitrary Contexts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.03545","noteUrl":"/review?paper=review-neurips-2026-149808","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149813","title":"Asymmetric Phase Coding Audio Watermarking","authors":["Guang Yang","Fengchen Liu","Amir Ghasemian","Zhong Wang","Ninareh Mehrabi","Homa Hosseinmardi"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Asymmetric Phase Coding Audio Watermarking","plainExplanation":"At an exact cell center, the sign-based parity correction can leave the wrong bit unchanged. Separately, opposite log-magnitude shifts preserve a product rather than the sum of squared amplitudes. The report gives a parity-preserving projection and the actual energy-change bound.","primaryUrl":"https://arxiv.org/html/2605.07241","noteUrl":"/review?paper=review-neurips-2026-149813","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149814","title":"Learning What to Predict: Downstream-Guided Task Design for Continued Pretraining","authors":["Shuqi Ke","Giulia Fanti"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Learning What to Predict: Downstream-Guided Task Design for Continued Pretraining","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.22108","noteUrl":"/review?paper=review-neurips-2026-149814","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149815","title":"Nash Social Welfare for Multi Armed Bandits: Trajectory-wise Expected and High Probability Regret","authors":["Avishek Ghosh"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Nash Social Welfare for Multi Armed Bandits: Trajectory-wise Expected and High Probability Regret","plainExplanation":"An algorithm that commits to one random arm has zero pseudo-regret with probability at least the inverse arm count, contradicting the proposed confidence quantifier. A direct argument yields a constant-probability lower bound of the intended order. The separate expected-regret comparison and upper bounds are unaffected by this witness.","primaryUrl":"https://arxiv.org/html/2610.07737","noteUrl":"/review?paper=review-neurips-2026-149815","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149817","title":"Support Before Frequency in Discrete Diffusion","authors":["Adrian Müller","Antoine Gonon","Zebang Shen","Ya-Ping Hsieh","Niao He"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Support Before Frequency in Discrete Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13999","noteUrl":"/review?paper=review-neurips-2026-149817","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149818","title":"GenRec: Knowing Where to Reconstruct and Where to Generate","authors":["Ata Çelen","Jaewoo Jung","Federico Tombari","Marc Pollefeys","Sunghwan Hong","Michael Niemeyer","Daniel Barath"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GenRec: Knowing Where to Reconstruct and Where to Generate","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.17832","noteUrl":"/review?paper=review-neurips-2026-149818","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149819","title":"Why Do Time Series Models Need Long Context Windows?","authors":["Luca Butera","Giovanni De Felice","Andrea Cini","Cesare Alippi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Why Do Time Series Models Need Long Context Windows?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.01999","noteUrl":"/review?paper=review-neurips-2026-149819","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149820","title":"RxGS: Receiver-Generalizable 3D Gaussian Splatting for Radio-Frequency Data Synthesis","authors":["Kang Yang","Mani Srivastava"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RxGS: Receiver-Generalizable 3D Gaussian Splatting for Radio-Frequency Data Synthesis","plainExplanation":"A line-of-sight path does not share the assigned scatterer’s final propagation segment. Separating it restores the stated factorization for groups with a common final scatterer; the conditional common-attenuation factorization is valid. Batching can share work and launch overhead, but it cannot remove the cost of writing a separate spectrum for every receiver.","primaryUrl":"https://arxiv.org/html/2605.24290","noteUrl":"/review?paper=review-neurips-2026-149820","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149821","title":"Learning Rate Transfer for Hybrid Transformer-SSM Architectures","authors":["Jimin Seo","Gyubok Lee","Yeonsik Jo","Kiwoong Yoo","Yeongoon Kim","Minhae Oh","Jin W Koo","Suhwan Kim","Nakyung Lee","Minsik Seol","Idris Nechnech","Jaehyeon Kim","LeeGiho","Jungwoo Lee"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Learning Rate Transfer for Hybrid Transformer-SSM Architectures","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01172","noteUrl":"/review?paper=review-neurips-2026-149821","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149822","title":"LASER: Latent Space Adjoint Matching for Support Constrained Entropy Regularized Offline RL","authors":["Songyuan Zhang","Oswin So","Eric Yu","Matthew Cleaveland","Peter Crowley-Dolen","Chuchu Fan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LASER: Latent Space Adjoint Matching for Support Constrained Entropy Regularized Offline RL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.08989","noteUrl":"/review?paper=review-neurips-2026-149822","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149823","title":"VETime: Vision Enhanced Zero-Shot Time Series Anomaly Detection","authors":["Yingyuan Yang","Tian Lan","Yifei Gao","Yimeng Lu","Xuming An","Wenjun He","Meng Wang","Chenghao Liu","Chen Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"VETime: Vision Enhanced Zero-Shot Time Series Anomaly Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.16681","noteUrl":"/review?paper=review-neurips-2026-149823","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149824","title":"The Override Gap: A Magnitude Account of Knowledge Conflict Failure in Hypernetwork-Based Instant LLM Adaptation","authors":["Shuaizhi Cheng","Xiang Shi","Zhiwei Zhang","Mingwei Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LoRA layer scaling does not imply exact output-margin scaling","plainExplanation":"The adapter contribution at one layer is linear in its scale for a fixed input. Nonlinear downstream layers, and scale-dependent inputs to later adapters, do not preserve that linearity at the output logits.","primaryUrl":"https://arxiv.org/html/2604.23750","noteUrl":"/review?paper=review-neurips-2026-149824","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149826","title":"An Õptimal Differentially Private PAC Learner for Concept Classes with VC Dimension 1","authors":["Chao Yan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The learner has local specification and proof gaps with concrete repairs","plainExplanation":"The final threshold does not follow from the event that was bounded. The exponential mechanism cannot guarantee a child below \\(\\alpha N\\) when every available child has larger weight.","primaryUrl":"https://arxiv.org/abs/2505.06581v2","noteUrl":"/review?paper=review-neurips-2026-149826","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149828","title":"Entropy Minimization without Model Collapse: Mitigating Prediction Bias in Medical Imaging","authors":["Tim Nielen","Sameer Ambekar","Daniel Lang","Julia Schnabel"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Entropy Minimization without Model Collapse: Mitigating Prediction Bias in Medical Imaging","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.02339","noteUrl":"/review?paper=review-neurips-2026-149828","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149829","title":"GRAPHLCP: Structure-Aware Localized Conformal Prediction on Graphs","authors":["Peyman Baghershahi","Fangxin Wang","Debmalya Mandal","Sourav Medya"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Calibration-fitted PCA is not calibration-role independent","plainExplanation":"Calibration-fitted PCA depends on calibration/test roles, leaving the claimed exchangeability justification incomplete. The walk recurrence also mixes row and column orientations.","primaryUrl":"https://arxiv.org/html/2605.08074","noteUrl":"/review?paper=review-neurips-2026-149829","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149834","title":"A Scalable Multi-Task Model for Virtual Sensors","authors":["Leon Götz","Lars Frederik Peiss","Erik Sauer","Andreas U Sass","Thorsten Bagdonat","Stephan Günnemann","Leo Schwinn"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Scalable Multi-Task Model for Virtual Sensors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.20634","noteUrl":"/review?paper=review-neurips-2026-149834","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149836","title":"FreeSpec: Training-Free Long Video Generation via Singular-Spectrum Reconstruction","authors":["Fangda Chen","Shanshan Zhao","Longrong Yang","Chuanfu Xu","Zhigang Luo","Long Lan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The local mask uses the native window length as a radius","plainExplanation":"The mask treats the stated total native length \\(W\\) as a radius, admitting \\(2W+1\\) interior tokens rather than \\(W\\). This is a width/radius inconsistency in the stated local operator. It does not identify the implementation’s actual mask.","primaryUrl":"https://arxiv.org/html/2605.06509","noteUrl":"/review?paper=review-neurips-2026-149836","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149837","title":"Near-Optimal Stochastic Linear Bandits with Delay","authors":["Ofir Schlisselberg","Mengxiao Zhang","Yishay Mansour"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Near-Optimal Stochastic Linear Bandits with Delay","plainExplanation":"The printed estimators mix upper and lower completions, global and armwise sums, or target and collected counts. Other corrections concern a main/formal elimination mismatch, a horizon restriction, a legal optimal-arm probability law and a nonnegative-variable concentration hypothesis. The report separates local constant and notation repairs from claims about the final regret rate.","primaryUrl":"https://arxiv.org/html/2606.16656","noteUrl":"/review?paper=review-neurips-2026-149837","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","incorrect_bound","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149838","title":"Rubato: Transcribing Piano Music with Timestamps","authors":["Nazif Can Tamer","Victoria Ebert","Guang Yang","Noah Smith"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rubato: Transcribing Piano Music with Timestamps","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24291","noteUrl":"/review?paper=review-neurips-2026-149838","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149839","title":"Neurosymbolic Learning for Inference-Time Argumentation","authors":["Gabriel Freedman","Adam Dejl","Adam Gould","Mansi -","Lihu Chen","Junqi Jiang","Francesca Toni"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Neurosymbolic Learning for Inference-Time Argumentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20098","noteUrl":"/review?paper=review-neurips-2026-149839","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149842","title":"Stability Regimes for Framing-Sensitive Fine-Tuning in Language Models","authors":["Victor De Lima","Jiqun Liu","Grace H Yang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Stability Regimes for Framing-Sensitive Fine-Tuning in Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.15273","noteUrl":"/review?paper=review-neurips-2026-149842","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149844","title":"CRePE: Curved Ray Expectation Positional Encoding for Unified-Camera-Controlled Video Generation","authors":["Seonghyun Jin","youngmin Kim","Sunwoo Park","Jong Chul Ye"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"CRePE: Curved Ray Expectation Positional Encoding for Unified-Camera-Controlled Video Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12938","noteUrl":"/review?paper=review-neurips-2026-149844","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149848","title":"Subcritical Signal Propagation at Initialization in Normalization-Free Transformers","authors":["Sergey Alekseev"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Subcritical Signal Propagation at Initialization in Normalization-Free Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.11890","noteUrl":"/review?paper=review-neurips-2026-149848","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149849","title":"Steering Frozen LLMs: Adaptive Social Alignment via Online Prompt Routing","authors":["Zeyu Zhang","Xiangxiang Dai","Ziyi Han","Xutong Liu","John C. S. Lui"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Steering Frozen LLMs: Adaptive Social Alignment via Online Prompt Routing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.15647","noteUrl":"/review?paper=review-neurips-2026-149849","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149852","title":"AudioGS: High-Fidelity Neural Audio Compression via Continuous Gaussian Splatting","authors":["Ron Aluf","Alon Canfi","Eliya Nachmani"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AudioGS: High-Fidelity Neural Audio Compression via Continuous Gaussian Splatting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04651","noteUrl":"/review?paper=review-neurips-2026-149852","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149853","title":"Navigating by Old Maps: The Pitfalls of Static Mechanistic Localization in LLM Post-Training","authors":["Hang Chen","Jiaying Zhu","Hongyang Chen","Hongxu Liu","Xinyu Yang","Wenya Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The denoising Taylor displacement has the opposite sign","plainExplanation":"The denoising Taylor expansion uses the opposite of the actual clean-minus-corrupted displacement.","primaryUrl":"https://arxiv.org/html/2605.06076","noteUrl":"/review?paper=review-neurips-2026-149853","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149858","title":"Beyond Isotropy in JEPAs: Hamiltonian Geometry and Symplectic Prediction","authors":["Robert Jenkinson Alvarez"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Beyond Isotropy in JEPAs: Hamiltonian Geometry and Symplectic Prediction","plainExplanation":"A smooth positive-definite Hamiltonian with position–momentum coupling gives a Jacobian determinant different from one. Restricting to separable Hamiltonians or using the stated implicit symplectic-Euler step repairs the claim. The specialized separable leapfrog predictor is a distinct, unaffected case.","primaryUrl":"https://arxiv.org/pdf/2605.20107","noteUrl":"/review?paper=review-neurips-2026-149858","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149859","title":"PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking","authors":["Junnan Nie","Jiayi Li","Jiachen Zhang","Junyi Lao","Chenghao Liu","Tianle Zhang","Songfang Huang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.00537","noteUrl":"/review?paper=review-neurips-2026-149859","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149862","title":"Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models","authors":["Ayan Sengupta","Vaibhav Seth","Tanmoy Chakraborty"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36734","noteUrl":"/review?paper=review-neurips-2026-149862","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149863","title":"DisRFM: Polar Riemannian Flow Matching for Structure-Preserving Graph Domain Adaptation","authors":["Yingxu Wang","Xinwang Liu","Siyang Gao","Mengzhu Wang","Nan Yin"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"DisRFM: Polar Riemannian Flow Matching for Structure-Preserving Graph Domain Adaptation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.00656","noteUrl":"/review?paper=review-neurips-2026-149863","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149865","title":"HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning","authors":["Shenzhi Wang","Shixuan Liu","Jing Zhou","Chang Gao","Xiong-Hui Chen","Binghai Wang","An Yang","Shiji Song","Bowen Yu","Gao Huang","Junyang Lin"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.17024","noteUrl":"/review?paper=review-neurips-2026-149865","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149868","title":"Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering","authors":["Jaewoo Jung","Hyeonseo Yu","Honggyu An","Jisang Han","Mungyeom Kim","Minkyeong Jeon","Heeseong Shin","WonJun Moon","Federico Tombari","Daniel Barath","Marc Pollefeys","Seungryong Kim","Sunghwan Hong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Gaussian head width cannot hold the stated SH attributes","plainExplanation":"The head has fewer channels per Gaussian than its stated spherical-harmonic color alone requires, before geometric attributes are included.","primaryUrl":"https://arxiv.org/html/2609.38177","noteUrl":"/review?paper=review-neurips-2026-149868","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149871","title":"Path-Guided Flow Matching for Dataset Distillation","authors":["xuhui li","Zhengquan luo","Zixu Wu","Xiwei Liu","Yongqiang Yu","Zixiang Hong","Zhiqiang Xu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Path-Guided Flow Matching for Dataset Distillation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.05616","noteUrl":"/review?paper=review-neurips-2026-149871","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149876","title":"PDF-HR: Pose Distance Fields for Humanoid Robots","authors":["Yi Gu","Yukang Gao","Yangchen Zhou","Xingyu Chen","Yixiao Feng","Mingle Zhao","Yunyang Mo","Zhaorui Wang","Lixin Xu","Renjing Xu"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"PDF-HR: Pose Distance Fields for Humanoid Robots","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04851","noteUrl":"/review?paper=review-neurips-2026-149876","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149877","title":"Model-Based Online Decision Making via Generative Trajectory Planning","authors":["Haldun Balim","Yilun Du","Na Li"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Model-Based Online Decision Making via Generative Trajectory Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.18677","noteUrl":"/review?paper=review-neurips-2026-149877","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149878","title":"Flow Matching from Viewpoint of Proximal Operators","authors":["Kenji Fukumizu","Wei Huang","Han Bao","Shuntuo Xu","Nisha Chandramoorthy"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Flow Matching from Viewpoint of Proximal Operators","plainExplanation":"A smooth function can make the regularized proximal objective unbounded below despite satisfying every stated local condition. Global convexity gives the existence and localization needed for the expansion. The objection is to the self-contained local formulation, not to the paper’s identities that explicitly assume convexity.","primaryUrl":"https://arxiv.org/html/2602.12683","noteUrl":"/review?paper=review-neurips-2026-149878","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149880","title":"MoRe: Modular Representations for Principled  Continual Representation Learning on Sequential Data","authors":["Jiaqi Sun","Boyang Sun","Rasmy M. H.","Xiangchen Song","Kun Zhang"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"MoRe: Modular Representations for Principled  Continual Representation Learning on Sequential Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14364","noteUrl":"/review?paper=review-neurips-2026-149880","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149882","title":"Bayesian Decision Making around Experts","authors":["Daniel Jarne Ornia","Joel Dyer","Nicholas Bishop","Anisoara Calinescu","Michael Wooldridge"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The countable-posterior proof needs a tail argument; several formulas need local corrections","plainExplanation":"The printed mixture is independent of \\(\\epsilon\\), so it cannot define a contamination threshold. A threshold comparing the wrong KL orientation would reverse which model is preferred.","primaryUrl":"https://arxiv.org/html/2510.08113","noteUrl":"/review?paper=review-neurips-2026-149882","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149885","title":"CoT-Guard: Small Models for Strong Monitoring","authors":["Nirav Diwan","Han Wang","Berkcan Kapusuzoglu","Ramin Moradi","Supriyo Chakraborty","Giridharan Iyengar","Sambit Sahu","Huan Zhang","Gang Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CoT-Guard: Small Models for Strong Monitoring","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12746","noteUrl":"/review?paper=review-neurips-2026-149885","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149886","title":"Revisiting Cross-View Completion: Self-Supervised Pre-Training via Reconstruction Error Comparison","authors":["Thibaut Loiseau","Guillaume Bourmaud","Vincent Lepetit"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Revisiting Cross-View Completion: Self-Supervised Pre-Training via Reconstruction Error Comparison","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.01530","noteUrl":"/review?paper=review-neurips-2026-149886","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149887","title":"Diff-Aid: Inference-time Adaptive Interaction Denoising for Rectified Text-to-Image Generation","authors":["Binglei Li","Mengping Yang","Zhiyu Tan","Qiang Xiang","Junping Zhang","Hao Li"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Diff-Aid: Inference-time Adaptive Interaction Denoising for Rectified Text-to-Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.13585","noteUrl":"/review?paper=review-neurips-2026-149887","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149889","title":"Fine-Tuning Language Models to Know What They Know","authors":["Sangjun Park","Elliot Meyerson","Xin Qiu","Risto Miikkulainen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Fine-Tuning Language Models to Know What They Know","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.02605","noteUrl":"/review?paper=review-neurips-2026-149889","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149891","title":"The Invisible Hand of Physics: When Video Diffusion Models Know More Than They Show","authors":["Parsa Esmati","Somjit Nath","Katja Hofmann","Derek Nowrouzezahrai","Samira Ebrahimi Kahou","Majid Mirmehdi"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"First differentiability does not justify the stated higher-order remainder","plainExplanation":"Continuous first derivatives give remainder \\(h\\omega(h)=o(h)\\), without a linear bound on the derivative modulus. The spatially constant velocity \\(v(t)=t^{3/2}\\) is \\(C^1\\) and Lipschitz in state. At zero, its first derivative vanishes, but the explicit/implicit discrepancy is \\(-h^{5/2}\\); divided by \\(h^3\\), its magnitude diverges. The higher-order remainder is unsupported. The leading global first-order result may still hold under weaker assumptions.","primaryUrl":"https://arxiv.org/html/2606.05328","noteUrl":"/review?paper=review-neurips-2026-149891","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149892","title":"CSLA: Sparse-Linear Attention with Learnable Routing and Quantization-aware Training","authors":["Jintao Zhang","Haoxu Wang","Kai Jiang","Kaiwen Zheng","Youhe Jiang","Ion Stoica","Jianfei Chen","Jun Zhu","Joseph Gonzalez"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"CSLA: Sparse-Linear Attention with Learnable Routing and Quantization-aware Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.12675v1","noteUrl":"/review?paper=review-neurips-2026-149892","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149893","title":"CurveRL: Principled Distribution-Aware Context Reweighting for LLM Reasoning","authors":["Ke Sun","Yizhou Zhao","Jiayi Xin","Qi Long","Weijie Su"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"CurveRL: Principled Distribution-Aware Context Reweighting for LLM Reasoning","plainExplanation":"A nonuniform reference can agree with the uniform CDF on every pass rate reached by the current policy family, giving the same nonzero parameter gradients. Uniformity remains sufficient. Necessity requires stronger pointwise functional-derivative quantifiers than the stated expected parameter-gradient comparison.","primaryUrl":"https://arxiv.org/html/2605.24331","noteUrl":"/review?paper=review-neurips-2026-149893","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149894","title":"MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction","authors":["Jianing Zhang","Chenhao Zheng","Yajun Yang","Rustin Soraki","Winson Han","Chun-Liang Li","Jason Ren","Max Argus","Jieyu Zhang","Ranjay Krishna"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.18558","noteUrl":"/review?paper=review-neurips-2026-149894","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149895","title":"Towards Reliable LLM Evaluation: Correcting the Winner’s Curse in Adaptive Benchmarking","authors":["Yang Xu","Jiefu Zhang","Haixiang Sun","Zihan Zhou","Tianyu Cao","Vaneet Aggarwal"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Towards Reliable LLM Evaluation: Correcting the Winner’s Curse in Adaptive Benchmarking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05973","noteUrl":"/review?paper=review-neurips-2026-149895","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149897","title":"Do Reasoning LLMs Refuse What They Infer in Long Contexts?","authors":["Yu Fu","Haz Sameen Shahgir","Zhipeng Wei","Huanli Gong","N. Benjamin Erichson","Yue Dong"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Do Reasoning LLMs Refuse What They Infer in Long Contexts?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.08874","noteUrl":"/review?paper=review-neurips-2026-149897","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149898","title":"HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling","authors":["Jonathan Cederlund","Axel Berg","Durmus Alp Emre Acar","Chuteng Zhou","Pontus Giselsson"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The head-count formula needs cardinality and feasibility bounds","plainExplanation":"The scalar memory inequality \\(Tc_k-N(c_k-c_s)\\leq bTC\\) gives a real lower bound, but a head count also obeys \\(0\\leq N\\leq T\\). If the layer already fits, the printed ceiling can be negative. If \\(bC<c_s\\), even pruning every head leaves the protected sink cost above budget, so no allowed count works. The printed minimum formula is incomplete on its stated domain. This concerns the formal budget rule, not measured compression results.","primaryUrl":"https://arxiv.org/html/2605.14877","noteUrl":"/review?paper=review-neurips-2026-149898","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149904","title":"Knowledge Transfer Scaling Laws for 3D Medical Imaging","authors":["Ho Hin Lee","Dongna Du","Chu Wang","Yuankai Huo","Shi Gu","James Gee","Yifan Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Knowledge Transfer Scaling Laws for 3D Medical Imaging","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06859","noteUrl":"/review?paper=review-neurips-2026-149904","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149905","title":"What to Remember, What to Reveal: Privacy-Aware Memory for Conversational Agents","authors":["Wenjie Wang","Wenhe Si","Xinyue Xu","Yue Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"What to Remember, What to Reveal: Privacy-Aware Memory for Conversational Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.16551","noteUrl":"/review?paper=review-neurips-2026-149905","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149908","title":"LOSCAR-SGD: Local SGD with Communication-Computation Overlap and Delay-Corrected Sparse Model Averaging","authors":["Peter Richtarik","Yassine Maziane","Ammar Mahran","Artavazd Maranjyan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LOSCAR-SGD: Local SGD with Communication-Computation Overlap and Delay-Corrected Sparse Model Averaging","plainExplanation":"Substituting the proof’s step size into its convergence bound gives an inverse-squared accuracy term, and a bounded-noise example rules out the stronger printed rate. The correction affects the first stochastic term; the following minibatch discussion already uses the squared denominator.","primaryUrl":"https://arxiv.org/html/2605.20866","noteUrl":"/review?paper=review-neurips-2026-149908","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149909","title":"AtomMOF: All-Atom Flow Matching for MOF-Adsorbate Structure Prediction","authors":["Nayoung Kim","Honghui Kim","Sihyun Yu","Minkyu Kim","Seongsu Kim","Sungsoo Ahn"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The particle transition and terminal target need two clarifications","plainExplanation":"As written, the transition can assign different deterministic drifts to identical resampled states because it mixes two particle ancestors.","primaryUrl":"https://arxiv.org/html/2602.07351","noteUrl":"/review?paper=review-neurips-2026-149909","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149910","title":"Efficient and Robust Physical 3DGS-MPM Simulation via Interior Filling and Text-Physics Optimization","authors":["Yikun Ma","Yiqing Li","Jingwen Ye","Zhongkai Wu","Weidong Zhang","Lin Gao","Zhi Jin"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The guidance rules use different derivative variables and targets","plainExplanation":"The guidance descriptions differentiate with respect to different material variables and use different targets. The strain-yield explanation also reverses its modulus dependence.","primaryUrl":"https://arxiv.org/html/2602.01723v1","noteUrl":"/review?paper=review-neurips-2026-149910","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149913","title":"Equilibrium Forcing: Adaptive Video Generation Without Noise Conditioning","authors":["Hansen Lillemark","Alex Rojas","Zachary Novack","Runqian (Ray) Wang","Yilun Du","Yian Ma","Taylor Berg-Kirkpatrick","Rose Yu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Equilibrium Forcing: Adaptive Video Generation Without Noise Conditioning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.14706","noteUrl":"/review?paper=review-neurips-2026-149913","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149914","title":"Decoupled Optimization for Teacher-Student Semi-Supervised Learning via a Pioneer Student","authors":["Haorong Han","Jidong Yuan","Chixuan Wei","Yongqi Sun"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Decoupled Optimization for Teacher-Student Semi-Supervised Learning via a Pioneer Student","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.09609","noteUrl":"/review?paper=review-neurips-2026-149914","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149915","title":"Causal Bias Detection in Generative Artifical Intelligence","authors":["Drago Plecko"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Positive exogenous support does not ensure cross-regime overlap","plainExplanation":"Let \\(X\\) be a fair binary exogenous variable, \\(Z=0\\), and \\(W=X\\). Model A has \\(Y=0\\); model B has \\(Y=\\mathbf1\\{X=1,W=0\\}\\). Both produce the identical observed law with \\(Y=0\\), but \\(Y_{1,W_0}\\) is zero in A and one in B, including when factual \\(X=0\\). Positive-support exogenous noise does not populate a parent configuration excluded by deterministic structural maps. The general identification claim lacks support assumptions. The decomposition on overlapping configurations is not refuted.","primaryUrl":"https://arxiv.org/pdf/2605.11365v2","noteUrl":"/review?paper=review-neurips-2026-149915","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149918","title":"M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals","authors":["Vinícius Da Silva","Isabelle R de Melo","Matheus Levy de Lima Bessa","Guilherme Schardong","Luiz Schirmer","Andre Araujo","Nuno Gonçalves","HELIO C LOPES","Alberto Raposo","Luiz Velho","Tiago Novello"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2609.28684","noteUrl":"/review?paper=review-neurips-2026-149918","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149921","title":"VerifyMAS: Hypothesis Verification for Failure Attribution in LLM Multi-Agent Systems","authors":["Hezhe Qiao","Hanghang Tong","Ee-peng Lim","Bing Liu","Guansong Pang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"VerifyMAS: Hypothesis Verification for Failure Attribution in LLM Multi-Agent Systems","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.17467","noteUrl":"/review?paper=review-neurips-2026-149921","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149922","title":"Fast-WAM: Do World Action Models Need Test-time Future Imagination?","authors":["Tianyuan Yuan","Zibin Dong","Yicheng Liu","Hang Zhao"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Fast-WAM: Do World Action Models Need Test-time Future Imagination?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.16666","noteUrl":"/review?paper=review-neurips-2026-149922","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149923","title":"CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models","authors":["Zhenghao He","Guangzhi Xiong","Boyang Wang","Sanchit Sinha","Aidong Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.15441","noteUrl":"/review?paper=review-neurips-2026-149923","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149924","title":"Manifold Sampling via Entropy Maximization","authors":["Cornelius Braun","Tilman Burghoff","Marc Toussaint"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Manifold Sampling via Entropy Maximization","plainExplanation":"No issue was found in the checked claims. A zero initial component mass violates the stated prerequisite for the theorem’s analysis. It therefore does not contradict the convergence statement in its surrounding context.","primaryUrl":"https://arxiv.org/html/2605.12338","noteUrl":"/review?paper=review-neurips-2026-149924","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149925","title":"What Is Worth Representing? Representational Empowerment for Continual Model Construction","authors":["Fei Dai","Hanqi Zhou","Alison Gopnik","Charley M Wu"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"What Is Worth Representing? Representational Empowerment for Continual Model Construction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.02322","noteUrl":"/review?paper=review-neurips-2026-149925","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149927","title":"Flexible Routing via Uncertainty Decomposition","authors":["Charlotte Peale","Siddartha Devic","Parikshit Gopalan","Udi Wieder","Aravind Gollakota"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Flexible Routing via Uncertainty Decomposition","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07805","noteUrl":"/review?paper=review-neurips-2026-149927","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149928","title":"SemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete Data","authors":["Wenhao Li","Zhibin Wu","Chong Xiao","Qiangchang Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30238","noteUrl":"/review?paper=review-neurips-2026-149928","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149929","title":"ChainFlow-VLA: Causal Flow Planning with Vision-Language Models","authors":["Xiyang Wang","Xinlin Wang","Tingguang Zhou","Gong Chen","Xingtai Gui","Zhi Xu","Hangning Zhou","Xiaolei Wu","Feiyang Tan","Mu Yang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ChainFlow-VLA: Causal Flow Planning with Vision-Language Models","plainExplanation":"No issue was found in the checked claims. Deterministic proposal centers do not make a discrete mixture invalid. The paper does not identify its final utility-maximizing decision with a random draw from that mixture.","primaryUrl":"https://arxiv.org/html/2605.23270v1","noteUrl":"/review?paper=review-neurips-2026-149929","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149930","title":"DynaTokens: Teaching Dynamics to Camera-Controlled Video Models at Test Time","authors":["Ziqi Ma","Hongqiao Chen","Georgia Gkioxari"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The observer-frame set must use the inverse translation","plainExplanation":"The static object’s displayed observer-frame set uses the forward translation instead of its inverse. Its motion is therefore opposite to the explicitly defined observer-to-world map. This is a local sign typo in the motivating example, rather than a defect in the camera-control model.","primaryUrl":"https://arxiv.org/html/2609.35704","noteUrl":"/review?paper=review-neurips-2026-149930","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149932","title":"MetaKE: Meta-Learning for Knowledge Editing Toward a Better Accuracy-Editability Trade-off","authors":["Shuxin Liu","Di Gao","Ou Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MetaKE: Meta-Learning for Knowledge Editing Toward a Better Accuracy-Editability Trade-off","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.12677","noteUrl":"/review?paper=review-neurips-2026-149932","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149934","title":"Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network","authors":["Grigory Bartosh","David Ruhe","Emiel Hoogeboom","Jonathan Heek","Thomas Mensink","Tim Salimans"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18190","noteUrl":"/review?paper=review-neurips-2026-149934","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149935","title":"EM-NeSy: Expectation Maximization for Neurosymbolic Learning","authors":["Annegret Seibt","Luc De Raedt","Giuseppe Marra"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The rejection-sampling gradient loses its leading minus sign","plainExplanation":"The rejection-sampling gradient drops the leading minus sign when a posterior-weighted sum is rewritten as an expectation. Changing the sampling measure does not remove that sign. This is a local sign slip in Section 4.2. The appendix retains the correct gradient; no implementation failure follows.","primaryUrl":"https://arxiv.org/html/2606.14463","noteUrl":"/review?paper=review-neurips-2026-149935","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-149938","title":"Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information","authors":["Guobin Shen","Xiang Cheng","Chenxiao Zhao","Lei Huang","Jindong Li","Dongcheng Zhao","XingYu Li"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information","plainExplanation":"Shared network parameters do not imply a coherent conditional–marginal probability law. Even with that law, a trajectory-dependent terminal potential can change optimal policies, and feedback-dependent context selection can reverse the expected sampled JSD direction. The fixed-context JSD identity remains valid. A separate local counting error affects the prompt-template reference answer.","primaryUrl":"https://arxiv.org/html/2605.11609","noteUrl":"/review?paper=review-neurips-2026-149938","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149939","title":"Large Discrete Policy: Advancing Explicit Behavior Modeling with Stochastic Iterative Scoring","authors":["Zhenxin Li","Nadine Chang","Xinglong Sun","Jingde Chen","Wenhao Yao","Zi Wang","Maying Shen","Yu-Gang Jiang","Zuxuan Wu","Shiyi Lan","Jose M. Alvarez"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Large Discrete Policy: Advancing Explicit Behavior Modeling with Stochastic Iterative Scoring","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.07049","noteUrl":"/review?paper=review-neurips-2026-149939","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149940","title":"Probabilistic Signature Inversion: Learning Conditional Distributions from Truncated Signatures","authors":["Junoh Kang","Kiseop Lee","Bohyung Han"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Probabilistic Signature Inversion: Learning Conditional Distributions from Truncated Signatures","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.15332","noteUrl":"/review?paper=review-neurips-2026-149940","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149942","title":"A Distribution Mapping Approach to Counterfactually Fair Reinforcement Learning","authors":["Jianhan Zhang","Jitao Wang","John D Piette","Donglin Zeng","Chengchun Shi","Zhenke Wu"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"A Distribution Mapping Approach to Counterfactually Fair Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.08743","noteUrl":"/review?paper=review-neurips-2026-149942","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149943","title":"Topological Periodicity Test (TopPT) via Confidence Bound of Time-Delay Embeddings","authors":["Donghyun Park","Junhyun An","Taehyoung Kim","Jisu Kim"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Topological Periodicity Test (TopPT) via Confidence Bound of Time-Delay Embeddings","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.06324","noteUrl":"/review?paper=review-neurips-2026-149943","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149946","title":"Stochastic Interpolants via Conditional Dependent Coupling","authors":["Chenrui Ma","Xi Xiao","Tianyang Wang","Xiao Wang","Yanning Shen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The cascade comparisons need their noise and runtime assumptions stated explicitly","plainExplanation":"The advantage is not uniform over every permitted noise schedule. At \\(\\gamma=1\\), extra stages can add noise cost without reducing the final-stage variance.","primaryUrl":"https://arxiv.org/html/2509.23122","noteUrl":"/review?paper=review-neurips-2026-149946","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149948","title":"Locking Pretrained Weights via Deep Low-Rank Residual Distillation","authors":["Keitaro Sakamoto","Pierre Ablin","Federico Danieli","Marco Cuturi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Locking Pretrained Weights via Deep Low-Rank Residual Distillation","plainExplanation":"The conditioning, activation-memory and utility claims need different scope corrections. The utility counterexample uses a permitted rank-one update that improves the printed objective while increasing a bounded task loss.","primaryUrl":"https://arxiv.org/html/2605.10777","noteUrl":"/review?paper=review-neurips-2026-149948","collection":"neurips-2026","severity":"major","issueTypes":["complexity_error","false_statement","formula_error","invalid_inference","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-149949","title":"Random Matrix Theory of Early-Stopped Gradient Flow: A Transient BBP Scenario","authors":["Florentin Coeurdoux","Grégoire Ferré","jean-philippe bouchaud"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The spectral formulas need a consistent resolvent sign and a teacher-orientation condition","plainExplanation":"The printed Dyson equation and asymptotic condition belong to the opposite resolvent convention. They cannot be combined with the stated negative-imaginary branch and density sign.","primaryUrl":"https://arxiv.org/html/2604.18450","noteUrl":"/review?paper=review-neurips-2026-149949","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149954","title":"S$^{2}$-PINN: Stochastic Separable Physics-Informed Neural Networks","authors":["Zhendong Li","Akwum Onwunta"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"S$^{2}$-PINN: Stochastic Separable Physics-Informed Neural Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03303","noteUrl":"/review?paper=review-neurips-2026-149954","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149956","title":"Entropy Dynamics of Agent Reinforcement Learning","authors":["Wendi Li","Shawn Im","Sharon Li"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Entropy Dynamics of Agent Reinforcement Learning","plainExplanation":"The derivative scales and the optimization coordinates matter alongside the probability coefficients. An admissible two-bit policy reverses the claimed dominance while satisfying the stated probability ordering. The exact product-rule identities remain valid; a comparison needs explicit derivative and update-metric assumptions.","primaryUrl":"https://arxiv.org/html/2605.27954v1","noteUrl":"/review?paper=review-neurips-2026-149956","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149957","title":"CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models","authors":["Joowon Kim","Seungho Shin","Joonhyung Park","Eunho Yang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The return must concatenate accepted history rather than generated indices","plainExplanation":"The return indexes generated clips using accepted-history length, which can select a rejected clip after an earlier rejection.","primaryUrl":"https://arxiv.org/html/2605.08735","noteUrl":"/review?paper=review-neurips-2026-149957","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149960","title":"CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models","authors":["Chunzheng Zhu","Jiaqi Zeng","Hongbo Zhao","Yihang Chen","Yijun Wang","Jianxin Lin"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The patch contrast is not a direct effect of conflict treatment","plainExplanation":"The equation changes \\(H\\) between clean and conflict values while holding conflict context fixed. A direct effect of conflict \\(T\\) instead compares \\(Y(1,h)\\) and \\(Y(0,h)\\) at the same \\(H=h\\). In \\(H=T+U,Y=2T+3H\\), paired \\(U=0\\) gives the activation-patch contrast at \\(T=1\\) as \\(-3\\), while the conflict direct effect at \\(H=0\\) is \\(2\\). The controlled head contrast can be valid and causally useful. It is a different estimand, so patching effectiveness is not refuted.","primaryUrl":"https://arxiv.org/html/2609.38810","noteUrl":"/review?paper=review-neurips-2026-149960","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149962","title":"Learnable Low-Rank Polynomial Sketch for Effective Linear Attention","authors":["Zongyue Qin","Haoran Deng","Jason Cong","Yizhou Sun"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Learnable Low-Rank Polynomial Sketch for Effective Linear Attention","plainExplanation":"Placing a CP coefficient in every factor raises it to the polynomial degree; placing it in one factor restores the projector identity. Separately, an included baseline can already represent the kernel exactly, so a larger feature family need not improve its optimal error strictly. Neither correction disproves the expressive power of unconstrained learned factors.","primaryUrl":"https://arxiv.org/html/2610.14219","noteUrl":"/review?paper=review-neurips-2026-149962","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149963","title":"Attention-Based Sampler for Diffusion Language Models","authors":["Yuyan Zhou","Kai Syun Hou","Weiyu Chen","James Kwok"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Attention-Based Sampler for Diffusion Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.08564","noteUrl":"/review?paper=review-neurips-2026-149963","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149965","title":"Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes","authors":["Yu Wang","Rui Dai","Wei Chen","Yi Wang","Bo Dang","Kaikui Liu","Xiangxiang Chu","Yansheng Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.18356","noteUrl":"/review?paper=review-neurips-2026-149965","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149966","title":"Joint Learning of Hierarchical Neural Options and Abstract World Model","authors":["Top Piriyakulkij","Wolfgang Lehrach","Kevin Ellis","Kevin Murphy"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Joint Learning of Hierarchical Neural Options and Abstract World Model","plainExplanation":"An option can preserve hidden state information while its terminal abstraction depends only on the current abstraction. A single weighting law inside each abstract class then cannot reconstruct every start-dependent terminal distribution. Full-state exactness needs the stronger common conditional-law assumption; abstract transition closure is a separate property.","primaryUrl":"https://arxiv.org/html/2602.02799","noteUrl":"/review?paper=review-neurips-2026-149966","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149969","title":"Identified-Set Geometry of Distributional Model Extraction under Top-K Censored API Access","authors":["Wenhua Nie","ZiCheng Zhu","Jianan Wu","Binhan Luo","Haoran Zheng","Jyh-Shing R Jang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Identified-Set Geometry of Distributional Model Extraction under Top-K Censored API Access","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10407","noteUrl":"/review?paper=review-neurips-2026-149969","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149970","title":"RECIPE: Procedural Planning via Grounding in Instructional Video","authors":["Luigi Seminara","Antonino Furnari","Lorenzo Torresani"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RECIPE: Procedural Planning via Grounding in Instructional Video","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19976","noteUrl":"/review?paper=review-neurips-2026-149970","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149971","title":"Think, Plan, Paint: Layout-Aware Reasoning for Controllable Image Generation in Unified Models","authors":["Junhao Liu","Jian-Wei Zhang","Tao Huang","Miles Yang","Zhao Zhong","Liefeng Bo"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Think, Plan, Paint: Layout-Aware Reasoning for Controllable Image Generation in Unified Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.16409","noteUrl":"/review?paper=review-neurips-2026-149971","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149972","title":"Mean-Field Control on Sparse Graphs: From Local Limits to GNNs via Neighborhood Distributions","authors":["Tobias Schmidt","Kai Cui"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Mean-Field Control on Sparse Graphs: From Local Limits to GNNs via Neighborhood Distributions","plainExplanation":"The locality proof uses an information cone with the wrong time origin, and the Bellman recursion mixes two discount conventions. In the finite infection model, disconnected components also prevent the scattered seeds from exceeding the stated strict threshold. The separate truncation-rate comparison establishes no error in the theorem’s exponential rate.","primaryUrl":"https://arxiv.org/html/2601.21477","noteUrl":"/review?paper=review-neurips-2026-149972","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149974","title":"AI Models Can Provably Hide Arbitrary Capabilities","authors":["Andis Draguns","Sumeet Motwani","Raymond Douglas","Christian Schroeder de Witt"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AI Models Can Provably Hide Arbitrary Capabilities","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.13928","noteUrl":"/review?paper=review-neurips-2026-149974","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149979","title":"Tunable Latent Generative Priors for Compressed Sensing and Inverse Problems","authors":["Sean Gunn","Jorio Cocola","Paul Hand","Oliver De Candido","Vaggos Chatziafratis"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Tunable Latent Generative Priors for Compressed Sensing and Inverse Problems","plainExplanation":"An SVD rotation preserves Gaussian distributions but need not preserve the estimator’s allowed coordinates. The risk formula is valid for a singular-coordinate estimator, rather than the printed fixed-coordinate restriction.","primaryUrl":"https://arxiv.org/html/2603.07357","noteUrl":"/review?paper=review-neurips-2026-149979","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference"],"classifications":["major"]},{"slug":"review-neurips-2026-149981","title":"SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators","authors":["Yuncong Yang","Zhengtao Han","Furkan Ozyurt","Zeyuan Yang","Han Yang","Junyi Cao","Haoyu Zhen","Yilun Du","Chuang Gan"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.09155","noteUrl":"/review?paper=review-neurips-2026-149981","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149985","title":"Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning","authors":["Yunan Wang","Minghui Song","Zihan Zhang","Shaohan Huang","Haizhen Huang","Furu Wei","Weiwei Deng","Feng Sun","Qi Zhang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.22995","noteUrl":"/review?paper=review-neurips-2026-149985","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-149988","title":"Generative Conformal Prediction with Optimized Coverage Allocation","authors":["Minxing Zheng","Shixiang Zhu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Generative Conformal Prediction with Optimized Coverage Allocation","plainExplanation":"The density-ranking and vanishing-radius arguments need stronger conditions. The finite proxy guarantee has a direct repair: cap each greedy jump’s gain at the remaining coverage quota.","primaryUrl":"https://arxiv.org/html/2410.13735","noteUrl":"/review?paper=review-neurips-2026-149988","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","invalid_inference","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-149989","title":"TrioPose: Native Triple-Stream Diffusion Transformers for Pose-Guided Text-to-Image Generation","authors":["dian gu","Zhengyi Yang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"TrioPose: Native Triple-Stream Diffusion Transformers for Pose-Guided Text-to-Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.07053","noteUrl":"/review?paper=review-neurips-2026-149989","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149990","title":"Modeling quantum neural network gradient with reinforcement learning","authors":["Nhan Luu","Trung D Luu","Ngoc Nam Pham","Thang C Truong"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Modeling quantum neural network gradient with reinforcement learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31066","noteUrl":"/review?paper=review-neurips-2026-149990","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149994","title":"GridProbe: Posterior-Probing for Adaptive Test-Time Compute in Long-Video VLMs","authors":["Mohamed Eltahir","Lama Ayash","Ali Habibullah","Tanveer Hussain","Naeemullah Khan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Nonuniform importance does not force a reduced effective budget","plainExplanation":"An admissible nonuniform importance map can retain every token, so nonuniformity alone does not guarantee subquadratic attention.","primaryUrl":"https://arxiv.org/html/2605.10762","noteUrl":"/review?paper=review-neurips-2026-149994","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-149996","title":"Witness Overlap: Directional Provenance Inside Open-Weight Model Families","authors":["Siyuan Li","Haoxuan Zeng","Xin Luo","Fernando Jia","Florence Li","Zhengyang Geng","Zico Kolter","Tai Sing Lee","Tianqin Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Witness Overlap: Directional Provenance Inside Open-Weight Model Families","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31784","noteUrl":"/review?paper=review-neurips-2026-149996","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149997","title":"RIZZ: Routing Interactions to Near Zero-Interference Zones for Continual Adaptation of Black-Box Agents","authors":["Sonali Goel","Pranav Vaidhyanathan","Lucas Schorling","Natalia Ares","Michael A Osborne"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RIZZ: Routing Interactions to Near Zero-Interference Zones for Continual Adaptation of Black-Box Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.20638","noteUrl":"/review?paper=review-neurips-2026-149997","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-149999","title":"Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws","authors":["Zhiwei Xu","Shihao Wu","Hanseul Cho","Wei Hu","Yixin Wang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06888","noteUrl":"/review?paper=review-neurips-2026-149999","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150000","title":"Retrieval-Centric Deep Learning in Growing Nonparametric Neural Networks","authors":["Maximilian Schlegel","Rajai Nasser","Seijin Kobayashi","Yanick Schimpf","Oliver Sieberling","Robert Obryk","Kazuki Irie","João Sacramento","Johannes von Oswald"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The momentum impulse enters one step after its index","plainExplanation":"The momentum appendix gives each gradient impulse one extra update of influence; the later appendix supplies a coherent creation-time convention.","primaryUrl":"https://arxiv.org/html/2610.03858","noteUrl":"/review?paper=review-neurips-2026-150000","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150001","title":"HyFAD: Hybrid Time-Frequency Diffusion with Frequency-Aware Embedding for Time Series Imputation","authors":["Hongfan Gao","Wangmeng Shen","Bin Yang","Jilin Hu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"HyFAD: Hybrid Time-Frequency Diffusion with Frequency-Aware Embedding for Time Series Imputation","plainExplanation":"Earlier time-domain noise must pass through later frequency attenuation, changing the accumulated noise and covariance. The printed training and sampling maps also need consistent cumulative means, noise-target normalization and frequency-noise factors.","primaryUrl":"https://arxiv.org/html/2606.05239","noteUrl":"/review?paper=review-neurips-2026-150001","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","mathematical_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150004","title":"Edge of Stochastic Stability: Revisiting the Edge of Stability for SGD","authors":["Arseniy Andreyev","Pierfrancesco Beneventano"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Edge of Stochastic Stability: Revisiting the Edge of Stability for SGD","plainExplanation":"Batch Sharpness does not provide the claimed general stochastic-instability certificate. Local moment, sampling and quantifier corrections retain the valid stationary identity when its mean is zero.","primaryUrl":"https://arxiv.org/html/2412.20553","noteUrl":"/review?paper=review-neurips-2026-150004","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","mathematical_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150007","title":"Energy-Tweedie: Score meets Score, Energy meets Energy","authors":["Andrej Leban"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Energy-Tweedie: Score meets Score, Energy meets Energy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.23818","noteUrl":"/review?paper=review-neurips-2026-150007","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150008","title":"Response Time Enhances Alignment with Heterogeneous Preferences","authors":["Federico Echenique","Alireza Fallah","Baihe Huang","Michael Jordan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Cauchy–Schwarz gives a square root, not its reciprocal","plainExplanation":"The proof prints the reciprocal of the Cauchy–Schwarz bound. Its negative exponent diverges when the error second moment vanishes, rather than yielding the claimed convergence to zero. This is a local exponent typo in the convergence proof. The preceding envelope argument supports the corrected step; the main consistency result is not refuted by it.","primaryUrl":"https://arxiv.org/html/2605.06987","noteUrl":"/review?paper=review-neurips-2026-150008","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150009","title":"Grounding 3D Affordance from Human-Object-Interaction Videos via Multimodal Large Language Model","authors":["Hanqing Wang","Mingyu Liu","Xiaoyu Chen","Chengwei MA","Yiming Zhong","Yuhao Liu","Wenti Yin","Xing Mu","Jiahao Yuan","Zhiqing Cui","Lu Dai","Zhiyuan Ma","Hui Xiong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The spatial weight needs an isolated-point convention, and attention needs explicit shapes","plainExplanation":"An empty neighborhood makes the quotient undefined. The later stabilizer in the Dice-loss denominator does not define this earlier weight.","primaryUrl":"https://arxiv.org/html/2602.09638","noteUrl":"/review?paper=review-neurips-2026-150009","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150010","title":"REACT: Physically and Chemically Consistent Reconstruction of Marine Active Tracers","authors":["Wenbin Dai","Hao Zheng","Shiyu Liang","Chaofan Sun","Xueying Zhang","HanBo Huang","Xuan Gong","Yiran Zhang","Enhui Liao"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"REACT: Physically and Chemically Consistent Reconstruction of Marine Active Tracers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03888","noteUrl":"/review?paper=review-neurips-2026-150010","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150011","title":"Don't Always Pick the Highest-Performing Model: An Information Theoretic View of LLM Ensemble Selection","authors":["Yigit Turkmen","Baturalp Buyukates","Melih Bastopcu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Don't Always Pick the Highest-Performing Model: An Information Theoretic View of LLM Ensemble Selection","plainExplanation":"Strictly degrading one channel need not strictly worsen Bayes error or mutual information when another channel already determines the decision or label. The weak dominance result and log-odds voting rule remain valid under the balanced independent binary-symmetric-channel model.","primaryUrl":"https://arxiv.org/html/2602.08003","noteUrl":"/review?paper=review-neurips-2026-150011","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150012","title":"Teaching Video Generators to Remember: Eliciting Dynamic Memory for Out-of-Sight State Evolution","authors":["Tianshuo Xu","Yichen Xie","Depu Meng","Chensheng Peng","Quentin HERAU","Bo Jiang","yihan hu","Wei Zhan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Teaching Video Generators to Remember: Eliciting Dynamic Memory for Out-of-Sight State Evolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25333","noteUrl":"/review?paper=review-neurips-2026-150012","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150016","title":"One Loss to Rule Them All: Marked Time-to-Event for Structured EHR Foundation Models","authors":["Zilin Jing","Vincent Jeanselme","Yuta Kobayashi","Simon Lee","Chao Pang","Aparajita Kashyap","Yanwei Li","Xinzhuo Jiang","Shalmali Joshi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Code-presence entropy must include the absence outcome","plainExplanation":"A presence probability \\(p(m)\\) describes a Bernoulli variable, whose entropy includes both \\(-p\\log p\\) and \\(-(1-p)\\log(1-p)\\). Different code-presence events do not form a categorical partition. The displayed conditional expression is \\(p(M=1)H(N\\mid M=1)\\), not \\(H(M\\mid N)\\). If \\(N\\) is always present and \\(P(M=1)=1/2\\), it is zero while the latter entropy is \\(\\log2\\). The entropy identities are incorrect. Renaming may preserve selection, while replacing them by true entropy can change rankings; actual implementation is not inferred.","primaryUrl":"https://arxiv.org/html/2602.00541","noteUrl":"/review?paper=review-neurips-2026-150016","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150017","title":"$\\boldsymbol{f}$-OPD: Stabilizing Long-Horizon On-Policy Distillation with Freshness-Aware Control","authors":["Xianwei Chen","Shimin Zhang","Jibin Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"$\\boldsymbol{f}$-OPD: Stabilizing Long-Horizon On-Policy Distillation with Freshness-Aware Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17862","noteUrl":"/review?paper=review-neurips-2026-150017","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150019","title":"GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis","authors":["Rasul Khanbayov","Erchin Serpedin","HASAN KURBAN"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30901","noteUrl":"/review?paper=review-neurips-2026-150019","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150020","title":"Asymptotically Optimal Best Arm Identification with Fixed-Budget under Differential Privacy","authors":["Keqin Chen","Jie Bian","Yulian Wu","Vincent Tan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Asymptotically Optimal Best Arm Identification with Fixed-Budget under Differential Privacy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04600","noteUrl":"/review?paper=review-neurips-2026-150020","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150022","title":"Select Smarter, Not More? Prompt-Aware Evaluation Scheduling with Submodular Guarantees","authors":["haoyue liu","Xiaoyu Ma","Yiwen li","Zhichao Wang","Shuguang Cui","Xiaoying Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Random warmup does not inherit the cold-start greedy guarantee","plainExplanation":"Random warmup does not supply the greedy approximation premise used to claim the warm-start guarantee.","primaryUrl":"https://arxiv.org/html/2604.11328","noteUrl":"/review?paper=review-neurips-2026-150022","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150023","title":"Beyond Raw Context Transfer: Representation-based Federated Retrieval-Augmented Generation","authors":["Can Peng","Yu Liu","Yingyu Yang","Anjie Le","Yuyuan Liu","Qianye Yang","Alison Noble"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Beyond Raw Context Transfer: Representation-based Federated Retrieval-Augmented Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.22162","noteUrl":"/review?paper=review-neurips-2026-150023","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150024","title":"Online Finetuning Decision Transformers with Pure RL Gradients","authors":["Junkai Luo","Yinglun Zhu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Online Finetuning Decision Transformers with Pure RL Gradients","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.00167","noteUrl":"/review?paper=review-neurips-2026-150024","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150025","title":"Accelerating LLM Pre-Training through Flat-Direction Dynamics Enhancement","authors":["Shuchen Zhu","Rizhen Hu","Mingze Wang","Mou Sun","xue wang","Kun Yuan","Zaiwen Wen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The current revision corrects the Riemannian transformation; three separate formulas still need adjustment","plainExplanation":"Differentiating the velocity relation produces a \\(\\dot\\gamma_t/\\gamma_t\\) damping term and divides \\(\\dot\\beta_t\\) by \\(\\gamma_t\\). Both are absent from the printed reduction.","primaryUrl":"https://arxiv.org/html/2602.22681","noteUrl":"/review?paper=review-neurips-2026-150025","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150028","title":"Computationally Efficient Replicable Learning of Parities and Applications","authors":["Moshe Noivirt","Jessica Sorrell","Eliad Tsfadia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Computationally Efficient Replicable Learning of Parities and Applications","plainExplanation":"The span-agreement proposition needs a lower bound on sample size, and the span procedure needs a convention for zero input vectors. The restricted sample-size argument and a zero-filtering rule address these local defects; an all-input replicability extension needs its own sensitivity argument.","primaryUrl":"https://arxiv.org/html/2602.09499","noteUrl":"/review?paper=review-neurips-2026-150028","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150029","title":"OmniMem: Scalable and Adaptive Memory Retrieval for Long Video Generation","authors":["Lin Zhao","Yushu Wu","Yifan Gong","Yanzhi Wang","Pu Zhao"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"An empty attention mask leaves the output denominator zero","plainExplanation":"An all-zero attention mask leaves both accumulator and normalizer zero, requiring an explicit empty-row convention.","primaryUrl":"https://arxiv.org/html/2605.30519","noteUrl":"/review?paper=review-neurips-2026-150029","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150031","title":"Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold","authors":["Ye He","Yitong Qiu","Molei Tao"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold","plainExplanation":"The ridge argument needs global geometric conditions and derivatives of the actual moving constraint. Several local formulas also have explicit corrections; entering a tube does not ensure that a diffusion remains there.","primaryUrl":"https://arxiv.org/html/2602.06021","noteUrl":"/review?paper=review-neurips-2026-150031","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","lemma_error","mathematical_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150033","title":"AnchorWorld: Embodied Egocentric World Simulation with View-based Evolution Customization","authors":["Yu Li","Menghan Xia","Gongye Liu","Xintao Wang","Conglang Zhang","Lei Ke","yuxuan lin","Ruihang Chu","Pengfei Wan","Kun Gai","Yujiu Yang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"AnchorWorld: Embodied Egocentric World Simulation with View-based Evolution Customization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.07326","noteUrl":"/review?paper=review-neurips-2026-150033","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150034","title":"RADAR: Text-Guided Medical Image Segmentation via Residual Aggregation and Dense Alignment Representations","authors":["Jie Gui","Hang Tu","Yunfeng Dou","Wen Sha","Xiuquan Du","Shuo Li"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"RADAR: Text-Guided Medical Image Segmentation via Residual Aggregation and Dense Alignment Representations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.14944","noteUrl":"/review?paper=review-neurips-2026-150034","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150036","title":"Beyond Bounded Variance: Variance-Reduced Normalized Methods for Nonconvex Optimization under Blum-Gladyshev Noise","authors":["Antesh Upadhyay","Arda Fazla","Abolfazl Hashemi"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The bounded-variance noise term has fourth-root time decay","plainExplanation":"The fresh-noise term has fourth-root rather than square-root time decay under the stated schedule. The cited finite-horizon experiment also violates the invoked stepsize condition.","primaryUrl":"https://arxiv.org/html/2605.15314","noteUrl":"/review?paper=review-neurips-2026-150036","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150037","title":"SGD in Multiclass Logistic Regression: Sequential Learning and Scaling Laws","authors":["Konstantinos Tsiolis","Denny Wu","Christos Thrampoulidis","Murat Erdogdu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SGD in Multiclass Logistic Regression: Sequential Learning and Scaling Laws","plainExplanation":"The compute-optimal extrapolation can leave the time window controlled by the learning theorem. Separately, a Gaussian sketch changes the noise covariance, so preserving mean differences does not preserve the original signal-to-noise ratio.","primaryUrl":"https://arxiv.org/html/2609.07868","noteUrl":"/review?paper=review-neurips-2026-150037","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150038","title":"ADAPT: Hybrid Prompt Optimization for LLM Feature Visualization","authors":["João N. Cardoso","Arlindo L Oliveira","Bruno Martins"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Clipping negative activation loss gives the wrong hardness metric","plainExplanation":"The hardness formula clips negative activation loss with the wrong sign, making every term zero for nonnegative features.","primaryUrl":"https://arxiv.org/html/2602.17867","noteUrl":"/review?paper=review-neurips-2026-150038","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150039","title":"Rethinking Psychometric Evaluation of LLMs: When and Why Self-Reports Predict Behavior","authors":["Rafal Kocielnik","Pengrui Han","Peiyang Song","Myrl Marmarelis","Ramit Debnath","Dean Mobbs","Animashree Anandkumar","R. Michael Alvarez"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The sensitivity ratio does not identify a stable latent source","plainExplanation":"Prompt effects alone can reproduce the proposed sensitivity ratio and within-session coherence, so the ratio does not identify a stable common latent source.","primaryUrl":"https://arxiv.org/html/2606.12730","noteUrl":"/review?paper=review-neurips-2026-150039","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150044","title":"Voronoi-Markov chain and spatial entropy for point pattern analysis","authors":["James C Mathews","Francisco Couzo","Aleksandr Petrov","Avijit Chatterjee","Saad Nadeem"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Voronoi-Markov chain and spatial entropy for point pattern analysis","plainExplanation":"An equilateral triangle already makes the printed area formula nonreal. A separate aperiodic-chain example distinguishes the claimed spectral gap from the actual contraction rate.","primaryUrl":"https://arxiv.org/html/2609.36158","noteUrl":"/review?paper=review-neurips-2026-150044","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","formula_error","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-150047","title":"Feature Learning Dynamics in Infinite-Depth Neural Networks","authors":["Zihan Yao","Ruoyu Wu","Tianxiang Gao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Feature Learning Dynamics in Infinite-Depth Neural Networks","plainExplanation":"The stated input training changes the first-layer innovation covariance. Its boundary update must also use the input inner product rather than an activation covariance.","primaryUrl":"https://arxiv.org/html/2512.21075","noteUrl":"/review?paper=review-neurips-2026-150047","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-150051","title":"B$^3$-PWL: GPU-Batched Branch-and-Bound for Piecewise-Linear Optimization with SOS2 Constraints","authors":["Yilin Guan","Shuqing Luo","Pingzhi Li","Tianlong Chen","Kaidi Xu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"B$^3$-PWL: GPU-Batched Branch-and-Bound for Piecewise-Linear Optimization with SOS2 Constraints","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.28988","noteUrl":"/review?paper=review-neurips-2026-150051","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150052","title":"Decoupling Time and Risk: Risk-Sensitive Reinforcement Learning with General Discounting","authors":["Mehrdad Moghimi","Anthony Coache","Hyejin Ku"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Decoupling Time and Risk: Risk-Sensitive Reinforcement Learning with General Discounting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04131","noteUrl":"/review?paper=review-neurips-2026-150052","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150055","title":"Next Forcing: Causal World Modeling with Multi-Chunk Prediction","authors":["Gangwei Xu","Qihang Zhang","Jiaming Zhou","Xing Zhu","Yujun Shen","Xin Yang","Yinghao Xu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Next Forcing: Causal World Modeling with Multi-Chunk Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11187","noteUrl":"/review?paper=review-neurips-2026-150055","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150056","title":"SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting","authors":["Weijie Shi","Qiang Xu","fan deng","Yaguang Wu","Jiarun Liu","Yehong Xu","Hao Chen","Jia Zhu","Jiajie Xu","Xiangjun Huang","jian yang","Xiaofang Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07243","noteUrl":"/review?paper=review-neurips-2026-150056","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150057","title":"MoCAR: Motion-code Coordinate-aware AutoRegression for Continuous Trajectory Forecasting","authors":["Yiming Xu","Hao Cheng","Monika Sester"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MoCAR: Motion-code Coordinate-aware AutoRegression for Continuous Trajectory Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06210","noteUrl":"/review?paper=review-neurips-2026-150057","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150058","title":"Disentangling Dual Image References in Frequency Aware Diffusion Models for Personalized Generation","authors":["Haipeng Liu","Yang Wang","Meng Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Tokenwise entropy needs scalar aggregation for a loop endpoint","plainExplanation":"The entropy rule produces one value per image token but is used as a scalar loop endpoint without a token-reduction definition.","primaryUrl":"https://arxiv.org/html/2610.07684","noteUrl":"/review?paper=review-neurips-2026-150058","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150059","title":"Communication-Efficient Personalized Adaptation via Federated-Local Model Merging","authors":["Yinan Zou","Md Kamran Chowdhury Shisher","Christopher Brinton","Vishrant Tripathi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The exact Gaussian model is incompatible with an almost-sure bounded basin","plainExplanation":"A nondegenerate Gaussian marginal has unbounded support, so these exact assumptions cannot both describe the same random task vectors. A Gaussian tail example does not refute a theorem whose basin premise it violates.","primaryUrl":"https://arxiv.org/html/2602.18658","noteUrl":"/review?paper=review-neurips-2026-150059","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150061","title":"Seg3DParts: Segmentation-Grounded Controllable Part-Level 3D Generation","authors":["Jiantao Lin","Meixi Chen","Yingjie Xu","Chenbo FU","Leyi Wu","Hao CHEN","Yinchuan Li","Yingcong Chen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Seg3DParts: Segmentation-Grounded Controllable Part-Level 3D Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36918","noteUrl":"/review?paper=review-neurips-2026-150061","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150062","title":"DiLaDiff: Distilled Latent-augmented Diffusion for Language Modeling","authors":["Jean-Marie Lemercier","Tomas Geffner","Morteza Mardani","Karsten Kreis","Arash Vahdat","Ante Jukić"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The posterior justification needs conditional latent weights, and the likelihood must select clean tokens","plainExplanation":"When \\(x_t\\) is generated from a clean sequence depending on \\(z\\), observing \\(x_t\\) generally changes the latent law. The conditional-independence lemma does not justify keeping prior latent weights.","primaryUrl":"https://arxiv.org/html/2605.23605","noteUrl":"/review?paper=review-neurips-2026-150062","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150063","title":"Optimal Learning-Augmented Algorithm for Online Bidding","authors":["Changyeol Lee","Dahoon Lee","Jongseo Lee","Yongho Shin","Changki Yun"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Optimal Learning-Augmented Algorithm for Online Bidding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07349","noteUrl":"/review?paper=review-neurips-2026-150063","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150068","title":"StreamOV: Streaming Omni-Video Understanding via Evidence-Guided Memory and Response Triggering","authors":["Ming Xie","Zizheng Huang","Xudong Tan","Chao Wang","Xiangyu Zeng","Wenxiao Wu","Tao Chen","Limin Wang","Yanwei Fu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"StreamOV: Streaming Omni-Video Understanding via Evidence-Guided Memory and Response Triggering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25621","noteUrl":"/review?paper=review-neurips-2026-150068","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150070","title":"Geometry-Aware Online Scheduling for LLM Serving: From Theoretical Bound to System Practice","authors":["Li Kong","Qi Qi","Yinyu Ye","Zijie Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An untruncated geometric residual law does not prove the finite-cap result","plainExplanation":"A geometric law on positive integers has residual mean \\(1/\\mu\\) at every age. Under a finite output cap \\(B\\), the residual at age \\(B-1\\) is instead one. Thus equation (99)’s moment substitution is not an identity for the capped model. Calling the law an approximation or bounding excluded tail mass alone does not establish the stated exact latency ratio. This is a model-to-proof gap, not a counterexample to the final latency inequality. Reference-model calculations and empirical results may remain valid.","primaryUrl":"https://arxiv.org/html/2606.22327","noteUrl":"/review?paper=review-neurips-2026-150070","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150073","title":"Learning to Solve Compositional Geometry Routing Problems","authors":["Mingfeng Fan","Jianan Zhou","Jiaqi Cheng","Yifeng Zhang","Jie Zhang","Guillaume Sartoretti"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning to Solve Compositional Geometry Routing Problems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18094","noteUrl":"/review?paper=review-neurips-2026-150073","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150076","title":"LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels","authors":["Lucas Maes","Quentin Le Lidec","Damien Scieur","Yann LeCun","Randall Balestriero"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"One tuning dimension alone does not justify logarithmic search","plainExplanation":"Two centered rows \\((-a,a)\\) and \\((a,-a)\\) give unbiased covariance diagonals \\(2a^2\\) and off-diagonals \\(-2a^2\\). The printed off-diagonal sum divided by dimension is \\(-2a^2\\), rather than a nonnegative dependence penalty. Increasing \\(a\\) also satisfies the minimum-variance hinge, so that hinge does not cap this component. The printed regularizer component has the wrong mathematical behavior. Other loss terms can affect optimization, and no implemented divergence is established.","primaryUrl":"https://arxiv.org/html/2603.19312","noteUrl":"/review?paper=review-neurips-2026-150076","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150077","title":"URDF-Anything+: End-to-End Generation for Simulation-Ready Articulated Assets","authors":["Zhuangzhe Wu","Yue Xin","Chengkai Hou","Minghao Chen","Yaoxu Lyu","Jieyu Zhang","Shanghang Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"URDF-Anything+: End-to-End Generation for Simulation-Ready Articulated Assets","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.14010","noteUrl":"/review?paper=review-neurips-2026-150077","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150079","title":"Continuous-Time Distribution Matching for Few-Step Diffusion Distillation","authors":["Tao Liu","Hao Yan","Mengting Chen","Taihang Hu","Zhengrong Yue","Zihao Pan","Jinsong Lan","Xiaoyong Zhu","Ming-Ming Cheng","Bo Zheng","Yaxing Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A fixed terminal anchor gives a discrete atom","plainExplanation":"The fixed terminal student anchor creates an atom absent from the claimed continuous law. Distribution matching also does not imply the proposed pointwise derivative alignment.","primaryUrl":"https://arxiv.org/html/2605.06376","noteUrl":"/review?paper=review-neurips-2026-150079","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150080","title":"Zero-Shot Instruction Following in RL via Structured LTL Representations","authors":["Mathias Jackermeier","Mattia Giuri","Jacques Cloete","Alessandro Abate"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Zero-Shot Instruction Following in RL via Structured LTL Representations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.14344","noteUrl":"/review?paper=review-neurips-2026-150080","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150081","title":"Curvature Beyond Positivity: Greedy Guarantees for Arbitrary Submodular Functions","authors":["Yixin Chen","Alan Kuhnle"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Information-matrix conditioning does not equal the stated curvature","plainExplanation":"The information-matrix condition number does not give the stated function curvature, even for a scalar A-optimal example. Zero-slope pruning also need not improve strictly.","primaryUrl":"https://arxiv.org/html/2605.07902","noteUrl":"/review?paper=review-neurips-2026-150081","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150083","title":"Sampling-Free Privacy Accounting for Matrix Mechanisms under Random Allocation","authors":["Jan Schuchardt","Nikita Kalinin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The product lemma repeats the last reference distribution","plainExplanation":"A destination receives predecessor weight multiplied by \\(e^\\Delta\\), so its log update must be \\(\\operatorname{LSE}(\\mathrm{old},w+\\Delta)\\). Line 28 instead uses \\(\\operatorname{LSE}(w,\\mathrm{old}+\\Delta)\\). At closure, the wrap factor multiplies each state: the correct accumulation is \\(\\operatorname{LSE}(\\mathrm{acc},\\mathrm{state}+\\mathrm{wrap})\\), not an extra standalone wrap summand. The admitted \\(p=1\\) case already exposes both defects with empty prefix and suffix, so no wrap-index ambiguity is needed. The literal pseudocode is not generally exact or an upper bound. Fixing just one line is insufficient; no uninspected repository behavior is inferred.","primaryUrl":"https://arxiv.org/html/2601.21636","noteUrl":"/review?paper=review-neurips-2026-150083","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150087","title":"Beyond Appearance Shifts: Task-Semantic Action Calibration for VLA Models","authors":["Shuaijun Liu","Feiyang You","Chengyu Wu","Shuyang Hao","Chenglong Zhang","Jingyao Cai","Xingwei Chen","Li Sun","Ningxin Su"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Beyond Appearance Shifts: Task-Semantic Action Calibration for VLA Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.23650","noteUrl":"/review?paper=review-neurips-2026-150087","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150088","title":"SPANUQ: Span-Level Uncertainty Quantification for Large Language Model Generation","authors":["Yimeng Zhang","Yingying Zhuang","Ziyi Wang","Yuxuan Lu","Pei Chen","Aman Gupta","Zhe Su","Ming Tan","Zhilin Zhang","Qun Liu","Manikandarajan Ramanathan","Rajashekar Maragoud","Edward Vul","Jing Huang","Dakuo Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"SPANUQ: Span-Level Uncertainty Quantification for Large Language Model Generation","plainExplanation":"One mistaken decisive judgment can change a retained span’s label by one. The printed pooling weights can also give low-validity queries substantial weight.","primaryUrl":"https://arxiv.org/html/2607.05721","noteUrl":"/review?paper=review-neurips-2026-150088","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound"],"classifications":["major"]},{"slug":"review-neurips-2026-150090","title":"Efficient Evaluation of LLM Performance with Statistical Guarantees","authors":["Skyler Wu","Yash Nair","Emmanuel Candes"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Efficient Evaluation of LLM Performance with Statistical Guarantees","plainExplanation":"An observed label can inform later sampling under the stated fixed-bank model. Ignoring that history loses the claimed adaptive oracle optimum.","primaryUrl":"https://arxiv.org/html/2601.20251","noteUrl":"/review?paper=review-neurips-2026-150090","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150092","title":"DPPE: Rethinking Camera-Based Positional Encoding for Scaling Multi-View Transformers","authors":["Shun Kenney","Teppei Suzuki"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"DPPE: Rethinking Camera-Based Positional Encoding for Scaling Multi-View Transformers","plainExplanation":"Pose recovery depends on the collection of probes. Repeating one direction can leave rotations and translations unidentified, while a sufficiently rich token can identify the pose without cross-token aggregation.","primaryUrl":"https://arxiv.org/html/2606.31585","noteUrl":"/review?paper=review-neurips-2026-150092","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150093","title":"DegradeQuery: Counterfactual Tuple Pretraining for Context-Aware PROTAC Degradation Prediction","authors":["Dong Xu","Zhangfan Yang","Jiantao Wu","Zexuan Zhu","Jianqiang Li","Junkai Ji"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DegradeQuery: Counterfactual Tuple Pretraining for Context-Aware PROTAC Degradation Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.10595","noteUrl":"/review?paper=review-neurips-2026-150093","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150095","title":"Extending Pretrained 10-Second ECG Foundation Models to Longer Horizons","authors":["Wei Tang","Jinpei Han","Kangning Cui","Mattia Carletti","Fredrik K. Gustafsson","Shreyank Gowda","Patitapaban Palo","Anshul Thakur","Lei Clifton","Jean-michel Morel","Raymond H Chan","David Clifton","Xiao Gu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Extending Pretrained 10-Second ECG Foundation Models to Longer Horizons","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16975","noteUrl":"/review?paper=review-neurips-2026-150095","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150099","title":"A Mean-Field Framework for Inference-Time Distributional Control of Diffusion Models","authors":["Samuel Howard","Nikolas Nüsken"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"A Mean-Field Framework for Inference-Time Distributional Control of Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.08770","noteUrl":"/review?paper=review-neurips-2026-150099","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150100","title":"Calibeating Prediction-Powered Inference","authors":["Lars van der Laan","Mark van der Laan"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Containing constants does not supply a constant-shift direction","plainExplanation":"A function class can contain every constant without allowing shifts of its nonconstant minimizer. The equality and exact empirical-equivalence claims require separate corrections.","primaryUrl":"https://arxiv.org/html/2604.21260","noteUrl":"/review?paper=review-neurips-2026-150100","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150101","title":"TabClustPFN: A Prior-Fitted Network for Tabular Data Clustering","authors":["Tianqi Zhao","Guanyang Wang","Yan Shuo Tan","Qiong Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TabClustPFN: A Prior-Fitted Network for Tabular Data Clustering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.21656","noteUrl":"/review?paper=review-neurips-2026-150101","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150103","title":"Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution","authors":["Jessie Chen","Zhuo Chen","Archer Wang","Jeff Gore","Bill Freeman","Congyue Deng","Marin Soljacic"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26032","noteUrl":"/review?paper=review-neurips-2026-150103","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150105","title":"Emergent Misalignment as Data-Mediated Transfer","authors":["Baris Askin","Muhammed Ustaomeroglu","Anupam Nayak","Gauri Joshi","Guannan Qu","Carlee Joe-Wong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Emergent Misalignment as Data-Mediated Transfer","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12798v1","noteUrl":"/review?paper=review-neurips-2026-150105","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150106","title":"Stage Light is Sequence$^2$: Multi-Light Control via Imitation Learning","authors":["Zijian Zhao","Dian Jin","Zijing Zhou","Xiaoyu Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The joint minimization loss reverses the PPO actor sign","plainExplanation":"The joint minimization objective uses the PPO actor surrogate with the wrong sign. A separate return display repeats the current reward instead of future rewards.","primaryUrl":"https://arxiv.org/html/2605.03660","noteUrl":"/review?paper=review-neurips-2026-150106","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150109","title":"Counterfactual Online Conformal Prediction Under Adaptive Logging","authors":["Xinyu Qiao","yichen lin","Kaihong Ji","xue wang","Tao Yao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Counterfactual Online Conformal Prediction Under Adaptive Logging","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30811","noteUrl":"/review?paper=review-neurips-2026-150109","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150110","title":"ChainSpace: A Chained-Reasoning Paradigm for Spatial Intelligence","authors":["Xiaohan Zhang","Feng Gu","Xudong Rao","Xuhao Pan","Tao Wei","Pan Zhou","Kun Zhan"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"ChainSpace: A Chained-Reasoning Paradigm for Spatial Intelligence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.15788","noteUrl":"/review?paper=review-neurips-2026-150110","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150111","title":"Does the Question Really Matter? Training-Free Data Selection for Vision-Language SFT","authors":["Peng Sun","Yi Yang","Huawen Shen","Yi Ban","Tianfan Fu","Yanbo Wang","Yuqiang Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Does the Question Really Matter? Training-Free Data Selection for Vision-Language SFT","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.09715","noteUrl":"/review?paper=review-neurips-2026-150111","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150113","title":"Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing","authors":["Yuxiao YE","Haoran He","Fangyuan Kong","Xintao Wang","Pengfei Wan","Kun Gai","Ling Pan"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Equal active rewards make the standardized advantages undefined","plainExplanation":"The active threshold permits a nonempty group of equal rewards. Its centered rewards and population standard deviation are all zero, producing \\(0/0\\). A singleton has the same problem. Thresholding supplies no reward-variation condition; an empty set is not needed for the counterexample. This is a printed normalization boundary gap. Actual runtime failure and empirical impact are not established.","primaryUrl":"https://arxiv.org/html/2606.05950","noteUrl":"/review?paper=review-neurips-2026-150113","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150114","title":"Move on Muon : A Hamiltonian probability gradient flow perspective of Muon optimizer","authors":["Aratrika Mustafi","Soumya Mukherjee","Bharath Sriperumbudur"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Move on Muon : A Hamiltonian probability gradient flow perspective of Muon optimizer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23871","noteUrl":"/review?paper=review-neurips-2026-150114","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150121","title":"Teach-to-Reason: Competition-Guided Reasoning with a Self-Improving Teacher","authors":["Xiao Han","Hao Liu","Zhimin Bao","Jile Jiao","Yue Wang","Hui Guo","Mou X Feng","Yi Xu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Teach-to-Reason: Competition-Guided Reasoning with a Self-Improving Teacher","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.25407","noteUrl":"/review?paper=review-neurips-2026-150121","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150122","title":"Stable and Scalable Probabilistic Numerical Solvers for Stiff and High-Dimensional ODEs","authors":["Nathanael Bosch"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Stable and Scalable Probabilistic Numerical Solvers for Stiff and High-Dimensional ODEs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2606.08203v1","noteUrl":"/review?paper=review-neurips-2026-150122","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150123","title":"Stochastic Dynamic Barrier Perturbed Gradient Methods for Nonconvex Simple Bilevel Optimization","authors":["Mohammad Mahdi Ahmadi","Jincheng Cao","Aryan Mokhtari","Erfan Yazdandoost Hamedani"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Stochastic Dynamic Barrier Perturbed Gradient Methods for Nonconvex Simple Bilevel Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.10957","noteUrl":"/review?paper=review-neurips-2026-150123","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150124","title":"SDFlow: Similarity-Driven Flow Matching for Time Series Generation","authors":["Wei Li","FENG SHIBO","Pengcheng Wu","Xingyu Gao","Min Wu","Peilin Zhao"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Semi-orthogonality does not remove the residual cross term","plainExplanation":"Take \\(V=(1,0)^\\top\\), \\(u^*=0\\), \\(u=-1\\), and residual \\((1,0)\\). Then \\(z=(1,0)\\), \\(z_0=(-1,0)\\), and coordinate mismatch and residual each have squared norm one, but \\(\\|z-z_0\\|^2=4>2\\). Orthogonality of the columns of \\(V\\) does not make an arbitrary approximation residual perpendicular to them. The orthogonal-residual version survives; the bounded-decomposition hypothesis alone does not justify the cross-term cancellation.","primaryUrl":"https://arxiv.org/html/2605.05736","noteUrl":"/review?paper=review-neurips-2026-150124","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150126","title":"PercepCap: Video Captioner with Structured Spatio-Temporal Perception","authors":["Yifan Xu","Wang Zihao","Zhixiao Wang","Jiaming Zhang","Yichun Yang","Desen Meng","Yuanxing Zhang","Pengfei Wan","Limin Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PercepCap: Video Captioner with Structured Spatio-Temporal Perception","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.20389","noteUrl":"/review?paper=review-neurips-2026-150126","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150127","title":"RAVEL: Rare Concept Generation and Editing via Graph-driven Relational Guidance","authors":["Kavana Venkatesh","Yusuf Dalva","Ismini Lourentzou","Pinar Yanardag"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The stated refinement threshold exceeds the score’s entire range","plainExplanation":"For every nonempty feature set, \\(GSI\\leq1\\), so the test \\(GSI\\geq23\\) can never pass. The exact named-v2 HTML states the number 23 in ordinary text, ruling out a lost fraction in extraction. The separate guidance scale has its own range and does not redefine this threshold. The GSI-based criterion is unreachable as printed. Other stopping conditions can still terminate; neither infinite looping nor failed experiments are inferred.","primaryUrl":"https://arxiv.org/html/2412.09614","noteUrl":"/review?paper=review-neurips-2026-150127","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150129","title":"Interpretability-by-Design with Accurate Locally Additive Models and Conditional Feature Effects","authors":["Vasilis Gkolemis","Loukas Kavouras","Dimitrios Kyriakopoulos","Konstantinos Tsopelas","Dimitrios Rontogiannis","Giuseppe Casalicchio","Theodore Dalamagas","Christos Diou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Interpretability-by-Design with Accurate Locally Additive Models and Conditional Feature Effects","plainExplanation":"The centering, identifiability, approximation and monotonicity claims need distinct conditions. Constrained backfitting, unique component allocation and joint-law projection require the stated additional assumptions.","primaryUrl":"https://arxiv.org/html/2602.16503","noteUrl":"/review?paper=review-neurips-2026-150129","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","incorrect_bound","missing_assumption","proof_gap","unsupported_claim"],"classifications":["major"]},{"slug":"review-neurips-2026-150130","title":"Efficient Prediction of Pass@k Scaling in Large Language Models","authors":["Joshua Kazdan","Colin Sullivan","Rylan Schaeffer","Youssef Allouah","Kyssen Yu","Noam Levi","Sanmi Koyejo"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Efficient Prediction of Pass@k Scaling in Large Language Models","plainExplanation":"Likelihood normalization, existence of fitted parameters and the allocation limit each need correction. The proportional allocation rule remains available under positive variances and a growing per-problem trial budget.","primaryUrl":"https://arxiv.org/html/2510.05197","noteUrl":"/review?paper=review-neurips-2026-150130","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","formula_error","missing_assumption","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-150132","title":"When Does Interaction Help? Representational Tradeoffs in Value-Based Imitation Learning","authors":["Luca Viano","Antoine Moulin","Audrey Huang","Volkan Cevher","Philip Amortila","Dylan J Foster"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"When Does Interaction Help? Representational Tradeoffs in Value-Based Imitation Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2607.29617v1","noteUrl":"/review?paper=review-neurips-2026-150132","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150134","title":"PSD: Pushing the Pareto Frontier of Diffusion LLMs via Parallel Speculative Decoding","authors":["Shengyin Sun","Yiming Li","Renxi Liu","Xinqi Li","Hui-Ling Zhen","Weizhe Lin","Chen Chen","Xianzhi Yu","Mingxuan Yuan","Chen Ma"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PSD: Pushing the Pareto Frontier of Diffusion LLMs via Parallel Speculative Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15609","noteUrl":"/review?paper=review-neurips-2026-150134","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150136","title":"Efficient Lookahead Encoding and Abstracted Width for Learning General Policies in Classical Planning","authors":["Michael Aichmüller","Simon Ståhlberg","Martin Funkquist","Hector Geffner"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Efficient Lookahead Encoding and Abstracted Width for Learning General Policies in Classical Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18674","noteUrl":"/review?paper=review-neurips-2026-150136","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150137","title":"Forest-Guided Semantic Transport for Label-Supervised Manifold Alignment","authors":["Adrien Aumon","Myriam Lizotte","Guy Wolf","Kevin Moon","Jake Rhodes"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The printed diversity statistic is identically one","plainExplanation":"The formula called inverse Simpson diversity omits squared label masses. Because unsquared masses sum to one, its displayed value is constant and cannot measure neighborhood diversity. The explicit formula fails the claimed diversity identity, rather than merely using a different scale. The cited metric library may already compute the intended statistic; no code error is established.","primaryUrl":"https://arxiv.org/html/2602.00974","noteUrl":"/review?paper=review-neurips-2026-150137","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150143","title":"Learning Subspace-Preserving Sparse Attention Graphs from Heterogeneous Multiview Data","authors":["Jie Chen","Yuanbiao Gou","Chuanbin Liu","Zhu Wang","Xi Peng"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Learning Subspace-Preserving Sparse Attention Graphs from Heterogeneous Multiview Data","plainExplanation":"A transpose repairs the local reconstruction typo. The substantive gate and support claims require additional conditions: entmax permits zeros without guaranteeing them for every score vector.","primaryUrl":"https://arxiv.org/html/2605.11881","noteUrl":"/review?paper=review-neurips-2026-150143","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-150144","title":"Frontier Language Models Struggle to Copy: Text Can Be Better Viewed in 2D","authors":["Haodong Wen","Yiran Zhang","Yingfa Chen","Kaifeng Lyu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Unequal generated prefixes do not cancel from the attention comparison","plainExplanation":"The copying contradiction omits the contribution of unequal generated-prefix tokens. Correcting its positional typo alone does not complete the proof.","primaryUrl":"https://arxiv.org/html/2607.16072","noteUrl":"/review?paper=review-neurips-2026-150144","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150145","title":"Diffusion Models without Classifier-free Guidance","authors":["Zhicong Tang","Dong Chen","Jianmin Bao","Baining Guo"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Mixture scores require posterior class weights","plainExplanation":"A mixture score averages conditional scores with posterior class weights, not uniform prior weights. The displayed vanilla-scale label is separately inconsistent.","primaryUrl":"https://arxiv.org/html/2502.12154","noteUrl":"/review?paper=review-neurips-2026-150145","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150146","title":"Not Only Where, But When: Temporal Scheduling for RLVR","authors":["Jinghao Zhang","Ruilin Li","Feng Zhao","Jiaqi Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A common decaying multiplier does not relax relative token weights","plainExplanation":"A common decaying multiplier preserves relative token weights and eventually suppresses all updates; it does not return them to ordinary unweighted optimization.","primaryUrl":"https://arxiv.org/html/2605.25381","noteUrl":"/review?paper=review-neurips-2026-150146","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150147","title":"Tight $L_\\infty$ Sample Complexity for Low-Degree and Sparse Boolean Polynomials","authors":["Jasper van Doornmalen","Mathieu Molina","Victor Verdugo","José Verschae"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Tight $L_\\infty$ Sample Complexity for Low-Degree and Sparse Boolean Polynomials","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.17319","noteUrl":"/review?paper=review-neurips-2026-150147","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150150","title":"GLARE: Generating Listening Heads with Appropriate REactions","authors":["Zikai Liao","Yumin Suh","Yi Ouyang","Yi-Lun Lee","Yi-Hsuan Tsai","Zhaozheng Yin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GLARE: Generating Listening Heads with Appropriate REactions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.40317","noteUrl":"/review?paper=review-neurips-2026-150150","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150151","title":"Multi-Part Object Representations via Graph Structures and Co-Part Discovery","authors":["Alex Foo","Wynne Hsu","Mong-Li Lee"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The grouping proof and complexity analysis need more precise output and cost definitions","plainExplanation":"A planar graph can have a vertex adjacent to almost every other vertex. The standard planar edge bound controls the degree sum, not each degree.","primaryUrl":"https://arxiv.org/html/2512.18192","noteUrl":"/review?paper=review-neurips-2026-150151","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150152","title":"From Articulated Kinematics to Routed Visual Control for Action-Conditioned Surgical Video Generation","authors":["Bohan Li","Shuojue Yang","Baorui Peng","Xianda Guo","Erli Zhang","Youqi Tao","Junfeng Duan","Daguang Xu","DOU QI","Xin Jin","Wenjun Zeng","Hao Zhao","Yueming Jin"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A zero kinematic token makes every routing energy zero","plainExplanation":"Each energy is a norm or absolute value of a channel. At \\(K=0\\), all five vanish, and every normalized coefficient is \\(0/0\\). Taking an expectation afterward does not define the per-token quotient. Positive depth on visible tools does not exclude zero-energy background tokens over the full stated routing domain. The manuscript lacks a boundary definition. Actual code guards and runtime NaNs are unknown.","primaryUrl":"https://arxiv.org/html/2605.08712","noteUrl":"/review?paper=review-neurips-2026-150152","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150153","title":"Operads for compositional reasoning in LLMs","authors":["Nathaniel Bottman","Yinhong Liu","Kyle Richardson"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The algebra definition needs coherence, and ordinary grammar ambiguity uses the start color","plainExplanation":"Evaluating a composite question directly can disagree with first evaluating a subquestion and substituting its answer. Arbitrary model outputs therefore need not satisfy the defining algebra axiom.","primaryUrl":"https://arxiv.org/html/2606.13634","noteUrl":"/review?paper=review-neurips-2026-150153","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_ambiguity","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150156","title":"Structure-Guided Masked Autoencoders for Ultra-High Resolution Scientific Image Understanding","authors":["Enzhi Zhang","Du Wu","Rui Zhong","Cong Ma","Isaac Lyngaas","Amir K Ziabari","Xiao Wang","Peng Chen","Tao Luo","Toshio Endo","Fumiyoshi Shoji","Kento Sato","Kentaro Uesugi","Takayuki Nonoyama","Ryuji Kiyama","Masahiro Yoshida","Tezuka Masaru","Tetsuya Ishikawa","Satoshi Matsuoka","Masaharu Munetomo","Mohamed Wahib"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Structure-Guided Masked Autoencoders for Ultra-High Resolution Scientific Image Understanding","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.30682","noteUrl":"/review?paper=review-neurips-2026-150156","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150157","title":"Procedural Refinement by LLM-driven Algorithmic Debugging for ARC-AGI-2","authors":["Yuning Qiu","Lin-Feng Zou","Jiong-Da Wang","Xue-Rong Yuan","Wang-Zhou Dai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Procedural Refinement by LLM-driven Algorithmic Debugging for ARC-AGI-2","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.20334","noteUrl":"/review?paper=review-neurips-2026-150157","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150158","title":"From Cursed to Competitive: Closing the ZO–FO Gap via Input-to-State Stability","authors":["Amir Ali Farzin","Philipp Braun","Iman Shames"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"From Cursed to Competitive: Closing the ZO–FO Gap via Input-to-State Stability","plainExplanation":"A bounded comparison difference can approach a limit without attaining it. At that endpoint the claimed finite input-to-state gain need not exist.","primaryUrl":"https://arxiv.org/html/2604.25372","noteUrl":"/review?paper=review-neurips-2026-150158","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-150161","title":"First-Order Trajectory Matching: Fast Ensemble Predictions of Chaotic, Turbulent, Stochastic Systems","authors":["Shreya Jha","Timo Schorlepp","Nicholas Geissler","Jules Berman","Benjamin Peherstorfer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"First-Order Trajectory Matching: Fast Ensemble Predictions of Chaotic, Turbulent, Stochastic Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11138","noteUrl":"/review?paper=review-neurips-2026-150161","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150164","title":"SAGE: Scalable Automated Robustness Augmentation for LLM Knowledge Evaluation","authors":["Xiaoyuan Li","Yuzhe Wang","Moxin Li","Keqin Bao","Rui Men","Yichang Zhang","Dayiheng Liu","Wenjie Wang","Fuli Feng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SAGE: Scalable Automated Robustness Augmentation for LLM Knowledge Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12022","noteUrl":"/review?paper=review-neurips-2026-150164","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150165","title":"Simulation-Informed Diffusion for Decentralized Multi-robot Motion Planning","authors":["JINHAO LIANG","Sven Koenig","Ferdinando Fioretto"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The arrival-time definition needs a branch for an unreachable finite-horizon goal","plainExplanation":"The set defining \\(a_i\\) can be empty even when a collision-free partial trajectory exists. Its minimum is then undefined, so the subsequent goal-holding constraint has not been specified.","primaryUrl":"https://arxiv.org/html/2605.27697","noteUrl":"/review?paper=review-neurips-2026-150165","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150166","title":"The Geometric Inductive Bias of Grokking: Bypassing Phase Transitions via Architectural Topology","authors":["Alper YILDIRIM"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Activation normalization does not exclude high-parameter-norm representatives","plainExplanation":"For a ReLU block \\(\\operatorname{ReLU}(hA+b)B+d\\), replace \\((A,b,B)\\) by \\((cA,cb,B/c)\\), with \\(c>0\\). Positive homogeneity keeps the whole block output exactly unchanged, and therefore preserves later normalized residuals and predictions, while parameter norm can diverge as \\(c\\to\\infty\\). Internal weights are not constrained by the stated activation normalization. The inference about all parameter-norm representatives fails. The example does not disprove a gap between minimal representatives or empirical acceleration.","primaryUrl":"https://arxiv.org/html/2603.05228","noteUrl":"/review?paper=review-neurips-2026-150166","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150167","title":"SkillMaster: Toward Autonomous Skill Mastery in LLM Agents","authors":["Min Yang","Jinghua Piao","Xu Xia","Yongshun Gong","Yong Li"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"SkillMaster: Toward Autonomous Skill Mastery in LLM Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08693","noteUrl":"/review?paper=review-neurips-2026-150167","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150172","title":"Hierarchical Conformal Classification","authors":["Floris den Hengst","Inès Blin","Majid Mohammadi","Syed I Shah","Taraneh Younesian"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Hierarchical Conformal Classification","plainExplanation":"The implementable predictor needs consistent nonconformity calibration and missed-recall loss. Separate enumeration and adaptive-selection claims need their own corrections; the contextual fixed-universe qualification remains valid.","primaryUrl":"https://arxiv.org/html/2508.13288","noteUrl":"/review?paper=review-neurips-2026-150172","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","complexity_error","definition_error","false_statement","formula_error","invalid_inference","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150174","title":"Geometry-Guided Semantic Reconstruction for 3D Instance Segmentation","authors":["Yuanhao Su","Shaofeng Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Geometry-Guided Semantic Reconstruction for 3D Instance Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16122","noteUrl":"/review?paper=review-neurips-2026-150174","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150176","title":"Diffusion fine-tuning with Rewarded Moment Matching Distillation","authors":["Alexis Jacq","Guillaume Couairon","Quentin Berthet","Valentin De Bortoli","Arnaud Doucet","Romuald Elie"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Diffusion fine-tuning with Rewarded Moment Matching Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30414","noteUrl":"/review?paper=review-neurips-2026-150176","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150178","title":"Perceive-then-Plan: Layout-as-Policy for Monocular 3D Scene Layout Estimation","authors":["Junwei Zhou","Yu-Wing Tai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Perceive-then-Plan: Layout-as-Policy for Monocular 3D Scene Layout Estimation","plainExplanation":"The current formula removes the mismatch identified in the earlier report. Zero modulation now returns the original visual features for every gate value.","primaryUrl":"https://arxiv.org/html/2605.25326","noteUrl":"/review?paper=review-neurips-2026-150178","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150181","title":"The Bayesian Origin of the Probability Weighting Function in Human Representation of Probabilities","authors":["Xin Tong","Thi Thu Uyen Hoang","Xue-Xin Wei","Michael Hahn"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The Bayesian Origin of the Probability Weighting Function in Human Representation of Probabilities","plainExplanation":"The variance coefficient, statistical-penalty domain and observation-model units need separate corrections. The leading variance term remains unchanged.","primaryUrl":"https://arxiv.org/html/2510.04698","noteUrl":"/review?paper=review-neurips-2026-150181","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","invalid_inference","lemma_error","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-150182","title":"AVIS: Adaptive Test-Time Scaling for Vision–Language Models","authors":["Ahmadreza Jeddi","Minh Le","Amirhossein Kazerouni","Hakki Karaimer","Hue Nguyen","Iqbal Mohomed","Michael Brudno","Alex Levinshtein","Kosta Derpanis","Babak Taati","Radek Grzeszczuk"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"AVIS: Adaptive Test-Time Scaling for Vision–Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.11576","noteUrl":"/review?paper=review-neurips-2026-150182","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150187","title":"Permutation-Invariant Spectral Learning via Dyson Diffusion","authors":["Tassilo Schwarz","Cai Dieball","Constantin Kogler","Renaud Lambiotte","Arnaud Doucet","Aljaz Godec","George Deligiannidis"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Permutation-Invariant Spectral Learning via Dyson Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.08535","noteUrl":"/review?paper=review-neurips-2026-150187","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150189","title":"The Scaling Laws of Skills in LLM Agent Systems","authors":["Qiguang Chen","Qiming Yu","Yuhang Gu","Zhuoye Huang","Hanjing Li","Hongyu Liu","Simin Liu","Jinhao Liu","Dengyun Peng","Jiangyi Wang","Zheng Yan","Fanqing Meng","Ethan Qin","Carl Che","Mengkang Hu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Independent scores need not give rare independent overtakes","plainExplanation":"Distractor overtakes share the gold score and can have first-order intersections despite independent raw scores, invalidating the claimed quadratic union remainder.","primaryUrl":"https://arxiv.org/html/2605.16508","noteUrl":"/review?paper=review-neurips-2026-150189","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150197","title":"Beyond MNIST: Limitations of Amplitude Encoding on Quantum Classification","authors":["Xin Wang","Yabo Wang","Rebing Wu"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Beyond MNIST: Limitations of Amplitude Encoding on Quantum Classification","plainExplanation":"","primaryUrl":"https://neurips.cc/virtual/2026/loc/sydney/poster/150197","noteUrl":"/review?paper=review-neurips-2026-150197","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150198","title":"Activation Steering of Video Generation Models via Reduced-Order Linear Optimal Control","authors":["Jihoon Hong","Alice Chan","Qiyue Dai","Julian Skifstad","Glen Chou"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Activation Steering of Video Generation Models via Reduced-Order Linear Optimal Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04775","noteUrl":"/review?paper=review-neurips-2026-150198","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150202","title":"Progressive Memory Transformer: Memory-Aware Attention for Time-Series","authors":["Tord Sture Stangeland","Andreas Köhler","Steffen Maeland","Adín Ramírez Rivera"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Progressive Memory Transformer: Memory-Aware Attention for Time-Series","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31351","noteUrl":"/review?paper=review-neurips-2026-150202","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150203","title":"Weighted Conformal Clustering","authors":["Anirban Nath","YoonHaeng Hur","Genevera Allen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Weighted Conformal Clustering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.00436","noteUrl":"/review?paper=review-neurips-2026-150203","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150204","title":"When do Prophets Profit in Prediction Markets?","authors":["Anri Gu","Nicole Kagan","Alec Sun","Jibang Wu","Haifeng Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"When do Prophets Profit in Prediction Markets?","plainExplanation":"The market proofs contain local errors in strict-convexity reasoning, conjugate-divergence order, binary Kelly stakes, execution-loss accounting and the sign assigned to a score difference. The repairs distinguish the intended profitability results from the particular algebraic steps used to justify them.","primaryUrl":"https://arxiv.org/html/2607.06166","noteUrl":"/review?paper=review-neurips-2026-150204","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150205","title":"Chance-constrained Flow Matching for High-Fidelity Constraint-aware Generation","authors":["JINHAO LIANG","Yixuan Sun","Anirban Samaddar","Sandeep Madireddy","Ferdinando Fioretto"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Chance-constrained Flow Matching for High-Fidelity Constraint-aware Generation","plainExplanation":"Integrating the reaction–diffusion equation gives a state-dependent mass term. Samplewise feasibility and Gaussian chance feasibility also need distinct definitions before their projection results can be compared.","primaryUrl":"https://arxiv.org/html/2509.25157","noteUrl":"/review?paper=review-neurips-2026-150205","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","invalid_inference","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150206","title":"Efficient Image Synthesis with Sphere Latent Encoder","authors":["Tung Do","Thuan H Nguyen","Hao Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Efficient Image Synthesis with Sphere Latent Encoder","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15592","noteUrl":"/review?paper=review-neurips-2026-150206","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150208","title":"CRAFT: Counterfactual-to-Interactive Reinforcement Fine-Tuning for Driving Policies","authors":["Keyu Chen","Nanfei Ye","Yida Wang","Wenchao Sun","Danqi Zhao","Hao Cheng","Sifa Zheng"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"CRAFT: Counterfactual-to-Interactive Reinforcement Fine-Tuning for Driving Policies","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.04470","noteUrl":"/review?paper=review-neurips-2026-150208","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150209","title":"Performance-Driven Policy Optimization for Speculative Decoding with Adaptive Windowing","authors":["Jie Jiang","Xing Sun","RuoTian Chen","Jianan Su","Kaixin Shen"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Performance-Driven Policy Optimization for Speculative Decoding with Adaptive Windowing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14978","noteUrl":"/review?paper=review-neurips-2026-150209","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150211","title":"Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation","authors":["宇航 他","Junfeng Zuo","Tianhao Chu","Si Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.13219","noteUrl":"/review?paper=review-neurips-2026-150211","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150212","title":"BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation","authors":["Ning Li","Zixuan Guo","Yan Xu","Wenbo Fei","Yifan Niu","Chang Luo","Yasheng Wang","Weiwen Liu","Yong Yu","Weinan Zhang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Independent token scores do not factor from token-dependent advantages","plainExplanation":"Let two iid Bernoulli \\(0.3\\) tokens share a scalar logit, so score derivatives \\(g_t=b_t-0.3\\) are independent with variance \\(0.21\\). If the first token is zero, label it hallucinated and the second faithful, giving advantages \\((-1,1)\\); otherwise use zero advantages. Direct enumeration gives second moment \\(\\mathbb E[\\mathbf1_{b_1=0}(g_2-g_1)^2]=0.21\\), while the factored expression is \\(0.21\\,\\mathbb E[2\\mathbf1_{b_1=0}]=0.294\\). Advantages depend on the same tokens as the scores. The displayed factorization fails even with exactly independent token scores. This does not refute every possible sparse-token variance bound.","primaryUrl":"https://arxiv.org/html/2606.15893","noteUrl":"/review?paper=review-neurips-2026-150212","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150217","title":"SliceWorld: A Predictive and Controllable World-State Model for CT Report Generation","authors":["Yuanhe Tian","Yan Song"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"SliceWorld: A Predictive and Controllable World-State Model for CT Report Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24371","noteUrl":"/review?paper=review-neurips-2026-150217","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150220","title":"FRUC: Feedforward Dynamic Scene Reconstruction from Uncalibrated Collaborative Driving Views","authors":["Yihang Tao","Yu Guo","Zhengru Fang","Haonan An","Yuguang Fang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"FRUC: Feedforward Dynamic Scene Reconstruction from Uncalibrated Collaborative Driving Views","plainExplanation":"No issue was found in the checked claims. Paper decomposition remains a real theoretical gap, but record demonstration and claimed universal inequality/repair are wrong.","primaryUrl":"https://arxiv.org/html/2605.29997","noteUrl":"/review?paper=review-neurips-2026-150220","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150223","title":"Matching-Based Few-Shot Semantic Segmentation Models Are Interpretable by Design","authors":["Pasquale De Marinis","Uzay Kaymak","Rogier Brussee","Gennaro Vessio","Giovanna Castellano"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The deletion operator gives the opposite faithfulness objective","plainExplanation":"The purported keep-top metric uses a delete-top operator, reversing the intended faithfulness interpretation.","primaryUrl":"https://arxiv.org/html/2511.18163","noteUrl":"/review?paper=review-neurips-2026-150223","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150224","title":"FLARE: Diffusion for Hybrid Language Model","authors":["Yuchen Zhu","Jing Shi","Chongjian GE","Hao Tan","Yiran Xu","Wanrong Zhu","Jason Kuen","Ryan Rossi","Koustava Goswami","Rajiv Jain","Yongxin Chen","Molei Tao","Jiuxiang Gu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"FLARE: Diffusion for Hybrid Language Model","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.01774","noteUrl":"/review?paper=review-neurips-2026-150224","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150225","title":"Learning to Align Generative Appearance Priors for Fine-grained Image Retrieval","authors":["Shijie Wang","Yadan Luo","Zijian Wang","Xin Yu","Zi Huang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Learning to Align Generative Appearance Priors for Fine-grained Image Retrieval","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09859","noteUrl":"/review?paper=review-neurips-2026-150225","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150226","title":"Search at the Cost of Sampling: Nearly-Instant Latent Space Bayesian Optimization","authors":["Donney Fan","Colin Doumont","Aleksandra Kalisz","Paul Duckworth","Jacob Gardner","Henry Moss","Geoff Pleiss"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Search at the Cost of Sampling: Nearly-Instant Latent Space Bayesian Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.19476","noteUrl":"/review?paper=review-neurips-2026-150226","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150228","title":"Neural Fractional Stochastic Differential Equations","authors":["Yongxin Yang","Timothy Hospedales"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neural Fractional Stochastic Differential Equations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://papers.ssrn.com/sol3/Delivery.cfm/5753542.pdf?abstractid=5753542&mirid=1","noteUrl":"/review?paper=review-neurips-2026-150228","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150229","title":"Batched Stochastic Linear Bandits with 1-Bit Communication Constraints","authors":["Ivan Lau","Daniel McMorrow","Kevin Jamieson","Jonathan Scarlett"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Batched Stochastic Linear Bandits with 1-Bit Communication Constraints","plainExplanation":"The median-of-means radius fails on the stated full failure-probability range because the block-count ceiling can leave one block while the printed radius approaches zero. A radius that retains the ceiling gives a sufficient bound over the full range.","primaryUrl":"https://arxiv.org/html/2605.30976","noteUrl":"/review?paper=review-neurips-2026-150229","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150230","title":"Inductive Deductive Synthesis: Enabling AI to Generate Formally Verified Systems","authors":["Shubham Agarwal","Alexander Krentsel","Shu Liu","Mert Cemri","Audrey Cheng","Rui Meng","Tomas Pfister","Chun-Liang Li","Sylvia Ratnasamy","Aditya Parameswaran","Matei A Zaharia","Ion Stoica","Mohsen Lesani"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The read-your-writes guard permits a stale read from another origin","plainExplanation":"The condition imposes no requirement when the stored value comes from a different client, including the initial origin. It can therefore allow a read that misses a completed session write.","primaryUrl":"https://arxiv.org/html/2605.23109","noteUrl":"/review?paper=review-neurips-2026-150230","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150232","title":"DynamicRad: Content-Adaptive Sparse Attention for Long Video Diffusion","authors":["Yongji Long","Shijun Liang","Jintao Li","Yun Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Masked-softmax renormalization doubles discarded-mass error","plainExplanation":"Renormalizing retained attention doubles its discarded-mass error, a factor missing from the deterministic tail justification.","primaryUrl":"https://arxiv.org/html/2604.20470","noteUrl":"/review?paper=review-neurips-2026-150232","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150233","title":"Stable GFlowNets with Probabilistic Guarantees","authors":["Zengxiang Lei","Ananth Shreekumar","Jonathan Rosenthal","Ruoyu Song","Alvaro A Cardenas","Daniel Fremont","Dongyan Xu","Satish Ukkusuri","Z. Berkay Celik"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The low-ratio admissibility branch needs no extra logarithmic cutoff","plainExplanation":"For \\(\\rho>0,c>0\\), the active low-ratio branch has \\(\\delta/R=(1-e^c\\rho)/(e^c-1)\\). Its numerator is strictly below one, so \\(\\delta/R<1/(e^c-1)\\) automatically. The inactive branch is zero and also satisfies the strict positive bound. The high-ratio branch’s different condition does not justify reciprocating \\(\\rho\\). The scalar certificate search is unnecessarily restricted; the underlying certificate is not shown invalid.","primaryUrl":"https://arxiv.org/html/2605.01729","noteUrl":"/review?paper=review-neurips-2026-150233","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150234","title":"AlphaPROBE: Alpha Mining via principled retriveval and on-graph biased evolution","authors":["Taian Guo","Haiyang Shen","Junyu Luo","Binqi Chen","Hongjun Ding","Jinsheng Huang","Luchen Liu","Yun Ma","Ming Zhang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The admitted search edges can give a negative likelihood","plainExplanation":"The novelty admission rule permits lower-quality children, which can make the quantity called a likelihood negative.","primaryUrl":"https://arxiv.org/html/2602.11917","noteUrl":"/review?paper=review-neurips-2026-150234","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150236","title":"LongSpike: Fractional Order Spiking State Space Models for Efficient Long Sequence Learning","authors":["Xinrui He","Qiyu Kang","Xuhao Li","Zheng-Jun Zha"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The FFT derivation needs the closed-loop input-to-state kernel","plainExplanation":"The convolution is valid for a known driving signal \\(g\\), but here \\(g\\) depends on the unknown state \\(h\\). Aggregating the open-loop trace kernels does not resolve that feedback or give the required convolution from the external input \\(x\\).","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2606.12895","noteUrl":"/review?paper=review-neurips-2026-150236","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150237","title":"SLDR: Defending Against Malicious Fine-tuning via Selective Layers Recovery and Dynamic Routing","authors":["Hui Zhang","Yachao Yuan","Jiayun Wang","Yuanzhuo Li","Hongtao Wang","Yali Yuan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SLDR: Defending Against Malicious Fine-tuning via Selective Layers Recovery and Dynamic Routing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10345","noteUrl":"/review?paper=review-neurips-2026-150237","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150238","title":"Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers","authors":["Luis Medrano-Navarro","Giacomo Baldan","Qiang Liu","Benjamin Holzschuh","Jan Hagnberger","Mathias Niepert","Nils Thuerey"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03363","noteUrl":"/review?paper=review-neurips-2026-150238","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150243","title":"A typed tensor language for federated learning","authors":["Theofilos Mailis","Kalliopi-Christina Despotidou","Konstantinos Filippopolitis","Yannis Foufoulas","Thanasis-Michail Karampatsis","Andreas Ktenidis","Evdokia Mailli","Theodore Papamarkou","Yannis Ioannidis"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A typed tensor language for federated learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21103v2","noteUrl":"/review?paper=review-neurips-2026-150243","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150244","title":"3D-VITAL: Visibility-aware Identity Training from Artificial Liftings of 3D Reconstruction model for Cross-View Object Re-Identification","authors":["William Grolleau","Astrid Sabourin","Guillaume Lapouge","Catherine Achard"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"3D-VITAL: Visibility-aware Identity Training from Artificial Liftings of 3D Reconstruction model for Cross-View Object Re-Identification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16692","noteUrl":"/review?paper=review-neurips-2026-150244","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150246","title":"Sharpness-Aware Hybrid Model Learning for Architecture-Agnostic Parameter Estimation","authors":["Naoya Takeishi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Sharpness-Aware Hybrid Model Learning for Architecture-Agnostic Parameter Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.06837","noteUrl":"/review?paper=review-neurips-2026-150246","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150249","title":"AdaPCLA: Curriculum Prior Internalization For Long-Tailed Longitudinal EHR Generation","authors":["shuai cui","Chen Wenxuan","Wenjie Du","Jian Lou","Dan Li","Wenjie Feng"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"AdaPCLA: Curriculum Prior Internalization For Long-Tailed Longitudinal EHR Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.12645","noteUrl":"/review?paper=review-neurips-2026-150249","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150250","title":"CITE: Anytime Valid Statistical Inference in LLM Self-Consistency","authors":["Hirofumi Ota","Naoto Iwase","Yuki Ichihara","Junpei Komiyama","Masaaki Imaizumi"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"CITE: Anytime Valid Statistical Inference in LLM Self-Consistency","plainExplanation":"The claimed strict inclusion becomes equality on permitted small alphabets. Non-strict inclusion holds generally, while strictness requires enough distinct labels to construct the residual-tail witness.","primaryUrl":"https://arxiv.org/html/2605.05873","noteUrl":"/review?paper=review-neurips-2026-150250","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150257","title":"Data-Driven Soft Labeling Scales DNA Read Classification to Whole-Body Cell-Type Deconvolution","authors":["Dmytro Rizdvanetskyi","Nathan Roos","Pavlo Lutsik"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Data-Driven Soft Labeling Scales DNA Read Classification to Whole-Body Cell-Type Deconvolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.04987","noteUrl":"/review?paper=review-neurips-2026-150257","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150259","title":"PPO in the Fisher-Rao geometry","authors":["Razvan-Andrei Lascu","David Siska","Lukasz Szpruch"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The practical FR-PPO objective needs a discount-dependent rescaling","plainExplanation":"The pairing contributes a factor \\((1-\\gamma)^{-1}\\) to the advantage term that is absent from the sampled loss. The penalty agrees, so this is not an overall scaling of the objective at fixed \\(\\tau\\).","primaryUrl":"https://arxiv.org/html/2506.03757","noteUrl":"/review?paper=review-neurips-2026-150259","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150261","title":"Signed Rectified Flow: Negativity Controlled Generation","authors":["Runlong Liao","Baiyu Su","Lizhang Chen","Qiang Liu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Signed Rectified Flow: Negativity Controlled Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.18516","noteUrl":"/review?paper=review-neurips-2026-150261","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150263","title":"TAPIOCA: Why Task- Aware Pruning Improves OOD model Capability","authors":["Krish Sharma","Omar Naim","Soumadeep SAHA","Vinija Jain","Aman Chadha","Nicholas Asher"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TAPIOCA: Why Task- Aware Pruning Improves OOD model Capability","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.14738","noteUrl":"/review?paper=review-neurips-2026-150263","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150266","title":"Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions","authors":["Zhenting Qi","Ao Qu","Huangyuan Su","Chenyu Wang","Yu Yao","Han Zheng","Kushal Chattopadhyay","Guowei Xu","Zihan Wang","Weirui Ye","Vijay Janapa Reddi","Ju Li","Paul Liang","Hima Lakkaraju","Sham Kakade","Yilun Du"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stationary-payoff proof needs stronger temporal assumptions","plainExplanation":"A stationary bounded process can remain perfectly dependent forever. Its time average need not converge to its unconditional mean, and a negative unconditional increment mean need not give the conditional exponential-moment bound used by the proof.","primaryUrl":"https://arxiv.org/html/2606.02859","noteUrl":"/review?paper=review-neurips-2026-150266","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150267","title":"Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models","authors":["Yusuf Sahin","Ahmed R Saikia","Volkan Cevher","Paolo Favaro"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The perturbation bound omits correlated cross terms","plainExplanation":"The norm square of \\(\\sum_sJ_se_s\\) includes cross moments. Take \\(n\\) positions with \\(A_s=1/n\\), \\(J_s=1/\\sqrt n\\), and identical errors \\(e_s=B\\), where \\(B\\) is Bernoulli \\(1/4\\). The stated sensitivity proxy with \\(L=1\\) holds, but expected squared total is \\(n/4\\), versus diagonal bound \\(1/4\\). Row-normalized attention does not permit Jensen with square-root attention coefficients. This is a local motivational-bound gap; it does not establish poor ADAS performance.","primaryUrl":"https://arxiv.org/html/2606.10829","noteUrl":"/review?paper=review-neurips-2026-150267","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150268","title":"DiffATS: Diffusion in Aligned Tensor Space","authors":["Jinhua Lyu","Tianmin Yu","Brian Kim","Lizhuo Zhou","Chanwook Park","Naichen Shi"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DiffATS: Diffusion in Aligned Tensor Space","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09275","noteUrl":"/review?paper=review-neurips-2026-150268","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150269","title":"$\\gamma$-weakly $\\theta$-up-concavity: A Unified Framework for Non-Convex Optimization Beyond DR-Submodular and OSS Functions","authors":["Mohammad Pedramfar","Vaneet Aggarwal"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"$\\gamma$-weakly $\\theta$-up-concavity: A Unified Framework for Non-Convex Optimization Beyond DR-Submodular and OSS Functions","plainExplanation":"The exact separation procedure never removes a dominated point. A certified frontier representation can supply a different oracle, while the weaker outer approximation remains available.","primaryUrl":"https://arxiv.org/html/2602.13506","noteUrl":"/review?paper=review-neurips-2026-150269","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-150273","title":"Externalized CPDAG Summaries Improve LLM Causal Deduction","authors":["Wentao Sun","João P Nogueira","Dominique Verchere","Mathieu Acher","Alonso Silva"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Externalized CPDAG Summaries Improve LLM Causal Deduction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2609.31071","noteUrl":"/review?paper=review-neurips-2026-150273","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150274","title":"EVOCHAMBER:  Test-Time Co-evolution of Multi-Agent System at Individual, Team, and Population Scales","authors":["Yaolun Zhang","Tianyi Xu","Shengyu Dai","Zhenwen Shao","Qingyun Wu","Huazheng Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EVOCHAMBER:  Test-Time Co-evolution of Multi-Agent System at Individual, Team, and Population Scales","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11136","noteUrl":"/review?paper=review-neurips-2026-150274","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150275","title":"DPIAgent: Divide, Protocol, Isolate for Agentic Reproduction Test Generation","authors":["Hao Liu","Steven Liu","Xin Zhang","Jane Luo","Yu Kang","Jie Wu","Fangkai Yang","Yangyu Huang","Pengfei Gao","Scarlett Li","Yan Lu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"DPIAgent: Divide, Protocol, Isolate for Agentic Reproduction Test Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.23341","noteUrl":"/review?paper=review-neurips-2026-150275","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150280","title":"RubiConv - Efficient Boundary-Respecting Convolutions","authors":["Linda Friso","Annie Marsden","Xinyi Chen","Arushi Gupta","Peter Bartlett","Mark Braverman","Elad Hazan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The twiddle factor must use the padded transform length","plainExplanation":"The packed FFT uses original-length twiddles inside a padded transform, crops frequencies under a full-DFT contract, and lacks the alignment needed to prevent cross-document butterflies.","primaryUrl":"https://arxiv.org/html/2605.08451","noteUrl":"/review?paper=review-neurips-2026-150280","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150281","title":"RESBev: Making BEV Perception More Robust","authors":["Lifeng Zhuo","Kefan Jin","Zhe Liu","Hesheng Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RESBev: Making BEV Perception More Robust","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.09529","noteUrl":"/review?paper=review-neurips-2026-150281","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150284","title":"HiPhy: Hierarchical Alignment for Physically-Plausible Multi-Principle Video Generation","authors":["Tahira Kazimi","Shubhankar Borse","Munawar Hayat","Fatih Porikli","Pinar Yanardag"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"HiPhy: Hierarchical Alignment for Physically-Plausible Multi-Principle Video Generation","plainExplanation":"Standardizing each reward stream does not remove covariance between streams evaluated on the same rollout. Dividing their sum by the square root of the stream count gives constant variance only under an additional covariance condition; subsequent batch normalization is a separate operation.","primaryUrl":"https://arxiv.org/html/2610.02197","noteUrl":"/review?paper=review-neurips-2026-150284","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150287","title":"Adaptive Delayed-Update Cyclic Algorithm for Variational Inequalities","authors":["Yi Wei","Xufeng Cai","Jelena Diakonikolas"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Adaptive Delayed-Update Cyclic Algorithm for Variational Inequalities","plainExplanation":"Equation (4.20) is false: substituting $v_1=u_0+(1-\\beta)(u_1-u_0)$ does not eliminate the factor $1/(1-\\beta)$ multiplying the initial distance. This propagates to the displayed bound in Theorem 4.4. The preceding bound (4.19) permits a direct correction of the constants, so this finding does not establish failure of the convergence rates.","primaryUrl":"https://arxiv.org/html/2603.29128","noteUrl":"/review?paper=review-neurips-2026-150287","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","incorrect_bound","invalid_reduction"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150288","title":"TSQAgent: Rating Time Series Data Quality via Dedicated Agentic Reasoning","authors":["Shunyu Wu","Dan Li","Haozheng Ye","Weibin Feng","Jian Lou","Bo Zhang","Wenjie Feng","Chenjuan Guo","See-Kiong Ng"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Within-task pairwise loss cannot identify cross-task offsets","plainExplanation":"Within-task pairwise losses leave task-specific score offsets unidentified, so the objective alone does not guarantee cross-dataset calibration.","primaryUrl":"https://arxiv.org/html/2606.03629","noteUrl":"/review?paper=review-neurips-2026-150288","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150291","title":"Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating","authors":["Haozhuo Zhang","Qiang Zhang","Jian Tang","Mingzhe Ni","Michele Caprio","Angelo Cangelosi","Wei Pan"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Previous-object gating does not inductively preserve all earlier objects","plainExplanation":"Gating only the previous object leaves older objects unconstrained at later stages, defeating the claimed recursive preservation argument.","primaryUrl":"https://arxiv.org/html/2610.08320","noteUrl":"/review?paper=review-neurips-2026-150291","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150294","title":"The Symmetries of Three-Layer ReLU Networks","authors":["Johanna Marie Gegenfurtner","Moritz Grillo","Guido Montufar"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"The Symmetries of Three-Layer ReLU Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18319","noteUrl":"/review?paper=review-neurips-2026-150294","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150295","title":"FlashBoB: I/O-Efficient Exact Backward-over-Backward for Softmax Attention","authors":["Anthony Givans","Michael Crawshaw","Mingrui Liu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"FlashBoB: I/O-Efficient Exact Backward-over-Backward for Softmax Attention","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.24089","noteUrl":"/review?paper=review-neurips-2026-150295","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150297","title":"Mitigating Over-squashing without Rewiring: A Sheaf Effective Resistance Perspective","authors":["André Ribeiro Guimarães","Germano Barcelos","Amauri Souza","Diego Mesquita","Ana Luiza Tenorio"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mitigating Over-squashing without Rewiring: A Sheaf Effective Resistance Perspective","plainExplanation":"The flow comparison needs feasibility, and the main spectral statement needs the appendix’s interior-spectrum convention. The separate zero-row-sum criticism is resolved by the ordinary combinatorial Laplacian definition.","primaryUrl":"https://arxiv.org/html/2610.04157","noteUrl":"/review?paper=review-neurips-2026-150297","collection":"neurips-2026","severity":"major","issueTypes":["missing_assumption","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-150300","title":"Attention Itself Could Retrieve. RetrieveVGGT: Training-Free Long Context Streaming 3D Reconstruction via Query-Key Similarity Retrieval","authors":["Zichen Zou","Xiaosong Jia","Zuxuan Wu","Yu-Gang Jiang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Strict above-average thinning does not bound the cache","plainExplanation":"When the occupied regions have equal counts, no region is compressed. Repeated balanced visits can therefore grow the retained KV cache without bound.","primaryUrl":"https://arxiv.org/html/2605.09644","noteUrl":"/review?paper=review-neurips-2026-150300","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150301","title":"Sliced Inner Product Gromov–Wasserstein Distances","authors":["Xiaoyun Gong","Gabriel Rioux","Ziv Goldfeld"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Sliced Inner Product Gromov–Wasserstein Distances","plainExplanation":"The distance distinction needs an asymmetric witness. Nonsmooth restriction and residual bounds need their stated sign and domain corrections.","primaryUrl":"https://arxiv.org/html/2605.08546","noteUrl":"/review?paper=review-neurips-2026-150301","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150302","title":"ThinkSafe: Self-Generated Safety Alignment for Reasoning Models","authors":["Seanie Lee","Sangwoo Park","Yumin Choi","Gyeongman Kim","Minki Kang","Jihun Yun","Dongmin Park","Jongho Park","Sung Ju Hwang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Different full policies can have identical safe-set conditionals","plainExplanation":"Safe-set conditioning divides safe probabilities by their total mass. Multiplying all safe coordinates by the same positive factor changes their full masses but leaves the conditional law unchanged. Teacher and reference can therefore have different unsafe mass and identical safe conditionals \\((0.4,0.6)\\), giving zero conditional KL. The earlier conditional criterion and refusal-tilt result can remain valid. The stronger full-policy sufficient condition is false.","primaryUrl":"https://arxiv.org/html/2601.23143","noteUrl":"/review?paper=review-neurips-2026-150302","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150306","title":"SIEVES: Selective Prediction Generalizes through Visual Evidence Scoring","authors":["Hector G. Rodriguez","Marcus Rohrbach"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SIEVES: Selective Prediction Generalizes through Visual Evidence Scoring","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.25855","noteUrl":"/review?paper=review-neurips-2026-150306","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150307","title":"Training-free Spatially Grounded Geometric Shape Encoding","authors":["Yuhang He"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Training-free Spatially Grounded Geometric Shape Encoding","plainExplanation":"Exact pose recovery and isometry require the correct indices, normalization, retained-span rank and physical-scale information. Separately storing geometry cannot recover pose phases carried only by zero geometry coefficients.","primaryUrl":"https://arxiv.org/html/2604.07522","noteUrl":"/review?paper=review-neurips-2026-150307","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","missing_assumption","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-150308","title":"ABC: Any-Subset Autoregression via Non-Markovian Diffusion Bridges in Continuous Time and Space","authors":["Gabe Guo","Thanawat Sornwanee","Lutong Hao","Elon Litman","Stefano Ermon","Jose Blanchet"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The unweighted bridge regression has divergent irreducible endpoint risk","plainExplanation":"For the admitted Brownian bridge from zero to \\(Y\\sim N(0,I_d)\\), \\(X_t=tY+\\sqrt{t(1-t)}Z\\). Then \\(\\operatorname{Var}(X_t)=tI_d\\) and \\(\\operatorname{Var}(Y\\mid X_t)=(1-t)I_d\\). The remaining target \\((Y-X_t)/(1-t)\\) has conditional mean zero and conditional variance \\(I_d/(1-t)\\). Every predictor thus has risk at least \\(d/(1-t)\\), whose time integral diverges. The continuous Gaussian witness proves residual-risk divergence, not merely target-moment divergence. Pointwise oracle regression and implementations with clipping are not refuted; no separate Novikov counterexample is asserted.","primaryUrl":"https://arxiv.org/html/2604.27443","noteUrl":"/review?paper=review-neurips-2026-150308","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150310","title":"Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation","authors":["Chenqi Li","Yu Liu","Shuo Zhang","Timothy Denison","Tingting Zhu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Stage-2 pseudocode omits a gate evaluation already defined in the text","plainExplanation":"The pseudocode omits the evaluation that binds \\(w_k\\) to the current Stage-2 sample. Freezing the gate parameters does not make its outputs constant across samples.","primaryUrl":"https://arxiv.org/html/2603.04478","noteUrl":"/review?paper=review-neurips-2026-150310","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["algorithm_specification"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150313","title":"CASL-VAE: Learning Structured Latent Variables from Unpaired Data for Semi-supervised Clustering and Paired Sample Generation","authors":["Sai Spandana Chintapalli","Pratik Chaudhari","Christos Davatzikos"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"CASL-VAE: Learning Structured Latent Variables from Unpaired Data for Semi-supervised Clustering and Paired Sample Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2607.08254v1","noteUrl":"/review?paper=review-neurips-2026-150313","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150315","title":"Cardinality-Decomposed Loss for Heterogeneous GNNs","authors":["Parul Maheshwari","Amulya Paruchuri","Alireza Sahami Shirazi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"One-to-one labels alone do not make BPR incoherent","plainExplanation":"Binary BPR and softmax cross-entropy coincide for the same logits, so one-to-one cardinality alone cannot make BPR structurally incoherent.","primaryUrl":"https://arxiv.org/html/2607.20737v1","noteUrl":"/review?paper=review-neurips-2026-150315","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150316","title":"A Locally Tokenized Generative Model for Robust Time-Series Watermarking","authors":["Dongbin Kim","Geonwoo Shin","Yujin Choi","Soyeon Park","Jaewook Lee"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"A Locally Tokenized Generative Model for Robust Time-Series Watermarking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.19727","noteUrl":"/review?paper=review-neurips-2026-150316","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150318","title":"Diffusion-DRF: Free, Rich, and Differentiable Reward for Video Diffusion Fine-Tuning","authors":["Yifan Wang","Yanyu Li","Guocheng Qian","Sergey Tulyakov","Yun Fu","Anil Kag"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Diffusion-DRF: Free, Rich, and Differentiable Reward for Video Diffusion Fine-Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.04153","noteUrl":"/review?paper=review-neurips-2026-150318","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150319","title":"Swimba: Switch Mamba Model Scales State Space Models","authors":["Zhixu Du","Krishna Teja Chitty-Venkata","Murali Emani","Venkatram Vishwanath","Hai Li","Yiran Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Swimba: Switch Mamba Model Scales State Space Models","plainExplanation":"With unnormalized top-one routing, the two models already differ after one step. State-average equality does not remove the additional routing factor in the mixed readout.","primaryUrl":"https://arxiv.org/html/2603.06938","noteUrl":"/review?paper=review-neurips-2026-150319","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-150322","title":"Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation","authors":["Bruce C Xu","Adarsh Kumarappan","Mu Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2606.09864v2","noteUrl":"/review?paper=review-neurips-2026-150322","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150323","title":"Atom-level Protein Representation Learning Improves Protein Structure Prediction","authors":["Taewon Kim","Hyosoon Jang","Hyunjin Seo","Seonghwan Seo","Hyeongwoo Kim","Wonho Zhung","Mingyeong Shin","Woo Youn Kim","Sungsoo Ahn"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Atom-level Protein Representation Learning Improves Protein Structure Prediction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22133","noteUrl":"/review?paper=review-neurips-2026-150323","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150326","title":"Beyond Single Expert: Harmonizing Diverse Visual Priors in MLLMs for Spatial Understanding","authors":["XIAO LIN","Xiaohu Huang","Kai Han"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond Single Expert: Harmonizing Diverse Visual Priors in MLLMs for Spatial Understanding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.15054","noteUrl":"/review?paper=review-neurips-2026-150326","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150330","title":"In-Context Optimization for Retrieval-Augmented Generation: A Gradient-Descent Perspective","authors":["mingchen li","Jiatan Huang","Chuxu Zhang","Liang Zhao","Hong Yu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The matching objective needs a zero-update convention","plainExplanation":"The matching loss is undefined at zero updates allowed by its own target construction: the cosine divides by zero and the log norm ratio also lacks a zero-case definition. The printed objective is not defined on all updates admitted by the construction. This does not establish zero initialization, occurrence in training, or an experimental failure.","primaryUrl":"https://arxiv.org/html/2605.26356","noteUrl":"/review?paper=review-neurips-2026-150330","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150331","title":"How Do Language Models Compose Functions?","authors":["Apoorv Khandelwal","Ellie Pavlick"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"How Do Language Models Compose Functions?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.01685","noteUrl":"/review?paper=review-neurips-2026-150331","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150334","title":"Stochastic Matching via Local Sparsification","authors":["Sara Ahmadian","Edith Cohen","Mohammad Roghani"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Stochastic Matching via Local Sparsification","plainExplanation":"No demonstrated false statement; optional clarification is not a numeric contradiction.","primaryUrl":"https://arxiv.org/html/2605.14195","noteUrl":"/review?paper=review-neurips-2026-150334","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150338","title":"DriveWAM: Video Generative Priors Enable Scalable World-Action Modeling for Autonomous Driving","authors":["Chen Shi","Jinrui Xu","Shaoshuai Shi","Kehua Sheng","Bo Zhang","Li Jiang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"DriveWAM: Video Generative Priors Enable Scalable World-Action Modeling for Autonomous Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28544","noteUrl":"/review?paper=review-neurips-2026-150338","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150339","title":"EVIDENT: Routing MLLM Adaptation through Entity-Grounded Visual Evidence for Cross-Domain Video Temporal Grounding","authors":["Geo Ahn","Jiwook Han","Youngrae Kim","Joonseok Lee","Jinwoo Choi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EVIDENT: Routing MLLM Adaptation through Entity-Grounded Visual Evidence for Cross-Domain Video Temporal Grounding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26104","noteUrl":"/review?paper=review-neurips-2026-150339","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150341","title":"Universal Time Series Generation with Neural Controlled Differential Equations","authors":["Torben Berndt","Elyes Farjallah","Leif Seute","RAEID SAQUR","Benjamin Walker","Jan Stühmer"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Universal Time Series Generation with Neural Controlled Differential Equations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.28507","noteUrl":"/review?paper=review-neurips-2026-150341","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150342","title":"Distilling Sequential Computation in Transformer Language Models","authors":["Zixuan Lan","Jessica Yang","Yanhong Li","Karen Livescu","Jiawei Zhou"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Distilling Sequential Computation in Transformer Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.27233","noteUrl":"/review?paper=review-neurips-2026-150342","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150344","title":"Beyond the Golden Teacher: Enhancing Graph Learning through LLM-GNN Co-teaching","authors":["Zhuoyi Peng","Hanlin Gu","Lixin Fan","Yi Yang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Beyond the Golden Teacher: Enhancing Graph Learning through LLM-GNN Co-teaching","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.11583","noteUrl":"/review?paper=review-neurips-2026-150344","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150346","title":"Trust Region Policy Distillation","authors":["Zhengpeng Xie","Li Lyna Zhang","Zeke Xie","Mao Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Trust Region Policy Distillation","plainExplanation":"The terminal return is now defined. The main monotonic-improvement claim still applies a fixed-reward performance identity to a reward that changes with the student policy; the explicit two-token example continues to contradict the claimed bound.","primaryUrl":"https://arxiv.org/html/2607.04751","noteUrl":"/review?paper=review-neurips-2026-150346","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","formula_error"],"classifications":["critical"]},{"slug":"review-neurips-2026-150348","title":"OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation","authors":["Jinghui Lu","Jiayi Guan","Zhijian Huang","Jinlong Li","Guang Li","Lingdong Kong","Yingyan Li","Han Wang","Shaoqing Xu","Yuechen Luo","Fang Li","Chenxu Dang","Junli Wang","Tao Xu","jing wu","Jianhua Wu","Xiaoshuai Hao","Wen Zhang","Tianyi Jiang","Lingfeng Zhang","Lei Zhou","Yingbo Tang","Jie Wang","Yinfeng Gao","Xi zhou Bu","Haochen Tian","Yihang Qiu","Feiyang Jia","Lin Liu","Yigu Ge","Hanbing Li","Jiahong Chen","Zihui Li","Shen Yuannan","Jianwei Cui","Hongwei Xie","Bing Wang","Haiyang Sun","Jingwei Zhao","JiaHui Huang","Pei Liu","zeyu zhu","Yuncheng JIANG","Zibin Guo","Hanchao Leng","Chuhong Gong","Kun Ma","Guang Chen","Kuiyuan Yang","Hangjun Ye","Long Chen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.18486","noteUrl":"/review?paper=review-neurips-2026-150348","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150349","title":"Closing the Approximation Gap in Simulation-free Latent SDEs","authors":["Henry Smith","Brian L Trippe","Scott Linderman"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Closing the Approximation Gap in Simulation-free Latent SDEs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.16138","noteUrl":"/review?paper=review-neurips-2026-150349","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150351","title":"Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs","authors":["Trung Nguyen","Yiming Gao","Fanyi Pu","Kaichen Zhang","Shuo Sun","Ziwei Liu"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13737","noteUrl":"/review?paper=review-neurips-2026-150351","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150353","title":"Near-Optimal Regret in Adversarial Kernel Bandits","authors":["Yu-Jie Zhang","Hao Qiu","Jonathan Scarlett","Kevin Jamieson"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Near-Optimal Regret in Adversarial Kernel Bandits","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26585","noteUrl":"/review?paper=review-neurips-2026-150353","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150354","title":"Understanding the Surprising Generalization Properties of Tabular Foundation Models","authors":["Nour Shaheen","Junwei (Jeremy) Ma","Alex Labach","Frank Hutter","Anthony L Caterini","Valentin Thomas"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Understanding the Surprising Generalization Properties of Tabular Foundation Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.17957","noteUrl":"/review?paper=review-neurips-2026-150354","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150356","title":"Breaking $\\textit{Winner-Takes-All}$: Cooperative Policy Optimization Improves Diverse LLM Reasoning","authors":["Haoxuan Chen","Tianming Liang","Wei-Shi Zheng","Jian-Fang Hu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Breaking $\\textit{Winner-Takes-All}$: Cooperative Policy Optimization Improves Diverse LLM Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11461","noteUrl":"/review?paper=review-neurips-2026-150356","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150357","title":"CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling","authors":["Yifan Zhang","Yutong Dai","Viraj Prabhu","Zhiyuan Hu","Ran Xu","Zeyuan Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The reward range and voting margin need explicit definitions","plainExplanation":"Individual weights bounded by one do not bound their unnormalized sum by one.","primaryUrl":"https://arxiv.org/html/2610.06829","noteUrl":"/review?paper=review-neurips-2026-150357","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_ambiguity","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150358","title":"Training Transformers for KV-Cache Compressibility","authors":["Yoav Gelberg","Yam Eitan","Michael Bronstein","Yarin Gal","Haggai Maron"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The exact antiparallel router initialization has zero gradient","plainExplanation":"At normalized antiparallel vectors, the cosine derivative is zero: \\((I-uu^\\top)w=0\\) when \\(w=-u\\). Consequently every parameter derivative of \\(p=(1-\\cos)/2\\) vanishes at that initialization. A threshold straight-through estimator passes the same zero derivative. Dense and masked outputs also coincide there, so their KL does not provide an escape direction. The exact printed initialization is stationary for the routers. Floating-point effects or an unstated perturbation may change actual behavior; they are not a stated training mechanism.","primaryUrl":"https://arxiv.org/html/2605.05971","noteUrl":"/review?paper=review-neurips-2026-150358","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150363","title":"Prediction Under Imperfect Compression: A Theory of Approximate MDL","authors":["Qian Li","Xinyu Mao","Shang-Hua Teng","Guangxu Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Prediction Under Imperfect Compression: A Theory of Approximate MDL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04834","noteUrl":"/review?paper=review-neurips-2026-150363","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150365","title":"FalconPerception-HD: High Density Perception via Reinforcement Learning","authors":["Sofian Chaybouti","YASSER ABDELAZIZ DAHOU DJILALI","Ngoc D Huynh","REDA ALAMI","Hilde Kuehne"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A per-token likelihood ratio does not reweight the whole group advantage","plainExplanation":"A token-level likelihood ratio leaves other trajectory and group variables under the rollout law, so the claimed exact group-relative gradient unbiasedness fails.","primaryUrl":"https://arxiv.org/html/2608.18881","noteUrl":"/review?paper=review-neurips-2026-150365","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150370","title":"A Private Empirical Defense Against Privacy Audits","authors":["Saloni Modi","Srivi Balaji","Yusong Zhu","Gautam Kamath","Kevin Tian"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The privacy definition needs the nonnegative boundary convention used elsewhere","plainExplanation":"The displayed quantity can be negative even when the least nonnegative DP parameter is zero.","primaryUrl":"https://arxiv.org/html/2608.28934","noteUrl":"/review?paper=review-neurips-2026-150370","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150374","title":"Dynamic Resolution Routing for Efficient Egocentric Grounding","authors":["Huixin Sun","Wangbo Zhao","Fanyue Wei","Qiuxia Lin","Pengzhan Sun","Angela Yao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Dynamic Resolution Routing for Efficient Egocentric Grounding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.01638","noteUrl":"/review?paper=review-neurips-2026-150374","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150376","title":"Phases of Muon: When Muon  Eclipses SignSGD","authors":["Elliot Paquette","Noah Marshall","Lucas Benigni","Guangyuan Wang","Atish Agarwala","Courtney Paquette"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Gaussian sign linearization gives an inverse-square-root scale","plainExplanation":"The intermediate Gaussian sign coefficient has the wrong noise-scale dependence. The next equation uses the corrected coefficient, and the contraction factor follows from the stated adaptive optimal stepsize.","primaryUrl":"https://arxiv.org/html/2605.09552","noteUrl":"/review?paper=review-neurips-2026-150376","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150378","title":"Network Intervention by Polling Strategic Agents","authors":["Chenyu Zhang","Rohit Parasnis","Saurabh Amin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The star-network example uses the wrong singular value","plainExplanation":"For the unit-weight undirected star on \\(n\\) vertices, the maximum degree is \\(n-1\\) and the singular value is \\((n-1)^{3/2}\\). The printed equality is incorrect.","primaryUrl":"https://arxiv.org/pdf/2610.08347v1","noteUrl":"/review?paper=review-neurips-2026-150378","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150379","title":"SeeSE3: The Emergence of 3D Space in Vision Features","authors":["Viorica Patraucean","Leonidas Guibas","Sayna Ebrahimi","Caroline Chen","Ming-Hsuan Yang","Maks Ovsjanikov","Fedor Kitashov"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The adapter results need stronger local regularity and a coordinate-domain restriction","plainExplanation":"Those assumptions do not place \\(K\\) inside one smooth logarithm chart or one ambient flattening chart. The proof’s global use of these coordinates is therefore unjustified.","primaryUrl":"https://arxiv.org/html/2607.14228","noteUrl":"/review?paper=review-neurips-2026-150379","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150380","title":"Unifying Contrastive and Generative Objectives for Visual Understanding and Text-to-Image Generation","authors":["Chao Li","Tianhong Li","Sai V Nuthalapati","Hong-You Chen","Satya Narayan Shukla","Jianpeng Cheng","Yonghuan Yang","Jun Xiao","Xiangjun Fan","Aashu Singh","Dina Katabi","Shlok K Mishra"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Unifying Contrastive and Generative Objectives for Visual Understanding and Text-to-Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.02667","noteUrl":"/review?paper=review-neurips-2026-150380","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150381","title":"SpecHop: Continuous Speculation for Accelerating Multi-Hop Retrieval Agents","authors":["Mehrdad Saberi","Keivan Rezaei","Soheil Feizi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The variance substitution changes direction for positive numerators","plainExplanation":"The variance substitution gives an upper bound only when its numerator is nonpositive, within the stated Gaussian approximation.","primaryUrl":"https://arxiv.org/html/2605.21965","noteUrl":"/review?paper=review-neurips-2026-150381","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150382","title":"Learning to Follow In-Context Watermark Instructions via Self-Distillation","authors":["Yepeng Liu","Tianyi Chen","Xuandong Zhao","Dawn Song","Yuheng Bu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The WIP null calculation must account for shared letter-key membership","plainExplanation":"Distinct tokens that share an initial have the same key-membership indicator. Distinct initials are also dependent because exactly 13 of 26 letters are selected. Independence of the response from the key does not remove either dependence or make its letter frequencies equal the vocabulary frequencies.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2608.29030","noteUrl":"/review?paper=review-neurips-2026-150382","collection":"neurips-2026","severity":"repairable","issueTypes":["calibration_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150384","title":"Alignment Dynamics in LLM Fine-Tuning","authors":["Yuhan Huang","Huanran Chen","Yinpeng Dong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The sequence-level rebound interpretation needs an overlap term","plainExplanation":"The specialization drops a cross term that is generally nonzero for next-token distributions.","primaryUrl":"https://arxiv.org/html/2605.18309","noteUrl":"/review?paper=review-neurips-2026-150384","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","missing_assumption","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150386","title":"Self-Improvement Imitation with Biologically Guided Search for Protein Design Under Oracle Budgets","authors":["Ashima Khanna-Reiter","Dominik G Grimm"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Self-Improvement Imitation with Biologically Guided Search for Protein Design Under Oracle Budgets","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26690","noteUrl":"/review?paper=review-neurips-2026-150386","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150388","title":"From Patches to Trajectories: Privileged Process Supervision for Software-Engineering Agents","authors":["Murong Ma","Tianyu Chen","Yun Lin","Shuai Lu","Qinglin Zhu","Yeyun Gong","Zhiyong Huang","Peng CHENG","Yan Lu","Jin Song Dong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Local segment choices do not minimize length over a prefix-dependent family","plainExplanation":"Local minimality compares segments from the same pool. An alternative trajectory can reach a different later pool, so the proof cannot compare their segment lengths pointwise.","primaryUrl":"https://arxiv.org/html/2605.21996","noteUrl":"/review?paper=review-neurips-2026-150388","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150390","title":"RipplePLM: Structural and Property Decoupling for Protein Mutation Effect Generation","authors":["Liuzhenghao Lv","Yuyang Liu","Yuyang Gao","Li Yuan","Yonghong Tian"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"RipplePLM: Structural and Property Decoupling for Protein Mutation Effect Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01891","noteUrl":"/review?paper=review-neurips-2026-150390","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150392","title":"PG-LRF: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation","authors":["Xiaoda Wang","Minxiao Wang","Kaiqiao Han","Defu Cao","Ching Chang","Yidan Shi","Runze Yan","Xiao Luo","Yan Liu","Xiao Hu","Yizhou Sun","Wei Wang","Carl Yang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Waveform fitting does not separately identify the pulse-arrival phase","plainExplanation":"A compensating phase shift preserves all simulated waveforms while changing the fitted pulse-arrival parameter, leaving its physiological value unidentified without a gauge or independent measurement.","primaryUrl":"https://arxiv.org/html/2605.12541","noteUrl":"/review?paper=review-neurips-2026-150392","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150393","title":"Evidence-RL: Towards Evidence-intensive Visual Reasoning","authors":["Haojie Huang","Xinlei Yu","Chengming Xu","Zhangquan Chen","Cheng Yang","Qingdong He","Yu Yang","Jiangning Zhang","Xiaobin Hu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Evidence-RL: Towards Evidence-intensive Visual Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.08021","noteUrl":"/review?paper=review-neurips-2026-150393","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150396","title":"Hide-and-Seek in Trajectories: Discovering Failure Signals for VLA Runtime Monitoring","authors":["Seongheon Park","Wendi Li","Changdae Oh","Samuel (Min-Hsuan) Yeh","Zsolt Kira","Michael Hagenow","Sharon Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Exchangeability does not prevent a zero conformal scale","plainExplanation":"For constant successful scores \\(s_t=c\\), the first-split mean is \\(c\\) and every selected absolute deviation is zero, so \\(\\sigma(t)=0\\). A constant second-split score gives \\(0/0\\) in equation (8). This degenerate law is exchangeable, and expanding the fallback index set still maximizes only zeros. This is a removable definition gap. Exchangeability alone does not supply positivity, and no actual NaN or empirical failure is asserted.","primaryUrl":"https://arxiv.org/html/2605.30834","noteUrl":"/review?paper=review-neurips-2026-150396","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150398","title":"ICAT: Incident-Case–Grounded Adaptive Testing for Physical-Risk Prediction in Embodied World Models","authors":["Zhenglin Lai","Sirui Huang","Yuteng Li","Changxin Huang","Jianqiang Li","Bingzhe Wu"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"ICAT: Incident-Case–Grounded Adaptive Testing for Physical-Risk Prediction in Embodied World Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.16405","noteUrl":"/review?paper=review-neurips-2026-150398","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150399","title":"On the Burden of Achieving Fairness in Conformal Prediction","authors":["Ziang Gao","Pengqi Liu","Archer Yang","Mouloud Belbahri","Jesse Cresswell","Masoud Asgharian"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"On the Burden of Achieving Fairness in Conformal Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14260","noteUrl":"/review?paper=review-neurips-2026-150399","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150400","title":"Layerwise Progressive Freezing: A Training Scaffold for Depth-Scalable Binary Networks","authors":["Evan G Smith","Bashima Islam"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Layerwise Progressive Freezing: A Training Scaffold for Depth-Scalable Binary Networks","plainExplanation":"Partial mask refresh makes the expected frozen fraction a moving average of past sampling probabilities. It generally lags an increasing schedule, so the current Bernoulli probability is not the current expected fraction.","primaryUrl":"https://arxiv.org/html/2606.27759","noteUrl":"/review?paper=review-neurips-2026-150400","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150402","title":"Co-optimization for Adaptive Conformal Prediction","authors":["Xiaoyi Su","Zhixin Zhou","Rui Luo"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The conformal quantile needs an infinity boundary convention","plainExplanation":"The conformal rule requests a nonexistent \\((n+1)\\)-st score when \\(\\alpha<1/(n+1)\\). The theorem’s full parameter range therefore requires an infinity convention or a restricted domain. This is a boundary specification omission. The standard infinity convention repairs this rank issue; it does not demonstrate undercoverage in the paper’s large-calibration-set experiments.","primaryUrl":"https://arxiv.org/html/2603.01719","noteUrl":"/review?paper=review-neurips-2026-150402","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150404","title":"Selection, Not Fusion: Radar-Modulated State Space Models for Radar-Camera Depth Estimation","authors":["Zhangcheng Hou","TOMOAKI OHTSUKI"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Selection, Not Fusion: Radar-Modulated State Space Models for Radar-Camera Depth Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11840","noteUrl":"/review?paper=review-neurips-2026-150404","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150405","title":"SGD at the Edge of Stability: The Stochastic Sharpness Gap","authors":["Fangshuo Liao","Afroditi Kolomvaki","Anastasios Kyrillidis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A drift-compensated supermartingale need not be nonnegative","plainExplanation":"Subtracting accumulated drift can make the supermartingale negative, invalidating the invoked maximal inequality; a finite-horizon shift supplies a valid repair.","primaryUrl":"https://arxiv.org/html/2604.21016","noteUrl":"/review?paper=review-neurips-2026-150405","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150406","title":"From Failure Taxonomy to Intervention: A Diagnostic Methodology for Industry-Scale AVLM in Video and Live-Streaming Platform Moderation","authors":["Shuchang Ye","Jinqiang Yu","Zhujun Xiao","Yajing Kong","Yist Y Lin","Yang Ma","Jiaxi Liu","Xiaolei XU","Zheng Yu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"From Failure Taxonomy to Intervention: A Diagnostic Methodology for Industry-Scale AVLM in Video and Live-Streaming Platform Moderation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30059","noteUrl":"/review?paper=review-neurips-2026-150406","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150409","title":"Generating from Discrete Distributions Using Diffusions: Insights from Random Constraint Satisfaction Problems","authors":["Alankrita Bhatt","Mukur Gupta","Germain Kolossov","Andrea Montanari"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"BP ordering guarantees depend on the propagation convention","plainExplanation":"The output-law support, local Bayes-attainment argument and BP parameter conventions need separate corrections. The Bayes-attainment issue remains a proof gap. A finite-radius BP control also gives different exact success probabilities for confidence-first and random ordering, while the same finite ensemble has equal success under hard BP closure.","primaryUrl":"https://arxiv.org/html/2603.20589","noteUrl":"/review?paper=review-neurips-2026-150409","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","false_statement","formula_error","missing_assumption","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150410","title":"Data Density Scaling Laws for Image Self-Distillation","authors":["Alvard Barseghyan","Ani V Vanyan","Hakob Tamazyan","Hrant Khachatrian"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Data Density Scaling Laws for Image Self-Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.19013v1","noteUrl":"/review?paper=review-neurips-2026-150410","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150411","title":"Winfree Oscillatory Neural Network","authors":["Jiawen Dai","Yue Song"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Winfree Oscillatory Neural Network","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20922","noteUrl":"/review?paper=review-neurips-2026-150411","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150413","title":"SVoT: State-aware Visualization-of-Thought for Spatial Reasoning via Reinforcement Learning","authors":["Chao Lei","Yanbei Jiang","Markus Hiller","Zhijian Zhou","Xunye Tian","Krista A. Ehinger","Nir Lipovetzky"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"SVoT: State-aware Visualization-of-Thought for Spatial Reasoning via Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11770","noteUrl":"/review?paper=review-neurips-2026-150413","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150414","title":"SetCon: Towards Open-Ended Referring Segmentation via Set-Level Concept Prediction","authors":["Zhixiong Zhang","Yizhuo Li","Shuangrui Ding","Yuhang Zang","Shengyuan Ding","Long Xing","Yibin Wang","Qiaosheng Zhang","Jiaqi Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SetCon: Towards Open-Ended Referring Segmentation via Set-Level Concept Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20110","noteUrl":"/review?paper=review-neurips-2026-150414","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150415","title":"LITMUS: Benchmarking Behavioral Jailbreaks of LLM Agents in Real OS Environments","authors":["Chiyu Zhang","Huiqin Yang","Bendong Jiang","Xiaolei Zhang","Yiran Zhao","Ruyi Chen","Lu Zhou","Xiaogang Xu","Jiafei Wu","Liming Fang","Zhe Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Ambiguous semantic labels must be filtered consistently","plainExplanation":"The error-rate definition admits ambiguous semantic labels that its binary confusion-matrix interpretation omits, so the claimed equality needs consistent filtering.","primaryUrl":"https://arxiv.org/html/2605.10779","noteUrl":"/review?paper=review-neurips-2026-150415","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150418","title":"How Much Is One Recurrence Worth? Iso-Depth Scaling Laws for Looped Language Models","authors":["Kristian Schwethelm","Daniel Rueckert","Georgios Kaissis"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"How Much Is One Recurrence Worth? Iso-Depth Scaling Laws for Looped Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.21106","noteUrl":"/review?paper=review-neurips-2026-150418","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150419","title":"DECEIVE-AFC: Adversarial Claim Attacks against Search-Enabled LLM-based Fact-Checking Systems","authors":["Haoran Ou","Kangjie Chen","Gelei Deng","Hangcheng Liu","Jie Zhang","Tianwei Zhang","Kwok-Yan Lam"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DECEIVE-AFC: Adversarial Claim Attacks against Search-Enabled LLM-based Fact-Checking Systems","plainExplanation":"No issue was found in the checked claims. The alleged reversal assumes that larger values of \\(\\mathrm{Corr}\\) mean greater relevance. The source does not impose that convention, so an incorrect inequality has not been established.","primaryUrl":"https://arxiv.org/html/2602.02569","noteUrl":"/review?paper=review-neurips-2026-150419","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150420","title":"ToMAToMP: Robust and Multi-Parameter Topological Clustering","authors":["Ludo Andrianirina","Mathieu Carrière"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"ToMAToMP: Robust and Multi-Parameter Topological Clustering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14824","noteUrl":"/review?paper=review-neurips-2026-150420","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150423","title":"Minimax Private Estimation of Smooth Optimal-Transport Maps","authors":["Clément Lalanne","David Rodríguez-Vítores","Franck Iutzeler","Jean-Michel Loubes"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Positive-part normalization can have zero mass with positive probability","plainExplanation":"Clipping noisy basis coefficients to a positive part can leave zero total mass with positive probability, making the normalized density undefined without a fallback.","primaryUrl":"https://arxiv.org/html/2606.04683","noteUrl":"/review?paper=review-neurips-2026-150423","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150425","title":"DeepArrhythmia: Segment-Contextualized ECG Arrhythmia Classification via Selective Evidence Acquisition","authors":["Jiahui Li","Ruili Fang","Zishuai Liu","WenZhan Song","Jin Lu","Fei Dou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DeepArrhythmia: Segment-Contextualized ECG Arrhythmia Classification via Selective Evidence Acquisition","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16441","noteUrl":"/review?paper=review-neurips-2026-150425","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150426","title":"AtlasVid: Efficient Ultra-High-Resolution Long Video Generation via Decoupled Global-Local Modeling","authors":["Ziyang Mai","Yuyao Zhang","Yu-Wing Tai"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"AtlasVid: Efficient Ultra-High-Resolution Long Video Generation via Decoupled Global-Local Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16649","noteUrl":"/review?paper=review-neurips-2026-150426","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150427","title":"Fractional State Space Transition for Long Sequence Modeling","authors":["Ivan Kobyzev","Abbas Ghaddar","Ali Nasiri-Sarvi","Lifeng Shang","Yufei CUI"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Fractional State Space Transition for Long Sequence Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36314","noteUrl":"/review?paper=review-neurips-2026-150427","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150429","title":"Escaping the Cognitive Well: Efficient Competition Math with Off-the-Shelf Models","authors":["Xingyu Dang","Rohit Agarwal","Rodrigo Porto","Anirudh Goyal","Liam Fowl","Sanjeev Arora"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Escaping the Cognitive Well: Efficient Competition Math with Off-the-Shelf Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.16793","noteUrl":"/review?paper=review-neurips-2026-150429","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150432","title":"ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models","authors":["Ye Li","Huanan Liu","Kangye Ji","Yuan Meng","Jiajun Fan","Yuansong Wang","Shiyu Qin","Chenglei Wu","Shu-Tao Xia","Zhi Wang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29438","noteUrl":"/review?paper=review-neurips-2026-150432","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150433","title":"Instruction Anchor: Dissecting the Mechanistic Dynamics of Modality Arbitration","authors":["Yu Zhang","Mufan Xu","Xuefeng Bai","Kehai Chen","Pengfei Zhang","Yang Xiang","Min zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Instruction Anchor: Dissecting the Mechanistic Dynamics of Modality Arbitration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.03677","noteUrl":"/review?paper=review-neurips-2026-150433","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150434","title":"Efficient Learning of Truncated Boolean Product Distributions: Influence to the Rescue","authors":["Rohan Chauhan","Ioannis Panageas"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Efficient Learning of Truncated Boolean Product Distributions: Influence to the Rescue","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.22889","noteUrl":"/review?paper=review-neurips-2026-150434","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150437","title":"Equivariant Reinforcement Learning for Clifford Quantum Circuit Synthesis","authors":["Richie Yeung","Aleks Kissinger","Rob Cornish"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Equivariant Reinforcement Learning for Clifford Quantum Circuit Synthesis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10910","noteUrl":"/review?paper=review-neurips-2026-150437","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150438","title":"A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks","authors":["Tomer Keren","Nitay Calderon","Asaf Yehudai","Yotam Perlitz","Michal Shmueli-Scheuer","Roi Reichart"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28556","noteUrl":"/review?paper=review-neurips-2026-150438","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150441","title":"SP$^2$ec: Adaptive Self-Speculative Decoding for Vision-Language Models","authors":["Yuqi Huang","Xingyao Li","Yunlong Hou","Fengzhuo Zhang","Jiachun Pan","Vincent Tan"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SP$^2$ec: Adaptive Self-Speculative Decoding for Vision-Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.18735","noteUrl":"/review?paper=review-neurips-2026-150441","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150442","title":"2-Step Agent: How a Bayesian Decision Maker Learns from AI-Decision Support","authors":["Otto Nyberg","Fausto Carcassi","Davide Tugnoli","Giovanni Cinà"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"2-Step Agent: How a Bayesian Decision Maker Learns from AI-Decision Support","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.21889","noteUrl":"/review?paper=review-neurips-2026-150442","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150443","title":"Does 1/2-Tsallis-INF Also Work Well for Best-Arm Identification?","authors":["Jingxin Zhan","Yuze Han","Zhihua Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Does 1/2-Tsallis-INF Also Work Well for Best-Arm Identification?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.15365","noteUrl":"/review?paper=review-neurips-2026-150443","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150444","title":"Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory","authors":["Haoran Sun","Wenjie Li","Yujie Zhang","Zekai Lin","Fanrui Zhang","Kaitao Chen","Xingqi He","Yichen Li","Mianxin Liu","Lei Liu","Yankai Jiang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.09365","noteUrl":"/review?paper=review-neurips-2026-150444","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150445","title":"Quasi-Linear ICA for Motor Unit Decomposition during Dynamic Contractions","authors":["Alexander K Clarke","Dimitrios Chalatsis","Agnese Grison","Irene Mendez Guerra","Noura Ezaz-Nikpay","Pranav Mamidanna","Shihan Ma","Silvia Muceli","Dario Farina"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Quasi-Linear ICA for Motor Unit Decomposition during Dynamic Contractions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2406.19581","noteUrl":"/review?paper=review-neurips-2026-150445","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150446","title":"Evolutionary Feature Engineering for Structured Data","authors":["Ege Onur Taga","Yilin Zhuang","Muhammed Emrullah Ildiz","Petros Mol","Abhimanyu Das","Karthikeyan Duraisamy","Samet Oymak"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Evolutionary Feature Engineering for Structured Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.01548","noteUrl":"/review?paper=review-neurips-2026-150446","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150447","title":"Calibration without labels in multiple testing","authors":["Adway S Wadekar","Jake A Soloff"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Calibration without labels in multiple testing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19737","noteUrl":"/review?paper=review-neurips-2026-150447","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150450","title":"DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing","authors":["Artyom Mazur","Nina Konovalova","Aibek Alanov"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing","plainExplanation":"No issue was found in the checked claims. The local inconsistency in v1 has been corrected in the current revision. The equation and its explanation now specify the same feature-selection rule.","primaryUrl":"https://arxiv.org/html/2606.15796","noteUrl":"/review?paper=review-neurips-2026-150450","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150451","title":"Detect Anything in Graphic Design: Element-Level Rewards for Autoregressive Detection","authors":["Jiangning Zhu","Bowen Li","Shenyu Qiao","Yima Gu","Zhao Zhang","Yuhui Yuan","Shixia Liu"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Detect Anything in Graphic Design: Element-Level Rewards for Autoregressive Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2609.07072","noteUrl":"/review?paper=review-neurips-2026-150451","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150453","title":"Higher-Order Cell Tracking Transformer","authors":["Jordão Bragantini","Ilan Theodoro","Loic A Royer"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Higher-Order Cell Tracking Transformer","plainExplanation":"No issue was found in the checked claims. Refute the supplied endpoint counterexamples and rationale; this does not certify all nearly-parallel inputs.","primaryUrl":"https://arxiv.org/html/2607.11754","noteUrl":"/review?paper=review-neurips-2026-150453","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150454","title":"WTF?! Simulation-Free Reinforcement Learning with Wasserstein-Tilted Flow Maps","authors":["Abbas Mammadov","Jerry Huang","Justin Lin","Partha Kaushik","Sheel Shah","Kartik Nair","Yee Whye Teh","Nicholas Boffi"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"WTF?! Simulation-Free Reinforcement Learning with Wasserstein-Tilted Flow Maps","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.27033","noteUrl":"/review?paper=review-neurips-2026-150454","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150458","title":"Sparse Attention as a Range Searching Problem: Towards an Inference-Efficient Index for KV Cache","authors":["Mohsen Dehghankar","Abolfazl Asudeh"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The threshold scan has no cluster-exhaustion exit","plainExplanation":"After all \\(K\\) clusters are scanned in every subspace, the live set already contains every key. But \\(U(K)=\\sum_s\\min_i f_{s,i}\\) need not be below \\(\\tau\\). The stated loop then requests cluster rank \\(K+1\\), which does not exist. A finite full live set does not force the numerical threshold condition. The pseudocode needs a termination boundary. The exhaustion repair preserves zero-false-negative filtering; actual implementation is unknown.","primaryUrl":"https://arxiv.org/html/2605.06763","noteUrl":"/review?paper=review-neurips-2026-150458","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150459","title":"CENDRe: Concept Extraction with Natural Domain Representations","authors":["Antonia Holzapfel","Andres Posada Moreno","Sebastian Trimpe"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"CENDRe: Concept Extraction with Natural Domain Representations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.29621","noteUrl":"/review?paper=review-neurips-2026-150459","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150461","title":"On the Token Value Inequality in Efficient Reasoning","authors":["Run Zeng","Hang Hua","Yiyang Liu","Zhiqiang Tao","Ruixiang Tang","Qifan Wang","Cheng Han","Dongfang Liu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"On the Token Value Inequality in Efficient Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.33970","noteUrl":"/review?paper=review-neurips-2026-150461","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150462","title":"Aligning Inductive Bias for Data-Efficient Generalization in State Space Models","authors":["Qiyu Chen","Guozhang Chen"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Aligning Inductive Bias for Data-Efficient Generalization in State Space Models","plainExplanation":"Interference between stable conjugate modes can move a response peak away from the individual modal center. Differentiating one summand cannot locate every extremum of their sum.","primaryUrl":"https://arxiv.org/html/2509.20789","noteUrl":"/review?paper=review-neurips-2026-150462","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-150465","title":"CAB: Accelerating Flow and Diffusion Sampling via Rectification and Corrected Adams-Bashforth","authors":["Anuska Roy","Pravin Ramachandran Nair"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A decreasing grid needs absolute steps in norm and stability bounds","plainExplanation":"For smooth rectified field \\(\\epsilon(\\rho,y)=\\rho\\), with initial condition \\(y(\\rho_0)=0\\), the exact solution is \\((\\rho^2-\\rho_0^2)/2\\). Its first-step difference from Euler is \\(h^2/2\\), for either sign of the increment. AB2 and AB3 integrate this linear field exactly afterward and the linear extrapolation defect is zero, leaving the initial \\(h^2/2\\) error. The global maximum already includes that first error, so it cannot be \\(O(h^3)\\) on uniform grids. The conditional theorem is not refuted. This is its missing startup condition in the described implementation; CAB-2 is unaffected.","primaryUrl":"https://arxiv.org/html/2605.16736","noteUrl":"/review?paper=review-neurips-2026-150465","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150469","title":"Beyond Masked Sparsity: SNACK Enables Truly Sparse Neural Networks on GPU","authors":["Jafar Badour","Elena Mocanu","Maurice Keulen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Beyond Masked Sparsity: SNACK Enables Truly Sparse Neural Networks on GPU","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.04093","noteUrl":"/review?paper=review-neurips-2026-150469","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150471","title":"Beyond ICA: Identifiability by Symmetry Breaking","authors":["Pengzhou Wu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Almost-everywhere affine equality does not imply every-point recovery","plainExplanation":"Affine decoder equality up to null boundaries supports almost-sure recovery, not recovery at every individual image point.","primaryUrl":"https://arxiv.org/html/2607.23182","noteUrl":"/review?paper=review-neurips-2026-150471","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150472","title":"Pruning and Distilling Mixture-of-Experts into Dense Language Models","authors":["Junhyuck Kim","Jihun Yun","Haechan Kim","Gyeongman Kim","Joonghyun Bae","Jaewoong Cho"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Pruning and Distilling Mixture-of-Experts into Dense Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28207","noteUrl":"/review?paper=review-neurips-2026-150472","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150479","title":"Chasing Label Shifters: A Change-Aware Framework for Dynamic Graph Node Classification","authors":["Junghoon Kim","Seungyoon Choi","Hyunsung Kim","Chanyoung Park"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Chasing Label Shifters: A Change-Aware Framework for Dynamic Graph Node Classification","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.17582","noteUrl":"/review?paper=review-neurips-2026-150479","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150480","title":"OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer","authors":["Boyang Wang","Guangyi Xu","Jiahui Zhang","Zhipeng Tang","Zezhou Cheng"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer","plainExplanation":"No issue was found in the checked claims. The objection treats the predictions as an unordered set, but the described targets are a temporally ordered sequence. Converting a regression output to a class label alone would not establish an assignment; the source supplies additional ordering context.","primaryUrl":"https://arxiv.org/html/2604.24762","noteUrl":"/review?paper=review-neurips-2026-150480","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150481","title":"Focal Reward: Balanced Reinforcement Learning under Rubric-Based Rewards","authors":["Yu Huang","Zihua Zhao","Zhaoxin Huan","Wanli Gu","Feng Hong","Xinmu Ge","Lin Yuan","Qiang Hu","Weichang Wu","Xiaolu Zhang","Jun Zhou","Jiangchao Yao"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Focal Reward: Balanced Reinforcement Learning under Rubric-Based Rewards","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26579","noteUrl":"/review?paper=review-neurips-2026-150481","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150482","title":"Distributionally Robust Mixture-of-Experts Training","authors":["Xin Teng","Muxiao Li","Hongyi Wen"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Distributionally Robust Mixture-of-Experts Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07207","noteUrl":"/review?paper=review-neurips-2026-150482","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150484","title":"TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering","authors":["Yuan Sui","Yulin Chen","Yibo Li","Xue Jiang","Yufei He","Yihong Dong","Xiaoxin He","Tianyu Gao","Bryan Hooi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05980","noteUrl":"/review?paper=review-neurips-2026-150484","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150486","title":"Bayesian Preference Learning for Test-Time Steerable Reward Models","authors":["Jiwoo Hong","Shao Tang","Zhipeng Wang"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Bayesian Preference Learning for Test-Time Steerable Reward Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.08819","noteUrl":"/review?paper=review-neurips-2026-150486","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150496","title":"Gaussian Density Splatting Network","authors":["Miao Shang","Yabin Wang","Xiaopeng Hong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Gaussian Density Splatting Network","plainExplanation":"No issue was found in the checked claims. The proposed unconstrained-weight counterexample violates the source’s explicit feasible domain. The fact that MLP heads produce the parameters does not cancel those constraints in the mathematical definition.","primaryUrl":"https://arxiv.org/html/2610.10396","noteUrl":"/review?paper=review-neurips-2026-150496","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150497","title":"A Near-optimal SQ Lower Bound for Smoothed Agnostic Learning of Boolean Halfspaces","authors":["Tim Sinen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Near-optimal SQ Lower Bound for Smoothed Agnostic Learning of Boolean Halfspaces","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.02350","noteUrl":"/review?paper=review-neurips-2026-150497","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150498","title":"Learning Pareto Stationary Fronts via Single-Pass Backpropagation","authors":["Elina Rojin Celik","Marcos M. Raimundo","Isabel Valera"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Expected feature energy and Dirichlet notation need distinct corrections","plainExplanation":"Geometric-mean growth does not give arithmetic expected feature energy. The valid eigenvalue range bound also does not make a random step eigenvalue equal to its expectation. The two-task Dirichlet variance formula needs an explicit symmetry and scalar-concentration convention.","primaryUrl":"https://arxiv.org/html/2610.06397","noteUrl":"/review?paper=review-neurips-2026-150498","collection":"neurips-2026","severity":"major","issueTypes":["definition_ambiguity","invalid_inference","lemma_error","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-150501","title":"Untrusted Content Masking for Web Agents with Security Guarantees","authors":["Kristina Nikolić","Egor Zverev","Javier Rando","Matthew Jagielski","Edoardo Debenedetti","Florian Tramer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Untrusted Content Masking for Web Agents with Security Guarantees","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.05277","noteUrl":"/review?paper=review-neurips-2026-150501","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150502","title":"Fast KVzip: Efficient and Accurate LLM Inference with Gated KV Eviction","authors":["Jang-Hyun Kim","Dongyoon Han","Sangdoo Yun"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Fast KVzip: Efficient and Accurate LLM Inference with Gated KV Eviction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.17668","noteUrl":"/review?paper=review-neurips-2026-150502","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150503","title":"Learning through Internalization","authors":["Nirmit Joshi","Marko Medvedev","Nikolaos Tsilivis","Julia Kempe","Nati Srebro"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The parity argument needs an all-row Phase-1 update and an affine-region proof","plainExplanation":"The literal update leaves every later output row unchanged, even though their terms appear in the Phase-1 loss.","primaryUrl":"https://arxiv.org/html/2606.20937","noteUrl":"/review?paper=review-neurips-2026-150503","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150504","title":"When Attention Collapses: Residual Evidence Modeling for Compositional Inference","authors":["Niklas Houba"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"When Attention Collapses: Residual Evidence Modeling for Compositional Inference","plainExplanation":"The clipped evidence recurrence reaches its floor under persistent attention, but the proof applies a logarithmic bound after clipping that need not hold. Positive attention also does not force distinct attention distributions or effective inputs.","primaryUrl":"https://arxiv.org/html/2605.02323","noteUrl":"/review?paper=review-neurips-2026-150504","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150506","title":"Intent Factored Generation: Unleashing the Diversity in Your Language Model","authors":["Eltayeb Ahmed","Uljad Berdica","Martha Elliott","Danijela Horak","Jakob Foerster"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The joint training objective has a response-loss subscript typo","plainExplanation":"The second term uses the example-index symbol in place of the defined response-loss label.","primaryUrl":"https://arxiv.org/html/2506.09659","noteUrl":"/review?paper=review-neurips-2026-150506","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150507","title":"DC-ViT: Modulating Spatial and Channel Interactions for Multi-Channel Images","authors":["Umar Marikkar","Syed S Husain","Muhammad Awais","Sara Atito"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The caption calls quadratic attention growth exponential","plainExplanation":"The caption misidentifies quadratic attention growth as exponential. The stated pairwise attention cost is proportional to the square of channel count with other dimensions fixed. The correction concerns the asymptotic complexity description. The later analysis and empirical efficiency comparison are not refuted.","primaryUrl":"https://arxiv.org/html/2603.14281","noteUrl":"/review?paper=review-neurips-2026-150507","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150508","title":"Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift","authors":["Sungjun Lim","Heedong Kim","Andrew Lee","Kyungwoo Song"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift","plainExplanation":"A diagonal eigenvalue crossing violates the global projector bound and its cross-block proof step. Approximate OOD subspace recovery also needs an estimation-error term before it can guarantee the full improvement margin of the exact population projector.","primaryUrl":"https://arxiv.org/html/2605.21849","noteUrl":"/review?paper=review-neurips-2026-150508","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150509","title":"Independent Learning of Nash Equilibria in Partially Observable Markov Potential Games with Decoupled Dynamics","authors":["Philip Jordan","Maryam Kamgarpour"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Independent Learning of Nash Equilibria in Partially Observable Markov Potential Games with Decoupled Dynamics","plainExplanation":"The concentrability assumption needs the harmonic-mean ratio used by the telescoping argument. The displayed policy update also needs a probability distribution over tied maximizers, and the filter-contraction lemma needs matching absolute times or stronger common-map assumptions.","primaryUrl":"https://arxiv.org/html/2605.06377","noteUrl":"/review?paper=review-neurips-2026-150509","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","definition_error","lemma_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150512","title":"Semiparametric Efficient Tests for Interpretable Distributional Treatment Effects","authors":["Houssam Zenati","Arthur Gretton"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Semiparametric Efficient Tests for Interpretable Distributional Treatment Effects","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08034","noteUrl":"/review?paper=review-neurips-2026-150512","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150514","title":"A Single Deep Preference-Conditioned Policy for Learning Pareto Coverage Sets","authors":["Akihiro Kubo","Kosuke Nakanishi","Shin Ishii"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The actor pseudocode omits current-policy entropy regularization","plainExplanation":"The actor algorithm omits current entropy regularization, and its soft target does not discount the complete next-state value under the declared reference convention.","primaryUrl":"https://arxiv.org/html/2605.08946","noteUrl":"/review?paper=review-neurips-2026-150514","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150516","title":"Neural Slack Variables for Shape Constraints","authors":["Ruben Wiedemann","Antoine Jacquier","Lukas Gonon"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neural Slack Variables for Shape Constraints","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.13803","noteUrl":"/review?paper=review-neurips-2026-150516","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150518","title":"Visual Instruction Tuning Aligns Modalities through Abstraction","authors":["Luis Palacios","Lorenzo Basile","Diego Doimo","Alberto Cazzaniga"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Visual Instruction Tuning Aligns Modalities through Abstraction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03871","noteUrl":"/review?paper=review-neurips-2026-150518","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150523","title":"Decentralized Aggregation of LLM Predictions via Wagering Mechanisms","authors":["Yuhong Luo","David Pennock","Xintong Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Decentralized Aggregation of LLM Predictions via Wagering Mechanisms","plainExplanation":"The mixture-regret, best-response, zero-wager and pooling guarantees need different conditions. The valid average-expert comparator and external-Bayesian property remain separate from the false claims.","primaryUrl":"https://arxiv.org/html/2607.04389","noteUrl":"/review?paper=review-neurips-2026-150523","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","false_statement","incorrect_bound","invalid_inference","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150529","title":"Certification from Examples is Hard for Circuits and Transformers under Minimal Overparametrization","authors":["Artur Back de Luca","Kimon Fountoulakis"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Certification from Examples is Hard for Circuits and Transformers under Minimal Overparametrization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22964","noteUrl":"/review?paper=review-neurips-2026-150529","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150532","title":"Open Vocabulary Domain Unlearning","authors":["Sumanth V Udupa","Mehrtash Harandi","Yadan Luo","Mahsa Baktashmotlagh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Open Vocabulary Domain Unlearning","plainExplanation":"Large forget-versus-retain loss sensitivity does not imply semantic independence. The mask can select a parameter that also changes an unseen retain-class decision.","primaryUrl":"https://arxiv.org/html/2609.31356","noteUrl":"/review?paper=review-neurips-2026-150532","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-150533","title":"FlashEvolve: Accelerating Agent Self-Evolution with Asynchronous Stage Orchestration","authors":["Zhengding Hu","Mingge Lu","Zhen Wang","Jixuan Ruan","Chang Chen","Zaifeng Pan","Yue Guan","Ruiyi Wang","Zhongkai Yu","Chao Zhang","Yufei Ding"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FlashEvolve: Accelerating Agent Self-Evolution with Asynchronous Stage Orchestration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08520v1","noteUrl":"/review?paper=review-neurips-2026-150533","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150534","title":"PLATO: Pointer Learner for Agent and Task Openness","authors":["Alireza Saleh Abadi","Leen-Kiat Soh","Daniel A Redder","Adam Eck","Prashant Doshi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The uniqueness proof’s final display asserts both distinct outputs","plainExplanation":"Uniqueness requires that two distinct outputs at one state cannot both lie in the graph. The displayed conclusion instead puts both outputs in the graph for every pair, asserting the opposite when they are distinct. The surrounding deterministic-forward argument gives the intended single-valued reasoning. This is a repeated quantifier/negation slip, not a counterexample to architecture well-definedness or intended uniqueness.","primaryUrl":"https://arxiv.org/html/2607.25082","noteUrl":"/review?paper=review-neurips-2026-150534","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150535","title":"Scalable Distributed Stochastic Optimization via Bidirectional Compression: Beyond Pessimistic Limits","authors":["Grigory Begunov","Alexander Tyurin"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Scalable Distributed Stochastic Optimization via Bidirectional Compression: Beyond Pessimistic Limits","plainExplanation":"No issue was found in the checked claims. The former worker-residual term on the right has been corrected to the global increment.","primaryUrl":"https://arxiv.org/html/2605.07795","noteUrl":"/review?paper=review-neurips-2026-150535","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150536","title":"AdaptNC: Adaptive Nonconformity Scores for Conformal Prediction under Distribution Shift","authors":["Renukanandan Tumu","Aditya Singh","Rahul Mangharam"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"AdaptNC: Adaptive Nonconformity Scores for Conformal Prediction under Distribution Shift","plainExplanation":"The replay comparison controls changes caused by a new score but omits the difference between the old-score replay and the actual chronological state. That reset residual needs a vanishing-average bound before the chronological coverage argument applies.","primaryUrl":"https://arxiv.org/html/2602.01629","noteUrl":"/review?paper=review-neurips-2026-150536","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150537","title":"CAM: Question Answering on Entity-Centric Videos with Continuous Extraction and Adaptive Querying","authors":["Yizhou Tian","Zizhe Chen","Shiyuan Deng","Garry YANG","Zijie Dai","Luohao Pan","Hao Lin","Peiqi Yin","Xiao Yan","James Cheng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CAM: Question Answering on Entity-Centric Videos with Continuous Extraction and Adaptive Querying","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.06504","noteUrl":"/review?paper=review-neurips-2026-150537","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150543","title":"Conservative Continuous-Time Treatment Optimization","authors":["Nora Schneider","Georg Manten","Niki Kilbertus"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Conservativeness requires a calibrated discrepancy coefficient","plainExplanation":"A positive discrepancy penalty certifies the claimed cost upper bound only when its coefficient covers the cost-class scale.","primaryUrl":"https://arxiv.org/html/2603.16789","noteUrl":"/review?paper=review-neurips-2026-150543","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150546","title":"LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL","authors":["Yujin Kim","Namgyu Ho","Sangmin Hwang","Joonkee Kim","Yongjin Yang","Sangmin Bae","Seungone Kim","Jaehun Jung","Se-Young Yun","Hwanjun Song"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.04412","noteUrl":"/review?paper=review-neurips-2026-150546","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150548","title":"Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It","authors":["Sebastian Diaz","Polina Golland","Elfar Adalsteinsson","Neel Dey"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.02564","noteUrl":"/review?paper=review-neurips-2026-150548","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150549","title":"Bridging Risk Approximation Gaps in Model Predictive Task Sampling via In-Context Modeling","authors":["Jiarong Wen","Qi Tao","Zhang Kaiyu","Yun Qu","Lizhou Cai","Heming Zou","Yuhang Jiang","Qi Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The ELBO-trained predictor needs an excess reconstruction-risk term","plainExplanation":"The Gaussian likelihood does not remove the KL penalty or make the fitted predictive mean a squared-error minimizer. Equation(11) uses that minimization property without establishing it. An infimum over the class cannot absorb the fitted model’s excess risk.","primaryUrl":"https://arxiv.org/html/2610.14913","noteUrl":"/review?paper=review-neurips-2026-150549","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150551","title":"QUTCC: Quantile Uncertainty Training and Conformal Calibration for Imaging Inverse Problems","authors":["Cassandra T Ye","Shamus Li","Tyler King","Kristina Monakhova"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The finite-sample correction must be applied to each half-budget","plainExplanation":"These operations do not commute. Algorithm1 permits \\(1/(2N_c)\\) more empirical miscoverage at each endpoint than the target stated in the prose.","primaryUrl":"https://arxiv.org/html/2507.14760","noteUrl":"/review?paper=review-neurips-2026-150551","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150552","title":"AdvJudge-Zero: Binary Decision Flips in LLM-as-a-Judge via Adversarial Control Tokens","authors":["Tung-Ling Li","Yuhao Wu","Hongliang Liu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"AdvJudge-Zero: Binary Decision Flips in LLM-as-a-Judge via Adversarial Control Tokens","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.17375","noteUrl":"/review?paper=review-neurips-2026-150552","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150553","title":"P$^{3}$: Joint Program-and-Proof Planning\\\\ for Verified Code Generation","authors":["Zenan Li","Ziran Yang","Peiyang Song","Zhaoyu Li","Kaiyu Yang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"P$^{3}$: Joint Program-and-Proof Planning\\\\ for Verified Code Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.09277","noteUrl":"/review?paper=review-neurips-2026-150553","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150558","title":"Ask, Answer, and Detect: Role-Playing LLMs for Personality Detection with Question-Conditioned Mixture-of-Experts","authors":["Yifan Lyu","Liang Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The fusion equation needs a common representation space","plainExplanation":"The gate can consume a concatenation of vectors with different dimensions, but its output does not itself map the two summands into the same vector space.","primaryUrl":"https://arxiv.org/html/2512.08814","noteUrl":"/review?paper=review-neurips-2026-150558","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_ambiguity"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150562","title":"LatentOmni: Rethinking Omni-Modal Understanding via Unified Audio-Visual Latent Reasoning","authors":["Yifan Dai","zhenhua wu","Bohan Zeng","Daili Hua","Jialing Liu","Bozhou Li","Yuran Wang","Chengzhuo Tong","Hao Liang","Xiaochen Ma","Junbo Niu","Tianyu Guo","Yang Shi","Yue Ding","Yiyan Ji","Bingyin Mei","Yushuo Guan","Yuanxing Zhang","Pengfei Wan","Fangcheng Fu","Wentao Zhang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"LatentOmni: Rethinking Omni-Modal Understanding via Unified Audio-Visual Latent Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22012","noteUrl":"/review?paper=review-neurips-2026-150562","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150563","title":"Automated Causal Effect Estimation through Self-Evolving AI","authors":["Can Wang","Hongyu Zhao","Yiqun Chen"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Automated Causal Effect Estimation through Self-Evolving AI","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.04274v1","noteUrl":"/review?paper=review-neurips-2026-150563","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150565","title":"Improving Consistency in Retrieval Augmented Systems  With Group Similarity Rewards","authors":["Faisal Hamman","Chenyang Zhu","Anoop Kumar","Xujun Peng","Sanghamitra Dutta","Daben Liu","Alfy Samuel"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Improving Consistency in Retrieval Augmented Systems  With Group Similarity Rewards","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.04392","noteUrl":"/review?paper=review-neurips-2026-150565","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150567","title":"Bayesian Optimization on Function Spaces via Sparse RKHS Manifolds","authors":["Davide Sartor","Meghan E Huber","Donghyun Kim","Nathan Wycoff"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Bayesian Optimization on Function Spaces via Sparse RKHS Manifolds","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07417","noteUrl":"/review?paper=review-neurips-2026-150567","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150568","title":"Parameter symmetries determine representational geometry in overparameterized nonlinear networks","authors":["Marvin Theiss","Lukas Braun","Andrew Saxe","Erin Grant"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Parameter symmetries determine representational geometry in overparameterized nonlinear networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39078","noteUrl":"/review?paper=review-neurips-2026-150568","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150569","title":"Latent Q-Barrier Shielding for Safe In-Context Reinforcement Learning","authors":["Minjae Kwon","Amir Moeini","Shangtong Zhang","Lu Feng"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Latent Q-Barrier Shielding for Safe In-Context Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25267","noteUrl":"/review?paper=review-neurips-2026-150569","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150572","title":"DySurface: Consistent 4D Surface Reconstruction via Bridging Explicit Gaussians and Implicit Functions","authors":["Minje Kim","Younghyun Noh","Jaesoon Kim","Tae-Kyun Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The cycle-loss display swaps its arrow labels and omits the paired coordinates","plainExplanation":"The arrow labels are inconsistent with the preceding definitions. The omitted arguments should also identify corresponding points, so that the expression is clearly a composition residual.","primaryUrl":"https://arxiv.org/html/2605.10360","noteUrl":"/review?paper=review-neurips-2026-150572","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150573","title":"OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting","authors":["Sanah Suri","Kieran Ringel","Maike Sonnewald"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A spatial difference does not rule out a signed target copy","plainExplanation":"A free concept can be a signed copy of the target while its unweighted spatial difference remains structured, so that diagnostic does not rule out leakage.","primaryUrl":"https://arxiv.org/html/2605.12639","noteUrl":"/review?paper=review-neurips-2026-150573","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150575","title":"Value Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models","authors":["Seong Hah Cho","Junyi Li","Anna Leshinskaya"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Value Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.19101","noteUrl":"/review?paper=review-neurips-2026-150575","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150576","title":"Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory","authors":["Jingwei Sun","Jianing Zhu","Jiangchao Yao","Tongliang Liu","Bo Han"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19952","noteUrl":"/review?paper=review-neurips-2026-150576","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150579","title":"TIDES: Implicit Time-Awareness in Selective State Space Models","authors":["Taylan Soydan","Miguel Bessa","Dirk Mohr","Rui Barreira"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TIDES: Implicit Time-Awareness in Selective State Space Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09742","noteUrl":"/review?paper=review-neurips-2026-150579","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150580","title":"LoopRPT: Reinforcement Pre-Training for Looped Language Models","authors":["Guo Tang","Shixin Jiang","Heng Chang","Zihan Zhang","Nuo Chen","Yuhan Li","HuiMing Fan","Jia Li","Ming Liu","Bing Qin"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"LoopRPT: Reinforcement Pre-Training for Looped Language Models","plainExplanation":"The lag weights, entropy inequality, regularizer sign and exit-threshold boundary have separate explicit corrections. The correctly derived EMA closed form and restricted drift arguments remain available.","primaryUrl":"https://arxiv.org/html/2603.19714","noteUrl":"/review?paper=review-neurips-2026-150580","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","lemma_error","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-150581","title":"Scalable Maximum Entropy Reinforcement Learning for Diffusion Policies via Adjoint Matching","authors":["Serge Thilges","Onur Celik","Denis Blessing","Emiliyan Gospodinov","Gerhard Neumann"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The general adjoint target has a sign error; exact policy iteration has a valid qualification","plainExplanation":"The target must contain the negative gradient of the terminal cost. The printed positive reference-density term gives a different control even at the optimal path law.","primaryUrl":"https://arxiv.org/html/2606.22630","noteUrl":"/review?paper=review-neurips-2026-150581","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150582","title":"ROTATE: Regret-driven Open-ended Training for Ad Hoc Teamwork","authors":["Caroline Wang","Muhammad Arrasy Rahman","Benjamin Nativi","Johnny Liu","Jiaxun Cui","Yoonchang Sung","Peter Stone"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Current-ego regret maximization does not prove fixed-observer strict novelty","plainExplanation":"The novelty definition fixes observer \\(t\\) and requires a future artifact with strictly larger regret than the current one. In a one-step two-action coordination game, deterministic teammate \\(T\\) can already have regret \\(1\\) against fixed observer \\(H\\), the global maximum. No future teammate can have strictly greater regret. Maximization against a different, updated ego cannot overcome that fixed-observer bound. The strict novelty proof fails independently of a particular cycling trace. Empirical diversity, competence and increasing evaluation returns are not disproved.","primaryUrl":"https://arxiv.org/html/2505.23686","noteUrl":"/review?paper=review-neurips-2026-150582","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150587","title":"Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing","authors":["Xiang Li","Xueheng Li","Yu Wang","Xuanhua He","Zhangchi Hu","Chengjun Xie"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.15356","noteUrl":"/review?paper=review-neurips-2026-150587","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150599","title":"BioBlobs: Unsupervised Discovery of Functional Substructures for Protein Function Prediction","authors":["Xin(Allen) Wang","Kaiwen Shi","Carlos Oliver"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"BioBlobs: Unsupervised Discovery of Functional Substructures for Protein Function Prediction","plainExplanation":"Distinct seeds and independent sigmoid memberships do not make the resulting residue blobs disjoint. Overlapping radius neighborhoods admit shared memberships; disjoint hard outputs need an explicit exclusive assignment or separated candidate neighborhoods.","primaryUrl":"https://arxiv.org/html/2510.01632","noteUrl":"/review?paper=review-neurips-2026-150599","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150600","title":"Length Generalization for Transformers via Compression","authors":["Georg Zetzsche","Hongjian Jiang","Andy J Yang","Pascal Bergsträßer","Marco Sälzer","David Chiang","Anthony Widjaja Lin"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Length Generalization for Transformers via Compression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.08851","noteUrl":"/review?paper=review-neurips-2026-150600","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150601","title":"Closing the Indexing-Decoding Gap in Multimodal Generative Retrieval via Prefix Retention Optimization","authors":["Yufei Chen","Zihan Wang","Yubao Tang","Yukun Zhao","Maarten Rijke","Zhaochun Ren"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Closing the Indexing-Decoding Gap in Multimodal Generative Retrieval via Prefix Retention Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.09241","noteUrl":"/review?paper=review-neurips-2026-150601","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150603","title":"Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training","authors":["Anshu Raina","Peyman Razaghi","Yuankai Chen","Cheng Yao","Yao Fu","Vidushi Goyal","Devang Patel","Wen Xie","Zhenyu Gu","Emad Barsoum"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.19423v1","noteUrl":"/review?paper=review-neurips-2026-150603","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150604","title":"Faithful Embeddings of Irregular and Asynchronous Data for Online Log-NCDEs","authors":["Benjamin Walker","Alexandre Bloch","Lingyi Yang","Sam Morley","Terry Lyons"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Faithful Embeddings of Irregular and Asynchronous Data for Online Log-NCDEs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30213","noteUrl":"/review?paper=review-neurips-2026-150604","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150607","title":"Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation","authors":["Zhanghao Hu","Qinglin Zhu","Runcong Zhao","Di Liang","Hanqi Yan","Yulan He","Lin Gui"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An integer upper bound must be rounded down","plainExplanation":"The Fano-based necessary upper bound is rounded upward, admitting an integer candidate count that violates the inequality. Under the stated parameters the bound allows eleven, not twelve. The issue is the rounding of a theoretical necessary bound under the assumed information budget. It does not validate that budget or refute an empirically useful cap of twelve.","primaryUrl":"https://arxiv.org/html/2602.02007","noteUrl":"/review?paper=review-neurips-2026-150607","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150608","title":"AnalogToBi: Device-Level Analog Circuit Topology Generation via Bipartite Graph and Grammar Guided Decoding","authors":["Seungmin Kim","Mingun Kim","Yuna Lee","Yulhwa Kim"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"AnalogToBi: Device-Level Analog Circuit Topology Generation via Bipartite Graph and Grammar Guided Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.08720","noteUrl":"/review?paper=review-neurips-2026-150608","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150611","title":"COSAC: Counterfactual Credit Assignment in Sequential Cooperative Teams","authors":["Shripad Deshmukh","Jayakumar Subramanian","Raghavendra Addanki","Nikos Vlassis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stated learnability optimum and coverage premise do not hold","plainExplanation":"The prefix-conditional baseline minimizes mean-square error, but need not maximize the paper’s variance-based learnability. The full one-hot Gram matrix also has unavoidable null directions, even with complete action coverage.","primaryUrl":"https://arxiv.org/html/2604.17693","noteUrl":"/review?paper=review-neurips-2026-150611","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","false_statement","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150612","title":"How to Instruct Your Robot: Dense Language Annotations Power Robot Policy Learning","authors":["Bosung Kim","Ruiyi Wang","David Acuna","Jaehun Jung","Alex Trevithick","Brandon Cui","Yejin Choi","Prithviraj Ammanabrolu"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"How to Instruct Your Robot: Dense Language Annotations Power Robot Policy Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17077","noteUrl":"/review?paper=review-neurips-2026-150612","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150616","title":"Learning the Context of Errors: Black-Box Online Adaptation of Time Series Foundation Models","authors":["Xilin Dai","Yiding Liu","Hongjie Xia","Yifan Hu","Zewei Dong","Jiang-Ming Yang","Qiang Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Learning the Context of Errors: Black-Box Online Adaptation of Time Series Foundation Models","plainExplanation":"EMA routing need not eliminate a permanently worse expert. The displayed sublinear static-regret bound therefore fails even on a stationary two-expert problem.","primaryUrl":"https://arxiv.org/html/2606.14222","noteUrl":"/review?paper=review-neurips-2026-150616","collection":"neurips-2026","severity":"major","issueTypes":["incorrect_bound","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-150619","title":"Tree-Structured Synergy of Large Language Models and Bayesian Optimization for Efficient CASH","authors":["Beicheng Xu","Weitong Qian","Lingching Tung","Yupeng Lu","Bin CUI"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Convergence requires the appendix assumptions and an adaptive-count repair","plainExplanation":"Fresh randomization at each visit does not remain independent after conditioning on a terminal visit count that depends on that randomization. The stated conditional binomial comparison can fail.","primaryUrl":"https://arxiv.org/html/2601.12355","noteUrl":"/review?paper=review-neurips-2026-150619","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150620","title":"MARS: Enabling Autoregressive Models Multi-Token Generation","authors":["Ziqi Jin","Lei Wang","Ziwei Luo","Aixin Sun"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MARS: Enabling Autoregressive Models Multi-Token Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.07023","noteUrl":"/review?paper=review-neurips-2026-150620","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150621","title":"Gym-Anything: Turn Any Software into an Agent Environment","authors":["Pranjal Aggarwal","Graham Neubig","Sean Welleck"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Gym-Anything: Turn Any Software into an Agent Environment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.06126","noteUrl":"/review?paper=review-neurips-2026-150621","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150623","title":"Pygmalion Effect in Vision: Image-to-Clay Translation for Reflective Geometry Reconstruction","authors":["Gayoung Lee","Junho Kim","Jin-Hwa Kim","Junmo Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Pygmalion Effect in Vision: Image-to-Clay Translation for Reflective Geometry Reconstruction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2511.21098","noteUrl":"/review?paper=review-neurips-2026-150623","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150624","title":"Optimal Contextual Pricing under Agnostic Non-Lipschitz Demand","authors":["Jianyu Xu","Yu-Xiang Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Optimal Contextual Pricing under Agnostic Non-Lipschitz Demand","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05609","noteUrl":"/review?paper=review-neurips-2026-150624","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150625","title":"DRScaffold: Boosting Dense-Scene Reasoning in Lightweight Vision Language Models","authors":["Xinrui Shi","Kai Liu","Ziqing Zhang","Jianze Li","Anqi Li","Yulun Zhang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"DRScaffold: Boosting Dense-Scene Reasoning in Lightweight Vision Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26038","noteUrl":"/review?paper=review-neurips-2026-150625","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150628","title":"Selling Information While Being an Interested Party","authors":["Francesco Bacchiocchi","Matteo Castiglioni","Alberto Marchesi","Giulia Romano","Nicola Gatti"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Selling Information While Being an Interested Party","plainExplanation":"Optimizing seller payoff averaged over buyer types does not preserve a guarantee for every type. The contract conversion also needs a valid mixing range, and the fixed-price program must distinguish the common posted price from individual affordability budgets.","primaryUrl":"https://arxiv.org/html/2301.13790","noteUrl":"/review?paper=review-neurips-2026-150628","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","lemma_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150631","title":"Target-Aligned Reinforcement Learning","authors":["Leonard S. Pleiss","James Harrison","Maximilian Schiffer"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Target-Aligned Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.29501","noteUrl":"/review?paper=review-neurips-2026-150631","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150635","title":"KV-PRM: Efficient Process Reward Modeling via KV-Cache Transfer for Multi-Agent Test-Time Scaling","authors":["Peng Kuang","Haibo Jin","Xiaoyu Han","Yanli Wang","Xiaopeng Yuan","Ye Yu","Kaidi Xu","Haohan Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"KV-PRM: Efficient Process Reward Modeling via KV-Cache Transfer for Multi-Agent Test-Time Scaling","plainExplanation":"Capacity, reachable support and retained reward information are different properties. The representation assumptions do not supply the stated guarantees for an arbitrary generated cache or readout.","primaryUrl":"https://arxiv.org/html/2607.09153","noteUrl":"/review?paper=review-neurips-2026-150635","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150636","title":"Oracle-Robust Online Alignment for Large Language Models","authors":["Zimeng Li","Mudit G Gaur","Vaneet Aggarwal"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Oracle-Robust Online Alignment for Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.20457","noteUrl":"/review?paper=review-neurips-2026-150636","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150641","title":"HOMIE: Human-object Centric Video Personalization via Multimodal Intelligent Enhancement","authors":["Yiyang Cai","Nan Chen","Rongchang Xie","Junwen Pan","Chunyang Jiang","Cheng CHEN","Chang Liu","Zhenbang Sun","Wei Xue","Wenhan Luo","Yike Guo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"HOMIE: Human-object Centric Video Personalization via Multimodal Intelligent Enhancement","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.18217","noteUrl":"/review?paper=review-neurips-2026-150641","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150642","title":"VLMs Trace Without Tracking: Diagnosing Failures in Visual Path Following","authors":["Hyesoo Hong","Min S Kim","Wonje Jeung","Yoon Sangyeon","Dongjae Jeon","Albert No"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"VLMs Trace Without Tracking: Diagnosing Failures in Visual Path Following","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15672","noteUrl":"/review?paper=review-neurips-2026-150642","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150643","title":"Remember the Decision, Not the Description: A Rate-Distortion Framework for Agent Memory","authors":["mingxi Zou","Zhihan Guo","Langzhang Liang","Zhuo Wang","Qifan Wang","Qingsong Wen","Irwin King","Lizhen Qu","Zenglin Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Remember the Decision, Not the Description: A Rate-Distortion Framework for Agent Memory","plainExplanation":"The statistical lower bound needs boundary and saturation restrictions. The restarted upper bound must retain the distortion actually certified by the algorithm unless its relation to the optimum is proved.","primaryUrl":"https://arxiv.org/html/2605.10870","noteUrl":"/review?paper=review-neurips-2026-150643","collection":"neurips-2026","severity":"major","issueTypes":["incorrect_bound","invalid_inference","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150644","title":"Tapes Together Strong: The Co-evolution of Computation and Cooperation","authors":["Kunal Jha","Francesco Cicala","Blaise Aguera y Arcas","Blake Richards","Natasha Jaques","Max Kleiman-Weiner","Eyvind Niklasson"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Positive cooperation drift requires population-weighted transition rates","plainExplanation":"A baseline energy inequality does not prevent replication from retained-energy states. The witness concerns the universal extension to retained-energy states. It does not refute a baseline-only theorem, aggregate expected drift, or the stated conjectural rebound interpretation.","primaryUrl":"https://arxiv.org/html/2609.10817","noteUrl":"/review?paper=review-neurips-2026-150644","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150646","title":"Amortized Vine Copulas for High-Dimensional Density and Information Estimation","authors":["Houman Safaai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Amortized Vine Copulas for High-Dimensional Density and Information Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.20568","noteUrl":"/review?paper=review-neurips-2026-150646","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150647","title":"Computing Thiele Rules on Interval Elections and their Generalizations","authors":["Dimitris Avramidis","Alexandra Anna Lassota","Ulrike Schmidt-Kraepelin","Adrian Vetta"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Computing Thiele Rules on Interval Elections and their Generalizations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.03067","noteUrl":"/review?paper=review-neurips-2026-150647","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150649","title":"Improving LLM Final Representations with Inter-Layer Geometry","authors":["Tom Ulanovski","Eyal Blyachman","Maya Bechler-Speicher"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Improving LLM Final Representations with Inter-Layer Geometry","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.22665","noteUrl":"/review?paper=review-neurips-2026-150649","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150650","title":"Understanding Schedule-Free Methods in Nonconvex Optimization: Rate Guarantees and Escaping Saddles","authors":["Jiseok Chae","Donghwan Kim"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Understanding Schedule-Free Methods in Nonconvex Optimization: Rate Guarantees and Escaping Saddles","plainExplanation":"The Lyapunov proposition omits the step-size bound already used in its proof and later corollaries. The smooth-interpolation condition separately needs the squared norm appearing in the cited theorem.","primaryUrl":"https://arxiv.org/html/2607.09167","noteUrl":"/review?paper=review-neurips-2026-150650","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150651","title":"TreeWalker: Partial Evaluation for Grouped Tree-Ensemble Inference","authors":["Durmus Karatay","Richard Newman"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"TreeWalker: Partial Evaluation for Grouped Tree-Ensemble Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03939","noteUrl":"/review?paper=review-neurips-2026-150651","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150652","title":"GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation","authors":["Jeonghyeok Do","Munchurl Kim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.01896","noteUrl":"/review?paper=review-neurips-2026-150652","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150654","title":"Majority Bit-Aware Watermarking for Large Language Models","authors":["Jiahao Xu","Rui Hu","Olivera Kotevska","Zikai Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Majority Bit-Aware Watermarking for Large Language Models","plainExplanation":"The printed decoder search omits allowed majority-count hypotheses. Its strict finite-codebook expectation comparison also fails for permitted small blocks; the probability law and usable-message exclusions must enter the exact calculation.","primaryUrl":"https://arxiv.org/html/2508.03829","noteUrl":"/review?paper=review-neurips-2026-150654","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","false_statement","incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150655","title":"Weisfeiler-Leman Is Incomplete on Simple Spectrum Graphs, so Canonicalize Them","authors":["Snir Hordan","Nadav Dym","Tim Seppelt"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Weisfeiler-Leman Is Incomplete on Simple Spectrum Graphs, so Canonicalize Them","plainExplanation":"The spectral construction must account for column norms when choosing distinct eigenvalues. Weighted-graph universality needs a continuous target, and the parity-check rank cannot be bounded by confusing a difference space with its complementary constraint space.","primaryUrl":"https://arxiv.org/html/2605.23446","noteUrl":"/review?paper=review-neurips-2026-150655","collection":"neurips-2026","severity":"repairable","issueTypes":["complexity_error","false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150657","title":"Where to Look Matters: Rethinking Sub-Volume Sampling in 3D Medical Self-supervised Learning","authors":["Junkai Liu","Le Zhang"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Where to Look Matters: Rethinking Sub-Volume Sampling in 3D Medical Self-supervised Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.13915","noteUrl":"/review?paper=review-neurips-2026-150657","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150658","title":"Certification-Enhanced Generalization Bounds","authors":["Leo Elmecker-Plakolm","Matthew R Wicker"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Certification-Enhanced Generalization Bounds","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06238","noteUrl":"/review?paper=review-neurips-2026-150658","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150659","title":"Stealth Apart, Harm Together: Skill Cascading Attacks on Skill-Based Agent Systems","authors":["Zihao Zhu","Siwei Lyu","Adel Bibi","Baoyuan Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Stealth Apart, Harm Together: Skill Cascading Attacks on Skill-Based Agent Systems","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.30383","noteUrl":"/review?paper=review-neurips-2026-150659","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150660","title":"AUTOMEM: Automated Learning of Memory as a Cognitive Skill","authors":["Shengguang Wu","Hao Zhu","Yuhui Zhang","Xiaohan Wang","Serena Yeung-Levy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AUTOMEM: Automated Learning of Memory as a Cognitive Skill","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.01224","noteUrl":"/review?paper=review-neurips-2026-150660","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150661","title":"RISE: Red-teaming via Iterative Strategy Evolution for Modern Text-to-Image Models","authors":["Dmitrii Kharlapenko","Sergei Bratchikov","Konstantin Korolev","Aleksandr Nikolich"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RISE: Red-teaming via Iterative Strategy Evolution for Modern Text-to-Image Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34920","noteUrl":"/review?paper=review-neurips-2026-150661","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150663","title":"Depth-Recurrent Attention Mixtures: Giving Latent Reasoning the Attention it Deserves","authors":["Jonas Knupp","Jan Metzen","Jeremias Bohn","Georg Groh","Kristian Kersting"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Median-sign updates do not preserve mean-zero bias when counts tie","plainExplanation":"Median-sign updates do not preserve mean-zero bias when counts tie. The printed mean-zero invariant is false on allowed integer usage counts; the routing mechanism need not otherwise fail.","primaryUrl":"https://arxiv.org/html/2601.21582","noteUrl":"/review?paper=review-neurips-2026-150663","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150664","title":"DAGent: Evaluate-then-Grow Planning for Deep Research Agents","authors":["Hanwen Liu","Yuanfu Sun","Qiaoyu Tan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DAGent: Evaluate-then-Grow Planning for Deep Research Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39154","noteUrl":"/review?paper=review-neurips-2026-150664","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150665","title":"AlphaQ: Calibration-Free Bit Allocation for Mixture-of-Experts Quantization","authors":["Wanqi Yang","Yuexiao Ma","Alexander Conzelmann","Xiawu Zheng","Michael Mahoney","T. Konstantin Rusch","Shiwei Liu"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The Hill estimator requires an explicit switch to an unbounded Pareto proxy","plainExplanation":"The Hill estimator requires an explicit switch to an unbounded Pareto proxy. The exact-rewriting interpretation is wrong; the proxy model and compression method are not invalidated merely by this distinction.","primaryUrl":"https://arxiv.org/html/2606.04980","noteUrl":"/review?paper=review-neurips-2026-150665","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150667","title":"HEED: Density-Weighted Residual Alignment for Hybrid Vision-Language Model Distillation","authors":["Yihao Liang","Niraj Jha"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"HEED: Density-Weighted Residual Alignment for Hybrid Vision-Language Model Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17093","noteUrl":"/review?paper=review-neurips-2026-150667","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150668","title":"Unified High-Probability Analysis of Stochastic Variance-Reduced Estimation","authors":["Zhankun Luo","Antesh Upadhyay","M. B Sahin","Sang Bin Moon","Anuran Makur","Abolfazl Hashemi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The reciprocal-product representation excludes zero momentum","plainExplanation":"The non-reset recurrence cannot be applied at a reset endpoint. The explicit endpoint equality and resulting Appendix B.3 all-time bound are invalid. Family-specific asymptotic optimization rates are not refuted.","primaryUrl":"https://arxiv.org/html/2605.15388","noteUrl":"/review?paper=review-neurips-2026-150668","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150671","title":"Federated Learning by Utility-Constrained Stochastic Aggregation for Improving Rational Participation","authors":["Yashwanth Mandula","Arunabh Singh","Ashok Nayak","Saikiran Bulusu","Anirban Chakraborty"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Federated Learning by Utility-Constrained Stochastic Aggregation for Improving Rational Participation","plainExplanation":"A feasible mixture already supplies a deterministic aggregate under the scalar concave-utility model. The componentwise utility inference, large-concentration Jensen-gap limit and fallback construction each need additional conditions beyond the stated balanced slack program.","primaryUrl":"https://arxiv.org/html/2605.18020","noteUrl":"/review?paper=review-neurips-2026-150671","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement","incorrect_bound","invalid_inference","lemma_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150673","title":"Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback","authors":["Ved Sriraman","Peihan Liu","Daniel Hsu","Adam Block"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30923","noteUrl":"/review?paper=review-neurips-2026-150673","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150675","title":"Learning Implicit Bias in Generative Spaces for Accelerating Protein Dynamics Emulation","authors":["Kaihui Cheng","Zhiqiang Cai","Wenkai Xiang","Zhihang Hu","Siyu Zhu","Tzuhsiung Yang","Yuan Qi"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Learning Implicit Bias in Generative Spaces for Accelerating Protein Dynamics Emulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.01833","noteUrl":"/review?paper=review-neurips-2026-150675","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150679","title":"Mixture-of-Top-$k$ Attention: Efficient Attention as Scalable Fast Weights","authors":["Qishuai Wen","Zhiyuan Huang","meng xianghan","Wei He","CHUNGUANG LI"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Mixture-of-Top-$k$ Attention: Efficient Attention as Scalable Fast Weights","plainExplanation":"The shared compressed expert changes the final softmax even when the routed branch contains every original key and value. Exact recovery of full attention therefore requires disabling that branch or proving its output agrees with full attention.","primaryUrl":"https://arxiv.org/html/2602.01219","noteUrl":"/review?paper=review-neurips-2026-150679","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150680","title":"The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models","authors":["Nguyen Phuc","Chinh D La","Duy M. H. Nguyen","Nitesh Chawla","Binh T. Nguyen","Khoa D Doan"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The greedy-failure indicator has both a type error and the wrong polarity","plainExplanation":"The greedy-failure indicator has both a type error and the wrong polarity. The equation contradicts the stated selection semantics; it does not establish that an executable implementation used the same gate.","primaryUrl":"https://arxiv.org/html/2510.02230","noteUrl":"/review?paper=review-neurips-2026-150680","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150682","title":"Per-Loss Adapters for Gradient Conflict in Physics-Informed Neural Networks","authors":["Bum Jun Kim","Gnankan Landry Regis N'guessan"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Per-Loss Adapters for Gradient Conflict in Physics-Informed Neural Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10136","noteUrl":"/review?paper=review-neurips-2026-150682","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150683","title":"World Models as Group Actions","authors":["Zijie Wang","Wei Zhang","Weiming Zhang","Fanqi Zhang","Xiao Tan","Yipeng Qin","Guanbin Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Group-action axioms do not calibrate commanded physical motion","plainExplanation":"Group-action axioms do not calibrate commanded physical motion. The group-action theorem remains correct; physical command agreement requires this additional interpretation.","primaryUrl":"https://arxiv.org/html/2605.24578","noteUrl":"/review?paper=review-neurips-2026-150683","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150684","title":"Mamba Can Learn Low-Dimensional Targets In-Context via Test-Time Feature Learning","authors":["Junsoo Oh","Wei Huang","Taiji Suzuki"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Mamba Can Learn Low-Dimensional Targets In-Context via Test-Time Feature Learning","plainExplanation":"The learning proof uses a gate-bias regime opposite to the theorem, a concentration lemma without tail-moment control, and a sphere contraction that retains an extra quadratic statistic. The supplied local repairs specify the missing conditions and a fixed coordinate reweighting without claiming the full learning theorem follows automatically.","primaryUrl":"https://arxiv.org/html/2510.12026","noteUrl":"/review?paper=review-neurips-2026-150684","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","lemma_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150685","title":"Predicting Only from Selected Evidence: A Tempered Product-of-Experts Bottleneck for Auditable EEG Diagnosis","authors":["Yinghao WANG","Shujian Yu","Duc-Han LE","Zhikai Yu","Changming Wang","Van-Tam Nguyen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Predicting Only from Selected Evidence: A Tempered Product-of-Experts Bottleneck for Auditable EEG Diagnosis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.24377","noteUrl":"/review?paper=review-neurips-2026-150685","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150686","title":"Free energy Estimation on Any State Space","authors":["Jiajun He","Zijing Ou","Francisco Vargas","Yingzhen Li","José Miguel Hernández-Lobato","Carles Domingo i Enrich","Yuanqi Du"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Real linear combinations of generators need not be Markov generators","plainExplanation":"Real linear combinations of generators need not be Markov generators. The all-real-coefficient dynamics claim is false; the work-cancellation algebra remains valid on its admissible domain.","primaryUrl":"https://arxiv.org/html/2605.31063","noteUrl":"/review?paper=review-neurips-2026-150686","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150688","title":"CLR-voyance : Reinforcing Open-Ended Reasoning for Inpatient Clinical Decision Support with Outcome-Aware Rubrics","authors":["Aishik Nagar","Arun-Kumar Kaliya-Perumal","Yu-Hsuan Han","Andrew Sheng-Han Huang","Kristen Kee","Yushi Cao","Yiming Chen","Hongchao Jiang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"CLR-voyance : Reinforcing Open-Ended Reasoning for Inpatient Clinical Decision Support with Outcome-Aware Rubrics","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.09584","noteUrl":"/review?paper=review-neurips-2026-150688","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150689","title":"Self-Rewarding Sequential Monte Carlo for Masked Diffusion Language Models","authors":["Ziwei Luo","Ziqi Jin","Lei Wang","Lidong Bing","Thomas Schön"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The confidence-based sampler needs consistent normalization and transition definitions","plainExplanation":"The selected set depends jointly on the provisional tokens. After averaging over that latent selection, the position transitions are dependent.","primaryUrl":"https://arxiv.org/html/2602.01849","noteUrl":"/review?paper=review-neurips-2026-150689","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150690","title":"COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection","authors":["JINWOO PARK","Hyeongwon Kang","Seung H Han","Pilsung Kang"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The budget optimization does not require the mandatory variable","plainExplanation":"The optimization ranges over every \\(B\\)-element subset without requiring index one. With \\(B=1\\) and absolute standardized deviations \\((10,0,1)\\), its unique minimizing subset is \\(\\{2\\}\\). The union with \\(\\{1\\}\\) in the separate threshold rule does not automatically carry into this alternative optimization. This concerns only the budget-rule constraint; the threshold rule already includes the mandatory variable.","primaryUrl":"https://arxiv.org/html/2602.01635","noteUrl":"/review?paper=review-neurips-2026-150690","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150691","title":"Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima","authors":["Huanran Chen","Huaqing Zhang","Xiao Li","Yinpeng Dong","Ke Shen","Jun Zhu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Normalized inner gradients do not give the stated raw-loss objective","plainExplanation":"For the normalized update, the squared-step term is \\(\\gamma^2\\), not \\(\\gamma^2\\|\\nabla L\\|^2\\), so the proof’s gradient-norm cancellation is unavailable. With one quadratic loss \\(L(\\theta)=\\theta^2/2\\), one inner step and allowed unit-step outer SGD, Nexus is \\(\\theta_{t+1}=\\theta_t-\\gamma\\operatorname{sign}\\theta_t\\). Starting at \\(\\theta_0=\\gamma/4\\), it cycles between \\(\\gamma/4\\) and \\(-3\\gamma/4\\), violating convergence to zero and the claimed geometric bound. The auxiliary theorem is misattributed and false for an admitted normalized sequence. This does not invalidate the empirical generalization claims.","primaryUrl":"https://arxiv.org/html/2604.09258","noteUrl":"/review?paper=review-neurips-2026-150691","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150692","title":"RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models","authors":["Hao Wu","Yuqi Li","Yuan Gao","Fan Xu","Fan Zhang","Kun Wang","Penghao Zhao","Qiufeng Wang","Yizhou Zhao","Weiyan Wang","YingLi Tian","Xian Wu","Xiaomeng Huang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models","plainExplanation":"Upper bounds on two prediction errors do not imply a lower bound on their difference. The printed straight-through quantizer also returns its continuous input in the forward pass, violating the finite-grid invariant required by the index.","primaryUrl":"https://arxiv.org/html/2605.03821","noteUrl":"/review?paper=review-neurips-2026-150692","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","formula_error","incorrect_bound","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150694","title":"Understanding Private Evolution as Learning-Augmented Clustering","authors":["Audra McMillan","Kunal Talwar","Felix Zhou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Understanding Private Evolution as Learning-Augmented Clustering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36678","noteUrl":"/review?paper=review-neurips-2026-150694","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150695","title":"Contrastive Identification and Generation in the Limit","authors":["Xiaoyu Li","Andi Han","Jiaojiao Jiang","Junbin Gao"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Contrastive Identification and Generation in the Limit","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06211","noteUrl":"/review?paper=review-neurips-2026-150695","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150696","title":"SBNO : Schrödinger Bridge Neural Operator for the Forward and Inverse Problems under Missing Information","authors":["Jaehyeon Park","Inhyeok Jeong","Noseong Park"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SBNO : Schrödinger Bridge Neural Operator for the Forward and Inverse Problems under Missing Information","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.13317","noteUrl":"/review?paper=review-neurips-2026-150696","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150698","title":"ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning Model","authors":["Haichao Zhang","Yijiang Li","Shwai He","Tushar Nagarajan","Mingfei Chen","Jianglin Lu","Ang Li","Yun Fu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning Model","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.22281","noteUrl":"/review?paper=review-neurips-2026-150698","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150701","title":"Reconciling Causality and Non-Equilibrium Thermodynamics with Hamiltonian Causal Models","authors":["Dario Rancati","Max Welling","Francesco Locatello"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The reversal identity orders its reversed endpoints incorrectly","plainExplanation":"The reversal identity orders its reversed endpoints incorrectly. This is a local reversibility-identity endpoint error; broader entropy-production claims require their own definitions.","primaryUrl":"https://arxiv.org/html/2606.04822","noteUrl":"/review?paper=review-neurips-2026-150701","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150704","title":"DegBins: Degradation-Driven Binning for Depth Super-Resolution","authors":["Zhiqiang Yan","Zhengxue Wang","Jian Yang","Gim Hee Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DegBins: Degradation-Driven Binning for Depth Super-Resolution","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.09628","noteUrl":"/review?paper=review-neurips-2026-150704","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150705","title":"Model-Agnostic FDR Control via Group Gaussian Mirror and Permutation SHAP","authors":["Jiaan Han","Junxiao Chen","Yanzhe Fu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Marginal mirror symmetry alone does not control a joint adaptive FDR threshold","plainExplanation":"Marginal mirror symmetry alone does not control a joint adaptive FDR threshold. The proof lacks a joint-dependence step. This abstract witness is not a full-generator FDR counterexample.","primaryUrl":"https://arxiv.org/pdf/2608.00989v1","noteUrl":"/review?paper=review-neurips-2026-150705","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150707","title":"Geometry over Density: Few-Shot Cross-Domain OOD Detection","authors":["Li Li","You Qin","Jiate Li","Charith Peris","Lisa Bauer","Roger Zimmermann","Yue Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Balanced transport from one point has no free target weights","plainExplanation":"For iid \\(\\epsilon_i\\sim N(0,\\theta I_d)\\), the sample mean is \\(M\\sim N(0,\\theta I_d/n)\\). Its score is \\(-d/(2\\theta)+n\\|M\\|^2/(2\\theta^2)\\). Since \\(\\operatorname{Var}\\|M\\|^2=2d\\theta^2/n^2\\), its Fisher information is \\(d/(2\\theta^2)>0\\). A parameter-independent expectation does not make the sampling distribution parameter independent. The sample-energy efficiency and Cramér–Rao result can remain valid. Only the sample-mean zero-information exclusion is false.","primaryUrl":"https://arxiv.org/html/2605.03410","noteUrl":"/review?paper=review-neurips-2026-150707","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150708","title":"RA-CFGCache: From Branch-Level Criteria to Guided-Risk Control under Classifier-Free Guidance","authors":["Yiming Liu","Ben Wan","Hui Chen","Ao Wang","Yuqi Xiong","Fan Zhang","Tongxuan Liu","guiguang ding"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RA-CFGCache: From Branch-Level Criteria to Guided-Risk Control under Classifier-Free Guidance","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36433","noteUrl":"/review?paper=review-neurips-2026-150708","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150709","title":"Bridging the Modality Bottleneck in Pathology MIL through Virtual Molecular Staining","authors":["Yucheng Xing","Pei Liu","Jingying Ma","Ruping Hong","Jiangdong Qiu","Tianyu Liu","Kai He","Ling Huang","Mengling Feng"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The separation identity replaces a pair mean by an unrelated uniform baseline","plainExplanation":"The separation identity replaces a pair mean by an unrelated uniform baseline. The displayed proof step is wrong; existence of increased separation and empirical gains are not ruled out.","primaryUrl":"https://arxiv.org/html/2605.16392","noteUrl":"/review?paper=review-neurips-2026-150709","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150710","title":"Rank-Constrained Adaptation for Reliable Real-World Performance","authors":["Abinitha Gourabathina","Hyewon Jeong","Teya Bergamaschi","Marzyeh Ghassemi","Collin Stultz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rank-Constrained Adaptation for Reliable Real-World Performance","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.06924","noteUrl":"/review?paper=review-neurips-2026-150710","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150712","title":"ViCO: A Training Strategy towards Semantic Aware Dynamic High-Resolution","authors":["Long Cui","Weiyun Wang","Jie Shao","Zichen Wen","Gen Luo","Zhang","Linfeng Zhang","Wenhai Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ViCO: A Training Strategy towards Semantic Aware Dynamic High-Resolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.12793","noteUrl":"/review?paper=review-neurips-2026-150712","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150714","title":"Distilling Graph Geometry: Knowledge Gap from GNNs to MLPs","authors":["Zhewei Chen","Hao Zhu","Jiaojiao Jiang","Ahad N. Zehmakan"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The covariance recursion uses a derivative kernel in place of an activation covariance","plainExplanation":"The covariance recursion uses a derivative kernel in place of an activation covariance. The printed covariance recursion is incorrect; a corrected recursion can still preserve structural locality.","primaryUrl":"https://arxiv.org/html/2610.10520","noteUrl":"/review?paper=review-neurips-2026-150714","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150719","title":"Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion","authors":["Van Chien Nguyen","Chaitra Hegde","Van-Cuong Pham","Ryan Rossi","Franck Dernoncourt","Thien H Nguyen"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Speculative rejection needs the appropriate correction distribution","plainExplanation":"Under standard speculative acceptance, rejection needs a positive-residual correction rather than a fresh independent target draw. The author-linked implementation already uses the ideal residual branch; a precisely specified target-first coupling is another valid construction.","primaryUrl":"https://arxiv.org/html/2605.12825","noteUrl":"/review?paper=review-neurips-2026-150719","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150722","title":"Conformal Agent Error Attribution","authors":["Naihe Feng","Yi Sui","Shiyi Hou","Ga Wu","Jesse Cresswell"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Conformal Agent Error Attribution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06788","noteUrl":"/review?paper=review-neurips-2026-150722","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150723","title":"MATO: Multi-objective Personalized Alignment with Test-time Optimization for Large Language Models","authors":["Linhao Luo","Trang Vu","Van-Anh Nguyen","Junae Kim","Reza Haffari","Dinh Phung"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The reward-density interpretation omits the anchor distribution","plainExplanation":"The reward-density interpretation omits the anchor distribution. The appendix interpretation differs from anchored decoding; Eq. (10) itself remains well defined.","primaryUrl":"https://arxiv.org/html/2605.25342","noteUrl":"/review?paper=review-neurips-2026-150723","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150724","title":"FacePhys: State of the Heart Learning","authors":["Jack Tang","Kegang Wang","Yuntao Wang","Xin Liu","Yuxuan Fan","Jiatong Ji","Yuanchun Shi","Daniel McDuff"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FacePhys: State of the Heart Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.06275","noteUrl":"/review?paper=review-neurips-2026-150724","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150725","title":"Optimistic Dual Averaging Unifies Modern Optimizers","authors":["Thomas Pethick","Wanyun Xie","Roman Machacek","Volkan Cevher"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Optimistic Dual Averaging Unifies Modern Optimizers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11172","noteUrl":"/review?paper=review-neurips-2026-150725","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150726","title":"Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes","authors":["LIU Hanqing","Jianjun Cao","Yuanze Li","Zijian Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes","plainExplanation":"The binary32 absorption threshold needs directional half-spacing and rounding conditions. Label-smoothed logit curvature also needs a nonzero contrast-space Jacobian before it implies positive parameter curvature.","primaryUrl":"https://arxiv.org/html/2605.06152","noteUrl":"/review?paper=review-neurips-2026-150726","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","invalid_inference","missing_assumption","numerical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150727","title":"Dynamic Regret in Online Convex Optimization with Indicator Switching Costs","authors":["Naram Mhaisen","George Iosifidis"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Dynamic Regret in Online Convex Optimization with Indicator Switching Costs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30556","noteUrl":"/review?paper=review-neurips-2026-150727","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150730","title":"Multivariate Time Series Forecasting needs Cross Variable Loss","authors":["Kuiye Ding","Yifan Hu","Hanchen Wang","Hao Xue"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Multivariate Time Series Forecasting needs Cross Variable Loss","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.05742","noteUrl":"/review?paper=review-neurips-2026-150730","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150731","title":"Property-Guided LLM Program Synthesis for Planning","authors":["André G. Pereira","Augusto B. Corrêa","Jendrik Seipp"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The directness definition is vacuous on unsolvable tasks","plainExplanation":"Alive means reachable, solvable and non-goal. If the initial task is unsolvable, no reachable state can reach a goal, so the alive set and \\(S_{\\mathrm{down}}\\) are empty. Every heuristic then satisfies the universal directness clauses. For an initial dead-end \\(s_0\\) and a separate unreachable goal \\(g\\), with no outgoing edges, the printed definition calls any heuristic direct although no plan exists. The background iff fails as defined. The operational DFS on solvable benchmarks is not refuted. Large saved enumeration counts depend on whether empty goal sets and self-loops were admitted and are not needed for this witness.","primaryUrl":"https://arxiv.org/html/2605.16142","noteUrl":"/review?paper=review-neurips-2026-150731","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150732","title":"Coherence Mechanisms for Provable Self-Improvement","authors":["Mehryar Mohri","Jon Schneider","Yifan Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Coherence Mechanisms for Provable Self-Improvement","plainExplanation":"The universal-distribution theorem has a point-mass proof that replaces the invalid spanning argument. Separate claims need to allow seed blocks to merge and to require a positive-definite Hessian before calling a Taylor quadratic Legendre.","primaryUrl":"https://arxiv.org/html/2511.08440","noteUrl":"/review?paper=review-neurips-2026-150732","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","lemma_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150736","title":"Soft-Radial Projection for Constrained End-to-End Learning","authors":["Philipp Schneider","Daniel Kuhn"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Soft-Radial Projection for Constrained End-to-End Learning","plainExplanation":"Absolute continuity of the inputs does not ensure a smooth expected objective after projection. Stationary cluster points also do not imply that the minimum Clarke-subgradient norm at every iterate tends to zero.","primaryUrl":"https://arxiv.org/html/2602.03461","noteUrl":"/review?paper=review-neurips-2026-150736","collection":"neurips-2026","severity":"repairable","issueTypes":["citation_hypothesis_mismatch","false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150737","title":"Breadcrumbing Search Agents: Per-Turn Scheming Over Long-Horizon Trajectories","authors":["Xuebin Li","Hanqing Zhao","Siyuan Liang","Kejiang Chen","Weiming Zhang","Dacheng Tao","Nenghai Yu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Breadcrumbing Search Agents: Per-Turn Scheming Over Long-Horizon Trajectories","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.04565v2","noteUrl":"/review?paper=review-neurips-2026-150737","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150740","title":"SPHQuant: Efficient extreme low bit weight quantization for Vision-Language Models","authors":["Kewei Zhang","郑 陈","Haotong Qin","Yulun Zhang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SPHQuant: Efficient extreme low bit weight quantization for Vision-Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.24875","noteUrl":"/review?paper=review-neurips-2026-150740","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150745","title":"AutoformBot: Formalizing Mathematics at Scale","authors":["Ahmad Rammal","Niket Patel","Fabian Gloeckle","Amaury Hayat","Julia Kempe","Remi Munos","Charles Arnal","Vivien Cabannes"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AutoformBot: Formalizing Mathematics at Scale","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29955","noteUrl":"/review?paper=review-neurips-2026-150745","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150747","title":"When Metropolis and Hastings Meet Bradley and Terry: Exact MCMC From Preference Voting","authors":["Ariel Smogorghevski","Nir Rosenfeld","Yaniv Romano"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"When Metropolis and Hastings Meet Bradley and Terry: Exact MCMC From Preference Voting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.00905","noteUrl":"/review?paper=review-neurips-2026-150747","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150748","title":"Sinkhorn Based Associative Memory Retrieval Using Spherical Hellinger Kantorovich Dynamics","authors":["Aratrika Mustafi","Soumya Mukherjee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The softmin proof puts an absolute value on a one-sided bound","plainExplanation":"The proof bounds an absolute softmin difference by a signed maximum that can be negative. The one-Lipschitz lemma is true, but this one-sided intermediate inequality needs the arguments swapped as well. This is a local sign/proof typo. The corrected argument preserves the lemma and its constant one.","primaryUrl":"https://arxiv.org/html/2603.20656","noteUrl":"/review?paper=review-neurips-2026-150748","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150749","title":"Label-Efficient Dataset Pruning via Semi-Supervised Pseudo-Labeling","authors":["Yeseul Cho","Baekrok Shin","Changmin Kang","Chulhee Yun"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Label-Efficient Dataset Pruning via Semi-Supervised Pseudo-Labeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23198","noteUrl":"/review?paper=review-neurips-2026-150749","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150750","title":"Robust Policy Optimization to Prevent Catastrophic Forgetting","authors":["Mahdi Sabbaghi","George J. Pappas","Adel Javanmard","Hamed Hassani"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Clipping makes the importance-ratio baseline expectation parameter dependent","plainExplanation":"Expanding \\(-\\mathbb E_x[\\lambda\\log\\mathbb E_{y\\mid x}e^{-r/\\lambda}]-\\lambda\\rho\\) gives \\(\\mathbb Er-\\mathbb E_x\\operatorname{Var}(r\\mid x)/(2\\lambda)-\\lambda\\rho\\). Total variance adds \\(\\operatorname{Var}_x\\mathbb E[r\\mid x]\\), which response reweighting cannot alter. For fair binary \\(x\\) and reward \\(r(x,y)=x\\), every response policy earns \\(1/2\\); conditional variance is zero while total variance is \\(1/4\\), producing a spurious penalty. The exact fixed-\\(\\lambda\\) conditional dual survives. The unconditional variance interpretation and associated tuning do not follow from it.","primaryUrl":"https://arxiv.org/html/2602.08813","noteUrl":"/review?paper=review-neurips-2026-150750","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150757","title":"Spectral-Sphere-Constrained Hyper-Connections","authors":["Zhaoyi Liu","Haichuan Zhang","Ang Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Spectral-Sphere-Constrained Hyper-Connections","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.20896","noteUrl":"/review?paper=review-neurips-2026-150757","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150758","title":"Palette: A Modular, Controllable, and Efficient Framework for On-demand Authorized Safety Alignment Relaxation in LLMs","authors":["Qitao Tan","Xiaoying Song","Arman Akbari","Arash Akbari","Yanzhi Wang","Xiaoming Zhai","Lingzi Hong","Zhen Xiang","Jin Lu","Geng Yuan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Palette: A Modular, Controllable, and Efficient Framework for On-demand Authorized Safety Alignment Relaxation in LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.24154","noteUrl":"/review?paper=review-neurips-2026-150758","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150759","title":"Learning Latency-Aware Orchestration for Multi-Agent Systems","authors":["Xi Shi","Mengxin Zheng","Qian Lou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Latency-Aware Orchestration for Multi-Agent Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2601.10560v2","noteUrl":"/review?paper=review-neurips-2026-150759","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150760","title":"Integrating Local and Global Entropy for Uncertainty Quantification in LLMs","authors":["Johanne Medina","Tianyi Zhou","Keivin Isufaj","Aristides Gionis","Sanjay Chawla"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Largest reliability values select the opposite uncertainty ordering","plainExplanation":"Largest reliability values select the opposite uncertainty ordering. The printed order-statistic convention is inconsistent; the text does not establish which ordering was executed.","primaryUrl":"https://arxiv.org/html/2606.09875","noteUrl":"/review?paper=review-neurips-2026-150760","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150762","title":"Panoptic Scene Program Diffusion Transformer","authors":["Chika Maduabuchi"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Panoptic Scene Program Diffusion Transformer","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.31780","noteUrl":"/review?paper=review-neurips-2026-150762","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150763","title":"Natural Language Actor-Critic: Scalable Off-Policy Learning in Language Space","authors":["Joey Hong","Kang Liu","Zhan Ling","Jiecao Chen","Sergey Levine"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Natural Language Actor-Critic: Scalable Off-Policy Learning in Language Space","plainExplanation":"The critic assumptions do not guarantee scalar-value recovery, and weak refinement does not guarantee optimal policies. The displayed divergence objective has the wrong optimization direction.","primaryUrl":"https://arxiv.org/html/2512.04601","noteUrl":"/review?paper=review-neurips-2026-150763","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","formula_error","missing_assumption"],"classifications":["critical"]},{"slug":"review-neurips-2026-150765","title":"A Free Lunch in LLM Compression: Revisiting Retraining after Pruning","authors":["Moritz Wagner","Christophe Roux","Max Zimmer","Sebastian Pokutta"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Free Lunch in LLM Compression: Revisiting Retraining after Pruning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.14444","noteUrl":"/review?paper=review-neurips-2026-150765","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150766","title":"Coarse-to-Real: Generative Rendering for Populated Dynamic Scenes","authors":["Gonzalo Gomez-Nogales","Yicong Hong","Chongjian GE","Peiye Zhuang","Marc Comino-Trinidad","Dan Casas","Yi Zhou"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Coarse-to-Real: Generative Rendering for Populated Dynamic Scenes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.22301","noteUrl":"/review?paper=review-neurips-2026-150766","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150768","title":"SkipSR: Faster Super-Resolution with Token Skipping","authors":["Rohan Choudhury","Shanchuan Lin","Jianyi Wang","Hao Chen","Qi Zhao","Feng Cheng","Lu Jiang","Kris Kitani","László Jeni"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"SkipSR: Faster Super-Resolution with Token Skipping","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.08799","noteUrl":"/review?paper=review-neurips-2026-150768","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150769","title":"LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling","authors":["Yuxin Chen","Chumeng Liang","Hangke Sui","Ruihan Guo","Chaoran Cheng","Jiaxuan You","Ge Liu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.11748","noteUrl":"/review?paper=review-neurips-2026-150769","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150770","title":"What do EEG Foundation Models Capture from Human Brain Signals?","authors":["Ling Tang","Qian Chen","Jilin Mei","Houshi Xu","Quanshi Zhang","Jing Shao","Na Zou","Xia Hu","Dongrui Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"What do EEG Foundation Models Capture from Human Brain Signals?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11410","noteUrl":"/review?paper=review-neurips-2026-150770","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150771","title":"Reducing Credit Assignment Variance via Counterfactual Reasoning Paths","authors":["fei ding","Yongkang Zhang","youwei wang","Zijian Zeng"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Reducing Credit Assignment Variance via Counterfactual Reasoning Paths","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16302","noteUrl":"/review?paper=review-neurips-2026-150771","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150772","title":"Fast and Stable Triangular Inversion for Delta-Rule Linear Transformers","authors":["Alek Sobczyk","Gioele Gottardo","Christos K Matzoros","Mirko De Vita","Filip Skogh","Anastasios Zouzias","Jiawei Zhuang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The triangular inverse factors double the identity input","plainExplanation":"Polynomial conditioning does not ensure the claimed scalar-initialization norm contraction. The claimed norm-based application is unsupported. The witness does not show Newton–Schulz divergence or realization by every DeltaNet architecture.","primaryUrl":"https://arxiv.org/html/2605.21325","noteUrl":"/review?paper=review-neurips-2026-150772","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150773","title":"Wavelet Flow Matching for Multi-Scale Physics Emulation","authors":["Gabriele Accarino","Juan Nathaniel","Carla Roesch","Pierre Gentine","Sara Shamekh","Duncan Watson-Parris","Viviana Acquaviva"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Wavelet Flow Matching for Multi-Scale Physics Emulation","plainExplanation":"Retained approximation bands are jointly dependent across wavelet scales, a true predictive distribution can have positive expected CRPS, and the written energy closure mixes internal with total energy. The generic weighted-regression example does not establish an error under complete deterministic PDE conditioning.","primaryUrl":"https://arxiv.org/html/2605.16573","noteUrl":"/review?paper=review-neurips-2026-150773","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","false_statement","formula_error","interpretation_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150774","title":"PaLoRA: Paced Low-Rank Adaptation for Continual Learning","authors":["Yuxuan Li","Fanhu Zeng","Hao Tang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PaLoRA: Paced Low-Rank Adaptation for Continual Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04226","noteUrl":"/review?paper=review-neurips-2026-150774","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150775","title":"PatchKV: Weight Space Compensation of KV Cache","authors":["Chanryeol Lee","Chanhyuk Lee","Yeonwoo Choi","Donggyun Kim","Seunghoon Hong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PatchKV: Weight Space Compensation of KV Cache","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39329","noteUrl":"/review?paper=review-neurips-2026-150775","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150776","title":"UnlearningSoup: Is Repeated Tuning Necessary for Large Language Model Unlearning?","authors":["Puning Yang","Qizhou Wang","Junchi Yu","Bo Han","Xiuying Chen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"UnlearningSoup: Is Repeated Tuning Necessary for Large Language Model Unlearning?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.37076","noteUrl":"/review?paper=review-neurips-2026-150776","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150777","title":"MobileMoE: Scaling On-Device Mixture of Experts","authors":["Yanbei Chen","Hanxian Huang","Ernie Chang","Jacob Szwejbka","Digant Desai","Zechun Liu","Vikas Chandra","Raghuraman Krishnamoorthi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Resource invariance does not preserve the loss-optimal expert count","plainExplanation":"Resource invariance does not preserve the loss-optimal expert count. Resource decoupling alone does not establish joint optimality; the empirical recommendation may still be useful.","primaryUrl":"https://arxiv.org/html/2605.27358","noteUrl":"/review?paper=review-neurips-2026-150777","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150778","title":"EgoForce: Robust Online Egocentric Motion Reconstruction via Diffusion Forcing","authors":["Inwoo Hwang","Donggeun Lim","Hojun Jang","Young Min Kim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"EgoForce: Robust Online Egocentric Motion Reconstruction via Diffusion Forcing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13041","noteUrl":"/review?paper=review-neurips-2026-150778","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150780","title":"AVSD: Adaptive-View Self-Distillation by Balancing Consensus and Teacher-Specific Privileged Signals","authors":["Duy Nguyen","Hanqi Xiao","Archiki Prasad","Zaid Khan","Anirban Das","Shi-Xiong Zhang","Sambit Sahu","Hyunji Lee","Elias Stengel-Eskin","Mohit Bansal"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AVSD: Adaptive-View Self-Distillation by Balancing Consensus and Teacher-Specific Privileged Signals","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20643","noteUrl":"/review?paper=review-neurips-2026-150780","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150783","title":"SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation","authors":["JungHun Oh","Sungyong Baik","Kyoung Mu Lee"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.16454","noteUrl":"/review?paper=review-neurips-2026-150783","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150785","title":"QMaxCal: Path-Space Regularization for Open Quantum Control via Girsanov's Theorem","authors":["Merijn Moody","Zier Mensch","Miranda Cheng","Peter G Bolhuis","Max Welling"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"QMaxCal: Path-Space Regularization for Open Quantum Control via Girsanov's Theorem","plainExplanation":"Constant measured homodyne drift does not force decoherence-free evolution for general complex monitored operators. In the STIRAP example, the drift measures a real coherence quadrature rather than excited-state population alone.","primaryUrl":"https://arxiv.org/html/2606.19947","noteUrl":"/review?paper=review-neurips-2026-150785","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","interpretation_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150786","title":"Rebellious Student: Reversing Teacher Signals for Reasoning Exploration with Self-Distilled RLVR","authors":["Jeonghye Kim","Jiwon Jeon","Dongsheng Li","Yuqing Yang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The valid influence bound has no converse for large reverse KL","plainExplanation":"A correct reference solution need not instantiate conditioning on the current rollout’s success. The ideal-to-practical guarantee needs an extra premise. This does not show that the actual context-sensitive teacher ignores solutions.","primaryUrl":"https://arxiv.org/html/2605.10781","noteUrl":"/review?paper=review-neurips-2026-150786","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150787","title":"Co-PiLOT: Constrained Physics-Informed Latent Optimization for Target-Driven Inverse Design","authors":["Mahish Kumar Guru","Mayank Nagar","Ayush vyas","Jan Bohlen","Roland Aydin","Noomane B Khalifa"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Co-PiLOT: Constrained Physics-Informed Latent Optimization for Target-Driven Inverse Design","plainExplanation":"The quaternion code’s physical rotation bound misses a factor of two. Multiplying a potentially negative log-improvement score by feasibility also reverses the intended gate; the author repository’s one-candidate acquisition already adds log feasibility.","primaryUrl":"https://arxiv.org/html/2609.37875","noteUrl":"/review?paper=review-neurips-2026-150787","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error","incorrect_bound","numerical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150788","title":"Beyond Spatial-Domain Supervision: A Relation Constrained Space for Multi-Modal Image Fusion","authors":["Zeyu Wang","Mingyu Ge","Haiyu Song","Haoran Duan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The main text and appendix normalize different reliability weights","plainExplanation":"The main text and appendix normalize different reliability weights. The written objectives disagree; neither formula identifies which version was implemented.","primaryUrl":"https://arxiv.org/html/2609.38968","noteUrl":"/review?paper=review-neurips-2026-150788","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150791","title":"Uncertainty Quantification for Large Language Diffusion Models","authors":["Artem Vazhentsev","Vladislav Smirnov","David Li","Maxim Panov","Timothy Baldwin","Artem Shelmanov"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The negative-log-likelihood score is missing its minus sign","plainExplanation":"The negative-log-likelihood score is missing its minus sign. This is a local score-definition error, rather than evidence about the reported implementation or tables.","primaryUrl":"https://arxiv.org/html/2605.14570","noteUrl":"/review?paper=review-neurips-2026-150791","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150794","title":"Who Needs Labels? Adapting Vision Foundation Models With the Metadata You Already Have","authors":["Elouan Gardes","Seung-Eun Yi","Huy V. Vo","Théo Moutakanni","Kartik Ahuja","Piotr Bojanowski","Wolfgang Pernice","Loic Landrieu","Camille Couprie"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Who Needs Labels? Adapting Vision Foundation Models With the Metadata You Already Have","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05107","noteUrl":"/review?paper=review-neurips-2026-150794","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150795","title":"Conformal Language Modeling via Posterior Sampling","authors":["Nicolas Emmenegger","Theo X. Olausson","Armando Solar-Lezama","Chara Podimata"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Conformal Language Modeling via Posterior Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03731","noteUrl":"/review?paper=review-neurips-2026-150795","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150796","title":"Prediction-Augmented Trees for Reliable Statistical Inference","authors":["Vikram Kher","Argyris Oikonomou","Manolis Zampetakis"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The regression variance estimate removes genuine score variation","plainExplanation":"Centering the score and removing a fitted regression effect are different operations. The displayed residual sandwich can remove all limiting residual variation while the score mean being estimated still has positive sampling variance. This occurs in one fixed leaf under an exact linear response model. The contradiction concerns the displayed regression variance estimator and the Wald intervals using it. The score central limit theorem and the population-mean estimator are not contradicted by this example.","primaryUrl":"https://arxiv.org/html/2510.16937","noteUrl":"/review?paper=review-neurips-2026-150796","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150800","title":"From Non-Convex to Strongly Convex: Curvature-Adaptive FTPL for Online Optimization","authors":["Moses Charikar","Chirag Pabbaraju","Ambuj Tewari"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"From Non-Convex to Strongly Convex: Curvature-Adaptive FTPL for Online Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02948","noteUrl":"/review?paper=review-neurips-2026-150800","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150802","title":"FrequencyBooster: Advancing Pixel Diffusion for High-Fidelity Image Generation","authors":["Lichen Ma","Zipeng Guo","Yu He","Xiaolong Fu","Luohang Liu","Jingling Fu","Junshi Huang","Yan Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"FrequencyBooster: Advancing Pixel Diffusion for High-Fidelity Image Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.17759v2","noteUrl":"/review?paper=review-neurips-2026-150802","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150804","title":"It Just Takes Two: Scaling Amortized Inference to Large Sets","authors":["Antoine Wehenkel","Michael Kagan","Lukas Heinrich","Chris Pollard"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"It Just Takes Two: Scaling Amortized Inference to Large Sets","plainExplanation":"Shared-nuisance marginal sufficiency does not force one set-size-independent exponential-family form. Optimal encoders can also retain nonlinear ancillary coordinates while recovering the same posterior, so common affine recovery need not be invertible.","primaryUrl":"https://arxiv.org/html/2605.07972","noteUrl":"/review?paper=review-neurips-2026-150804","collection":"neurips-2026","severity":"repairable","issueTypes":["citation_hypothesis_mismatch","false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150805","title":"The parameters in weight-sparse transformers are interpretable","authors":["Arnau Marin-Llobet","Stefan Heimersheim"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"The parameters in weight-sparse transformers are interpretable","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.02964","noteUrl":"/review?paper=review-neurips-2026-150805","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150806","title":"Mechanism Design for AI Overviews: Creator Incentives and Long-Term Profit","authors":["Yihang Wu","Jiajun Tang","Jinfei Liu","Haifeng Xu","Fan Yao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mechanism Design for AI Overviews: Creator Incentives and Long-Term Profit","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.22493","noteUrl":"/review?paper=review-neurips-2026-150806","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150808","title":"Contrastive Distribution Matching for Amortized Sequential Monte Carlo in Discrete Diffusion","authors":["Jaihoon Kim","Taehoon Yoon","Prin Phunyaphibarn","Seungjun Kim","Morteza Mardani","Minhyuk Sung"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Contrastive Distribution Matching for Amortized Sequential Monte Carlo in Discrete Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23346","noteUrl":"/review?paper=review-neurips-2026-150808","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150810","title":"Machine Unlearning in Low-Dimensional Feature Subspace","authors":["Kun Fang","Qinghua Tao","Junxu Liu","Yaxin Xiao","Yulin Jin","Qingqing Ye","Jian Sun","Haibo Hu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Machine Unlearning in Low-Dimensional Feature Subspace","plainExplanation":"No issue was found in the checked claims. The revision supplies the centering premise and weakens the conclusion to the expectation controlled by the covariance trace. The former constant-offset and rare-outlier counterexamples no longer violate the revised statement.","primaryUrl":"https://arxiv.org/html/2601.22456","noteUrl":"/review?paper=review-neurips-2026-150810","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150814","title":"Hyperparameter Transfer for Dense Associative Memories","authors":["Roi Holtzman","Dmitry Krotov","Boris Hanin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The displayed global energy omits its interaction term","plainExplanation":"The displayed global energy omits its interaction term. The global-energy display is incomplete; the later cancellation indicates a local omission rather than a demonstrated failure of the intended dynamics.","primaryUrl":"https://arxiv.org/html/2605.10164","noteUrl":"/review?paper=review-neurips-2026-150814","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150815","title":"A Communication-Theoretic Framework for LLM Agents: Cost-Aware Adaptive Reliability","authors":["Hamed Omidvar","Vahideh Akhlaghi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A Communication-Theoretic Framework for LLM Agents: Cost-Aware Adaptive Reliability","plainExplanation":"Independent channel randomness does not make information about a shared input additive. Under conditional independence, output redundancy must be subtracted from the sum of individual mutual informations.","primaryUrl":"https://arxiv.org/html/2605.09121","noteUrl":"/review?paper=review-neurips-2026-150815","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150817","title":"Smart Picks in the Dark: Towards Efficient RLVR for Reasoning via Tracing Metacognitive Pivots","authors":["Guangcheng Zhu","Shenzhi Yang","Haobo Wang","Xing Zheng","Yingfan Ma","Xuening Feng","Zhongqi Chen","Bowen Song","Weiqiang Wang","Gang Chen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Smart Picks in the Dark: Towards Efficient RLVR for Reasoning via Tracing Metacognitive Pivots","plainExplanation":"The Fisher approximation gives a local quadratic surrogate optimum, not an exact finite-KL improvement bound. A Bernoulli update violates the stated bound; binary-KL inversion supplies the sharp constrained alternative.","primaryUrl":"https://arxiv.org/html/2606.04503","noteUrl":"/review?paper=review-neurips-2026-150817","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150820","title":"Foresight-over-Graph: Reasoning Beyond Local Horizons for Knowledge Base Question Answering","authors":["Yang Hong","Yajun Yang","Xin Wang","Liping Jing","Qinghua Hu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Foresight-over-Graph: Reasoning Beyond Local Horizons for Knowledge Base Question Answering","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.08388","noteUrl":"/review?paper=review-neurips-2026-150820","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150824","title":"ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening","authors":["Antoine de Mathelin","Christopher Tosh","Wesley Tansey"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.12219","noteUrl":"/review?paper=review-neurips-2026-150824","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150825","title":"Paradoxes of Game Theoretic Equilibria and Price of Anarchy","authors":["Georgios Piliouras","Ian Gemp","Siqi Liu","Luke Marris"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Paradoxes of Game Theoretic Equilibria and Price of Anarchy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.11752","noteUrl":"/review?paper=review-neurips-2026-150825","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150828","title":"SOC-ICNN: From Polyhedral to Conic Geometry for Learning Convex Surrogate Functions","authors":["Kang Liu","Jianchen Hu","Wei Peng"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"SOC-ICNN: From Polyhedral to Conic Geometry for Learning Convex Surrogate Functions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.22355","noteUrl":"/review?paper=review-neurips-2026-150828","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150829","title":"Diffusion Subgoal Planning for Long-Horizon Offline Goal-Conditioned Reinforcement Learning","authors":["Hengrui Zhang","Yuhu Cheng","C.L.Philip Chen","Xuesong Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The likelihood interpretation needs to stay distinct from discounted return","plainExplanation":"The retained issue is the identification with the return-based advantage. It is not an error in Bayes’ rule or in the monotonicity of the logarithm for the positive relabeling quantity defined in the appendix.","primaryUrl":"https://arxiv.org/html/2609.34575","noteUrl":"/review?paper=review-neurips-2026-150829","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["interpretation_error","missing_assumption"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-150832","title":"A Few Teacher Steps Go a Long Way: Cost-Efficient On-Policy Data Augmentation for Agent Post-Training","authors":["Junze Ye","Jiayi Cheng","Miao Lu","Michal Mankowski","Jose Blanchet","Mohsen Bayati"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Few Teacher Steps Go a Long Way: Cost-Efficient On-Policy Data Augmentation for Agent Post-Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.04574","noteUrl":"/review?paper=review-neurips-2026-150832","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150835","title":"Signal-Adaptive Trust Regions for Gradient-Free Optimization of Recurrent Spiking Neural Networks","authors":["Jinhao Li","Yuhao Sun","Zhiyuan Ma","Hao He","Xinche Zhang","Xing Chen","Li Jin","Sen Song"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Signal-Adaptive Trust Regions for Gradient-Free Optimization of Recurrent Spiking Neural Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.21572","noteUrl":"/review?paper=review-neurips-2026-150835","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150839","title":"Exemplar2VQA: A Scalable Exemplar-Driven Visual Question Answering Generation Framework via Multi-Agent Coding","authors":["Jiayu Ying","Qijian Tian","Ruijie Xu","Xinnan Zhu","Daoguo Dong","Jiachen Xu","Xin Tan"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Exemplar2VQA: A Scalable Exemplar-Driven Visual Question Answering Generation Framework via Multi-Agent Coding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.37655","noteUrl":"/review?paper=review-neurips-2026-150839","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150840","title":"Geometry-Aware Similarity Metrics for Neural Representations on Riemannian and Statistical Manifolds","authors":["N Alex Cayco Gajic","Arthur Pellegrino"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Left whitening does not justify Euclidean Rayleigh bounds","plainExplanation":"Left whitening does not justify Euclidean Rayleigh bounds. The displayed proof step is invalid; Proposition 2.2 admits the congruence repair.","primaryUrl":"https://arxiv.org/html/2603.28764","noteUrl":"/review?paper=review-neurips-2026-150840","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150842","title":"The Multi-Block DC Function Class: Theory, Algorithms, and Applications","authors":["Sayedpouria Fatemi","Hoomaan Maskan","Alp Yurtsever","Suvrit Sra"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Multi-Block DC Function Class: Theory, Algorithms, and Applications","plainExplanation":"The algorithm can remain at a point that is critical in each coordinate but is not Clarke stationary. The stated rates require additional compatibility between block subgradient choices.","primaryUrl":"https://arxiv.org/pdf/2604.17560v1","noteUrl":"/review?paper=review-neurips-2026-150842","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-150843","title":"PULSE: Identifying Demonstration-Utility Features with Sparse Autoencoders","authors":["Chenduo Hao","Chuanbao Gao","Pinjun Zeng","Jingze Zhu","Liu Chonghan","Zidong Liu","Xu Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PULSE: Identifying Demonstration-Utility Features with Sparse Autoencoders","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.32469","noteUrl":"/review?paper=review-neurips-2026-150843","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150846","title":"SCALECUA : Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL","authors":["Bowen Lv","Xiao Liu","Yanyu Ren","Hanyu Lai","Bohao Jing","Hanchen Zhang","Yanxiao Zhao","顺天 姚","Jie Tang","Yuxiao Dong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SCALECUA : Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.11185","noteUrl":"/review?paper=review-neurips-2026-150846","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150847","title":"Sample-Efficient Optimization over Generative Priors via Coarse Learnability","authors":["Pranjal Awasthi","Sreenivas Gollapudi","Ravi Kumar","Kamesh Munagala"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Sample-Efficient Optimization over Generative Priors via Coarse Learnability","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2503.06917","noteUrl":"/review?paper=review-neurips-2026-150847","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150849","title":"Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation","authors":["Hugues Van Assel","Edward De Brouwer","Saeed Saremi","Gabriele Scalia","Aviv Regev"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The coupling recovery mixes two Sinkhorn normalization conventions","plainExplanation":"The coupling recovery mixes two Sinkhorn normalization conventions. The printed algorithm does not return the claimed marginals; this does not identify the implementation convention.","primaryUrl":"https://arxiv.org/html/2606.00514","noteUrl":"/review?paper=review-neurips-2026-150849","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150851","title":"WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation","authors":["Baining Zhao","jiacheng xu","Weicheng Feng","Xin Zhang","Zhaolu Wang","Haoyang Wang","Shilong Ji","Ziyou Wang","Jianjie Fang","Zhiheng Zheng","Weichen Zhang","Yu Shang","Wei Wu","Chen Gao","Xinlei Chen","Yong Li"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15964","noteUrl":"/review?paper=review-neurips-2026-150851","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150857","title":"Load Balancing Mixture of Experts with Similarity Preserving Routers","authors":["Nabil Omi","Siddhartha Sen","Ali Farhadi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Orthonormal columns preserve projected rather than arbitrary token geometry","plainExplanation":"Routed dot products are \\(xRR^\\top y^\\top\\). With \\(E<D\\), \\(RR^\\top\\) is a rank-\\(E\\) projector, not \\(I_D\\), so components in the router’s nullspace disappear. Column orthonormality preserves coefficient geometry within the column space, not arbitrary input geometry. The global similarity-preservation rationale fails. Observed load-balancing improvements are not refuted.","primaryUrl":"https://arxiv.org/html/2506.14038","noteUrl":"/review?paper=review-neurips-2026-150857","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150860","title":"Baba in Wonderland: Online Self-Supervised Dynamics Discovery for Executable World Models","authors":["SeungWon Seo","DongHeun Han","SeongRae Noh","Hyeongyeop Kang"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Baba in Wonderland: Online Self-Supervised Dynamics Discovery for Executable World Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16725","noteUrl":"/review?paper=review-neurips-2026-150860","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150862","title":"Advancing Entropy-Level Credit Assignment in RLVR via Proximal Entropy Policy Optimization","authors":["Yun Kim","Nojun Kwak"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Advancing Entropy-Level Credit Assignment in RLVR via Proximal Entropy Policy Optimization","plainExplanation":"Smooth monotonicity alone does not impose an inverse-square sequence-length curvature rate. Ordinary symmetric window averaging also preserves an affine positional trend; detrending requires subtraction or a different operator.","primaryUrl":"https://arxiv.org/html/2609.39402","noteUrl":"/review?paper=review-neurips-2026-150862","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","interpretation_error","invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150865","title":"Test-Time Speculation","authors":["Avinash Kumar","Sujay Sanghavi","Poulami Das"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A self-reference KL has zero gradient in the specified single step","plainExplanation":"At fixed context, the logit gradient of \\(D_{\\mathrm{KL}}(q_t\\|q_\\theta)\\) is \\(q_\\theta-q_t\\), hence zero at the starting policy \\(q_\\theta=q_t\\). All weighted contexts therefore contribute zero through parameter sharing as well. The coefficient of this term cannot affect that single evaluated step. The local single-step regularization claim is false. A different actual implementation or empirical effect is not inferred.","primaryUrl":"https://arxiv.org/html/2605.09329","noteUrl":"/review?paper=review-neurips-2026-150865","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150868","title":"AdaState: Self-Evolving Anchors for Streaming Video Generation","authors":["Yusuf Dalva","Pinar Yanardag"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"AdaState: Self-Evolving Anchors for Streaming Video Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30349","noteUrl":"/review?paper=review-neurips-2026-150868","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150872","title":"Falcon-X: A Time Series Foundation Model for Heterogeneous Multivariate Modeling","authors":["Yiding Liu","Yifan Hu","Hongjie Xia","Peiyuan Liu","Hongzhou Chen","Xilin Dai","Zewei Dong","Jiang-Ming Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Standardization does not put observations in the real arcsine domain","plainExplanation":"For observations \\((0,0,3)\\), the largest standardized value exceeds one under either sample or population standard deviation. Its real arcsine is undefined. The proposed inverse \\(\\sigma\\sin(\\cdot)+\\mu\\) is also bounded above by \\(\\mu+\\sigma\\), so it cannot invert all standardized observations without an additional bounded map. The printed real-valued pipeline has a domain defect. Deployed code behavior remains unknown.","primaryUrl":"https://arxiv.org/html/2605.27286","noteUrl":"/review?paper=review-neurips-2026-150872","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150874","title":"Mitigating Object Hallucination in Large Vision-Language Models via False Discovery Controlled Visual Data Splitting","authors":["Chang Liu","Yu Tian","Rui Xie"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Mitigating Object Hallucination in Large Vision-Language Models via False Discovery Controlled Visual Data Splitting","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.38979","noteUrl":"/review?paper=review-neurips-2026-150874","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150875","title":"Optimization Dynamics Imprint Semantic Specificity in Contrastive Embedding Norms","authors":["Ziwei Su","Junyu Ren","Victor Veitch"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Optimization Dynamics Imprint Semantic Specificity in Contrastive Embedding Norms","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30625","noteUrl":"/review?paper=review-neurips-2026-150875","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150878","title":"How Data Augmentation Shapes Neural Representations","authors":["Tianxiao He","Alex Williams","Sarah Harvey"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The zero-augmentation limit need not equal the cross-seed distance","plainExplanation":"The zero-augmentation limit need not equal the cross-seed distance. The asserted equality is false; empirical comparisons between the two scales are not invalidated.","primaryUrl":"https://arxiv.org/html/2605.15306","noteUrl":"/review?paper=review-neurips-2026-150878","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150882","title":"Sequential Behavioral Watermarking for LLM Agents","authors":["Hyeseon An","Shinwoo Park","Dongsu Kim","Yo-Sub Han"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The mean-shift coefficient depends on the action probabilities","plainExplanation":"The expected first-order hit increase depends on the elicited action probabilities, and reusing a context can prevent score variability from decreasing with sequence length. These corrections leave the exchangeable-key p-value argument intact. The KL expansion explicitly assumes a uniform action law; the normalized probability floor preserves that law, so the nonuniform-floor objection does not refute the stated expansion.","primaryUrl":"https://arxiv.org/html/2605.11036","noteUrl":"/review?paper=review-neurips-2026-150882","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150883","title":"Guaranteed Noisy CP Tensor Recovery via Riemannian Optimization on the Segre Manifold","authors":["Ke Xu","Yuefeng Han"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Guaranteed Noisy CP Tensor Recovery via Riemannian Optimization on the Segre Manifold","plainExplanation":"The tensor analysis loses a power of cross-overlap in its tangent estimate, normalizes the quadratic error rate incorrectly at iteration zero, and replaces a restricted Gram inverse with a projected ambient inverse. The supplied local corrections do not by themselves establish the full noisy convergence analysis.","primaryUrl":"https://arxiv.org/html/2510.00569","noteUrl":"/review?paper=review-neurips-2026-150883","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","invalid_inference","lemma_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150884","title":"CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery","authors":["Zi-Rong Li","Si-Yang Liu","Tian-Zuo Wang","Han-Jia Ye"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31156","noteUrl":"/review?paper=review-neurips-2026-150884","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150886","title":"i-DEQ: A stable inertial Deep Equilibrium model for image restoration","authors":["Antonin Clerc","Marien Renaud","Baudouin Denis de Senneville","Nicolas Papadakis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"i-DEQ: A stable inertial Deep Equilibrium model for image restoration","plainExplanation":"The MRI proximal denominator belongs in Fourier coordinates, and an inertial state derivative need not vanish at equilibrium. The accelerated stationarity rate also requires the cited theorem’s coupled parameter schedule and averaged output convention.","primaryUrl":"https://arxiv.org/html/2605.19705","noteUrl":"/review?paper=review-neurips-2026-150886","collection":"neurips-2026","severity":"repairable","issueTypes":["citation_hypothesis_mismatch","false_statement","formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150887","title":"Learning Polyhedral Conformal Sets for Robust Optimization","authors":["Shuyi Chen","Wenbin Zhou","Shixiang Zhu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Learning Polyhedral Conformal Sets for Robust Optimization","plainExplanation":"A local density bound needs a localization condition and conformal rank correction before it controls quantile error. The optimization procedures also need a lower-bounded initial master and derivatives that account for movement of the learned uncertainty set.","primaryUrl":"https://arxiv.org/html/2605.08506","noteUrl":"/review?paper=review-neurips-2026-150887","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error","incorrect_bound","invalid_inference","mathematical_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150891","title":"PRIM: Meta-Learned Bayesian Root Cause Analysis","authors":["Christopher Lohse","Anish Dhir","Amadou Ba","Bradley Eck","Marco Ruffini","Jonas Wahl"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"PRIM: Meta-Learned Bayesian Root Cause Analysis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08786","noteUrl":"/review?paper=review-neurips-2026-150891","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150892","title":"LOTION: Smoothing the Optimization Landscape for Quantized Training","authors":["Mujin Kwun","Depen Morwani","Huangyuan Su","Stephanie Gil","Nikhil Anand","Sham Kakade"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A fixed finite codebook cannot be unbiased over all real weights","plainExplanation":"A fixed finite codebook cannot be unbiased over all real weights. The all-real-input unbiasedness setup is inconsistent; this is not evidence of an experimental quantizer failure.","primaryUrl":"https://arxiv.org/html/2510.08757","noteUrl":"/review?paper=review-neurips-2026-150892","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150893","title":"SynIB: Informational Bottleneck for Maximizing Synergy in Multimodal Learning","authors":["Konstantinos Kontras","Teodora Popordanoska","Thomas Strypsteen","Christos Chatzichristos","Matthew Blaschko","Maarten De Vos","Paul Liang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SynIB: Informational Bottleneck for Maximizing Synergy in Multimodal Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.09853","noteUrl":"/review?paper=review-neurips-2026-150893","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150894","title":"DSAQuant: Denoising-Stage-Aligned Quantization-Aware Training for Video Generation","authors":["Shuaiting Li","Zelin Gao","Haibin Shen","Yujun Shen","Haotong Qin","Yinghao Xu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DSAQuant: Denoising-Stage-Aligned Quantization-Aware Training for Video Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.04031","noteUrl":"/review?paper=review-neurips-2026-150894","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150895","title":"CORP: Closed-Form One-shot Representation-Preserving Structured Pruning for Transformers","authors":["Boxiang Zhang","Baijian Yang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"CORP: Closed-Form One-shot Representation-Preserving Structured Pruning for Transformers","plainExplanation":"The residual covariance has a matrix-order error already avoided by the paper’s direct expansion. Its calibration bound also omits fixed-ridge bias and conditioning-sensitive drift, while zero aggregate attention gain allows cancellation across samples.","primaryUrl":"https://arxiv.org/html/2602.05243","noteUrl":"/review?paper=review-neurips-2026-150895","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150897","title":"SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference","authors":["Yaosheng Fu","Guangxuan Xiao","Xin Dong","Song Han","Oreste Villa"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04511","noteUrl":"/review?paper=review-neurips-2026-150897","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150898","title":"Understanding Goal Generalisation in Sequential Reinforcement Learning","authors":["Jason Brown","Edward Young"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Understanding Goal Generalisation in Sequential Reinforcement Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.23565","noteUrl":"/review?paper=review-neurips-2026-150898","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150901","title":"Privacy by Postprocessing the Discrete Laplace Mechanism","authors":["Quentin Hillebrand","Jacob Imola","Rasmus Pagh","Sia Sejer"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The partition-function correction scales finite variance downward","plainExplanation":"The partition-function correction scales finite variance downward. This concerns this partition-function estimator, not a universal variance claim about every debiasing problem.","primaryUrl":"https://arxiv.org/html/2605.06502","noteUrl":"/review?paper=review-neurips-2026-150901","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150903","title":"Joint Treatment Effect Estimation from Incomplete Healthcare Data: Temporal Causal Normalizing Flows with LLM-driven Evolutionary MNAR Imputation","authors":["Olivia Jullian Parra","Sara Zoccheddu","David C Cerezo","Tom Forzy","Franziska S Ulrich","William Sutcliffe","Jakob M Burgstaller","Oliver Senn","Patrick Owen","Nicola Serra"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Joint Treatment Effect Estimation from Incomplete Healthcare Data: Temporal Causal Normalizing Flows with LLM-driven Evolutionary MNAR Imputation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.05125","noteUrl":"/review?paper=review-neurips-2026-150903","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150906","title":"Ripple in Still Water: Zero-Shot Clustering in Heterogeneous Federated Learning with Wavelet Scattering Transform","authors":["Alessandro Licciardi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The convergence application and excess-risk corollary need stronger hypotheses","plainExplanation":"This is a Polyak–Łojasiewicz inequality, not a consequence of smoothness. Randomly selecting an iterate converts an average squared-gradient bound into expected stationarity; it does not turn stationarity into optimality.","primaryUrl":"https://arxiv.org/html/2610.03054","noteUrl":"/review?paper=review-neurips-2026-150906","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150907","title":"EpicWorldModel: Exploration-driven Planning with Latent World Models","authors":["Bowen Feng","Julian Ost","Zhiting Mei","Anirudha Majumdar","Felix Heide"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EpicWorldModel: Exploration-driven Planning with Latent World Models","plainExplanation":"The covariance, finite-ensemble uncertainty, information and planning claims concern different quantities. Their repairs require explicit model sensitivity, predictive-law assumptions, observation resolution or a consistent objective.","primaryUrl":"https://arxiv.org/html/2610.05996","noteUrl":"/review?paper=review-neurips-2026-150907","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference"],"classifications":["major"]},{"slug":"review-neurips-2026-150908","title":"Conceal, Reconstruct, Jailbreak: Exploiting the Reconstruction--Concealment Tradeoff in MLLMs","authors":["Md Farhamdur Reza","Richeng Jin","Tianfu Wu","Huaiyu Dai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Conceal, Reconstruct, Jailbreak: Exploiting the Reconstruction--Concealment Tradeoff in MLLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05709","noteUrl":"/review?paper=review-neurips-2026-150908","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150909","title":"Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders","authors":["William Lehn-Schiøler","Magnus R Kjaer","Rahul Thapa","Magnus G Pedersen","Anton M Storgaard","Nick Williams","Andreas Brink-Kjaer","Tue Lehn-schiøler","Sadasivan Puthusserypady","James Zou","Lars Kai Hansen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13930v3","noteUrl":"/review?paper=review-neurips-2026-150909","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150910","title":"Entangled Schrödinger Bridge Matching","authors":["Sophia Tang","Yinuo Zhang","Pranam Chatterjee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Entangled Schrödinger Bridge Matching","plainExplanation":"The stated endpoint tilt and constrained control family do not solve the general bridge problem. The displayed force cost, density potential and replay probabilities also need distinct formula corrections.","primaryUrl":"https://arxiv.org/html/2511.07406","noteUrl":"/review?paper=review-neurips-2026-150910","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","invalid_inference","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-150911","title":"GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction","authors":["Katharina Sophie Schmid","Nicolas von Lützow","Angela Dai","Jozef Hladký","Matthias Niessner"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23888","noteUrl":"/review?paper=review-neurips-2026-150911","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150912","title":"Agentic Abstention: Do Agents Know When to Stop Instead of Act?","authors":["Han Luo","Bingbing Wen","Lucy Lu Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Absolute-step recall and timely recall are different events","plainExplanation":"Absolute-step recall and timely recall are different events. The metric definitions are incompatible; this does not determine the actual benchmark implementation.","primaryUrl":"https://arxiv.org/html/2606.28733","noteUrl":"/review?paper=review-neurips-2026-150912","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150914","title":"Forced Deferral: Manipulating Routing Decisions in Multimodal LLM Cascades","authors":["Zhongye Liu","Yaopei Zeng","Yurui Chang","Lu Lin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Forced Deferral: Manipulating Routing Decisions in Multimodal LLM Cascades","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.15308","noteUrl":"/review?paper=review-neurips-2026-150914","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150917","title":"From Denoising to Refining: A Corrective Framework for Vision-Language Diffusion Model","authors":["Yatai Ji","Teng Wang","Yuying Ge","Zhiheng Liu","Sidi Yang","Ying Shan","Ping Luo"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The clean-token objective can have infinite expectation near time zero","plainExplanation":"The clean-token objective can have infinite expectation near time zero. The continuous printed objective has an integrability gap for ordinary imperfect predictors; training-code failure is not established.","primaryUrl":"https://arxiv.org/html/2510.19871","noteUrl":"/review?paper=review-neurips-2026-150917","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150918","title":"COMPOT: Calibration-Optimized Matrix Procrustes Orthogonalization for Transformers Compression","authors":["Denis Makhov","Dmitriy Shopkhoev","Magauiya Zhussip","Ammar Ali","Baher Mohammad","Stamatios Lefkimmiatis"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"COMPOT: Calibration-Optimized Matrix Procrustes Orthogonalization for Transformers Compression","plainExplanation":"No issue was found in the checked claims. The source no longer claims that a pooled singular-value order solves the heterogeneous-storage knapsack problem. The earlier counterexample addressed precisely that removed optimality claim.","primaryUrl":"https://arxiv.org/html/2602.15200","noteUrl":"/review?paper=review-neurips-2026-150918","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150919","title":"SR-GRPO: Stable Rank as an Intrinsic Geometric Reward for Large Language Model Alignment","authors":["Yixuan Tang","Yi Yang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"SR-GRPO: Stable Rank as an Intrinsic Geometric Reward for Large Language Model Alignment","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.02807","noteUrl":"/review?paper=review-neurips-2026-150919","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150922","title":"Towards Scalable Data Diversification for Language Model Pretraining via Leverage Score Sampling","authors":["Zailin Ma","Quzhe Huang","Yujun Li","Congyuan Rao","Yaodong Yang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Normalized regularized insertion changes the determinant-ranking matrix","plainExplanation":"Normalized regularized insertion changes the determinant-ranking matrix. The determinant lemma itself is correct. Its exact actual-pool insertion interpretation needs qualification.","primaryUrl":"https://arxiv.org/html/2609.32484","noteUrl":"/review?paper=review-neurips-2026-150922","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150926","title":"Quest: Training Frontier Deep Research Agents with Fully Synthetic Tasks","authors":["Jian Xie","Tianhe Lin","Zilu Wang","Yuting Ning","Yuekun Yao","Tianci Xue","Zhehao Zhang","Zhongyang Li","Kai Zhang","Yufan Wu","Shijie Chen","Boyu Gou","Mingzhe Han","Yifei Wang","Vint Lee","Xinpeng Wei","XJ Wang","Yu Su","Huan Sun"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Quest: Training Frontier Deep Research Agents with Fully Synthetic Tasks","plainExplanation":"No issue was found in the checked claims. The retained missing-boundary issue is resolved by the inclusive top interval.","primaryUrl":"https://arxiv.org/html/2605.24218","noteUrl":"/review?paper=review-neurips-2026-150926","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150928","title":"AETDICE: Unified Framework and Offline Optimization for Nonlinear Multi-Objective RL","authors":["Woosung Kim","Youngjun Suh","Jinho Lee","Jongmin Lee","Byung-Jun Lee"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Telescoping requires every prefix transform to be finite and defined","plainExplanation":"The appendix reward sum omits the initial transform offset. The printed algorithm optimizes a shifted objective; Proposition 4.1 under its main-text convention is not refuted.","primaryUrl":"https://arxiv.org/html/2606.31178","noteUrl":"/review?paper=review-neurips-2026-150928","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150929","title":"Beyond Pairs: Your Language Model is Secretly Optimizing a Preference Graph","authors":["Ning Liu","Chuanneng Sun","Kristina L Klinkner","Shervin Malmasi"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Beyond Pairs: Your Language Model is Secretly Optimizing a Preference Graph","plainExplanation":"Response-dependent length normalization can make the expected KL surrogate negative. The graph-score gradient also needs the node’s own loss, and preference ties alone do not make policy-dependent edge weights uniform.","primaryUrl":"https://arxiv.org/html/2605.08037","noteUrl":"/review?paper=review-neurips-2026-150929","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150932","title":"CASPIAN: Online Detection and Attribution of Cascade Attacks in LLM Multi-Agent Systems via Cross-Channel Causal Monitoring","authors":["Kavana Venkatesh","Jafar Isbarov","Saad Amin","Murat Kantarcioglu","Jiaming Cui"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The phase-shift rule loses dependence on positive growth magnitude","plainExplanation":"The phase-shift rule loses dependence on positive growth magnitude. The printed confirmation rule is nonselective in the ordinary positive-growth interior; other detector gates can still reject events.","primaryUrl":"https://arxiv.org/html/2605.19240","noteUrl":"/review?paper=review-neurips-2026-150932","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150933","title":"Adaptive Prior Selection in Gaussian Process Bandits with Thompson Sampling","authors":["Jack Sandberg","Morteza Haghir Chehreghani"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The posterior-mean bound omits the additive prior mean","plainExplanation":"A GP with constant mean \\(10\\) and zero covariance is identically \\(10\\), and every posterior mean is \\(10\\), regardless of observation noise. With noise standard deviation \\(100\\) and \\(T=2\\), equation (117)’s right side is \\(\\sqrt2(10+10+100)/100<10\\). The analogous squared bound also fails. The setup allows nonzero means and a zero kernel; a small positive kernel preserves the initial-round contradiction. The displayed manuscript bounds are invalid in the allowed nonzero-mean setting. No implementation or empirical failure is inferred.","primaryUrl":"https://arxiv.org/html/2502.01226","noteUrl":"/review?paper=review-neurips-2026-150933","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150936","title":"From Density Matrices to Phase Transitions in Deep Learning: Spectral Early Warnings and Interpretability","authors":["Max Hennick","Guillaume Corlouer"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"From Density Matrices to Phase Transitions in Deep Learning: Spectral Early Warnings and Interpretability","plainExplanation":"A diverging covariance signal need not rise monotonically, and selective gradients need not make its curvature correction rank one. Window behavior and statistical sample cost depend on the observed process, rather than unused parameter dimensions alone.","primaryUrl":"https://arxiv.org/html/2603.29805","noteUrl":"/review?paper=review-neurips-2026-150936","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","interpretation_error","missing_assumption","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150938","title":"PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers","authors":["Haopeng Li","Shitong Shao","Wenliang Zhong","zikai zhou","Lichen Bai","Hui Xiong","Zeke Xie"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers","plainExplanation":"The approximate-block numerator uses value sums while the displayed denominator counts one centroid per block. The current total-error theorem and online-rescaling correction are credited. The previously removed subquadratic wording remains resolved in the v4 body.","primaryUrl":"https://arxiv.org/html/2602.01077","noteUrl":"/review?paper=review-neurips-2026-150938","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150941","title":"Visual-ERM: Reward Modeling for Visual Equivalence","authors":["Ziyu Liu","Shengyuan Ding","Xinyu Fang","Xuanlang Dai","Penghui Yang","Jiaqi Wang","Kai Chen","Dahua Lin","Yuhang Zang"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Visual-ERM: Reward Modeling for Visual Equivalence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.13224","noteUrl":"/review?paper=review-neurips-2026-150941","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150944","title":"Guaranteed Jailbreaking Defense via Disrupt-and-Rectify Smoothing","authors":["Zheng Lin","Zhenxing Niu","haoxuan ji","Haichang Gao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The radius argument needs explicit growth and edit assumptions","plainExplanation":"Edit distance obeys the triangle inequality, not equality. A rectifier may undo disruption edits, so a positive number of rectification edits does not imply a positive net radius increment.","primaryUrl":"https://arxiv.org/html/2605.10582","noteUrl":"/review?paper=review-neurips-2026-150944","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150945","title":"Do Thinking Tokens Help with Safety?","authors":["Narutatsu Ri","Abhishek Panigrahi","Sanjeev Arora"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Do Thinking Tokens Help with Safety?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.25013","noteUrl":"/review?paper=review-neurips-2026-150945","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150946","title":"Frozen Memory Is Not Enough: Rethinking External Memory as Extraction","authors":["Mingyuan Li","Guangsheng Yu","Xu Wang","Shaoxiong Ji"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Frozen Memory Is Not Enough: Rethinking External Memory as Extraction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.17050","noteUrl":"/review?paper=review-neurips-2026-150946","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150947","title":"Hyperbolic Graph Neural Networks Under the Microscope: The Role of Geometry–Task Alignment","authors":["Dionisia Naddeo","Jonas Linkerhägner","Nicola Toschi","Geri Skenderi","Veronica Lachi"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Realized distortion is not equal to a universal radius certificate","plainExplanation":"The spatial squared norm is one, while the time squared coordinate is \\(\\|\\phi\\|^2-1/c\\). Its Lorentz norm is therefore \\(-\\|\\phi\\|^2+1/c+1\\), not identically \\(-1/c\\). At \\(c=1,\\phi=(2,0)\\), the output \\((\\sqrt3,1,0)\\) has norm \\(-2\\), rather than \\(-1\\). Below the root’s domain it is not even real. This is a printed lift defect. The theory explicitly excludes Lorentz in the relevant appendix, so no theorem or implementation failure is inferred.","primaryUrl":"https://arxiv.org/html/2602.01828","noteUrl":"/review?paper=review-neurips-2026-150947","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150949","title":"Nearest-Neighbor Radii under Dependent Sampling","authors":["Yuanyuan Gao","Yilong Hou","Zhexiao Lin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Nearest-Neighbor Radii under Dependent Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14343","noteUrl":"/review?paper=review-neurips-2026-150949","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150950","title":"Finding Koopman Invariant Subspaces via Personalized PageRank","authors":["Hyukpyo Hong","Qin Li","Matthew J Colbrook","Hanbaek Lyu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A sufficient mixing condition is used as a necessary one","plainExplanation":"After Corollary 16.1, the paper states that a positive auxiliary PageRank gap requires the mixing condition of Theorem 3(i). Its own Example B has a positive auxiliary gap while that mixing condition fails. Theorem 3 itself is correctly stated as sufficient.","primaryUrl":"https://arxiv.org/html/2605.24666","noteUrl":"/review?paper=review-neurips-2026-150950","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150951","title":"World–Value–Action Model: Implicit Planning for Vision–Language–Action Systems","authors":["Runze Li","Hongyin Zhang","Junxi Jin","Qixin Zeng","Zifeng Zhuang","Yiqi Tang","Shangke Lyu","Donglin Wang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The exponential feasibility comparison uses the wrong sampling measure","plainExplanation":"Uniform actions followed by true dynamics can produce feasible trajectories with probability one, even when the dynamics graph has tiny ambient volume. The ambient-volume bound therefore does not establish the claimed exponential advantage for latent sampling.","primaryUrl":"https://arxiv.org/html/2604.14732","noteUrl":"/review?paper=review-neurips-2026-150951","collection":"neurips-2026","severity":"major","issueTypes":["invalid_reduction","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-150952","title":"SnapAudit: Active Auditing of Differentially Private In-Context Learning via Snapshot-Based Simulation","authors":["Yuyang Xia","Ruixuan Liu","Li Xiong"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SnapAudit: Active Auditing of Differentially Private In-Context Learning via Snapshot-Based Simulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.13502","noteUrl":"/review?paper=review-neurips-2026-150952","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150956","title":"COMPOSE: Hypergraph Cover Optimization for Multi-view 3D Human Pose Estimation","authors":["Tony Danjun Wang","Tolga Birdal","Nassir Navab","Lennart Bastian"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"COMPOSE: Hypergraph Cover Optimization for Multi-view 3D Human Pose Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.09698","noteUrl":"/review?paper=review-neurips-2026-150956","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150958","title":"Sharper Guarantees for Misspecified Kernelized Bandit Optimization","authors":["Davide Maran","Csaba Szepesvari"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Sharper Guarantees for Misspecified Kernelized Bandit Optimization","plainExplanation":"Inherited data cannot be counted as a single budget across every depth. A separate Chebyshev construction defeats the example used to claim a persistent approximation floor.","primaryUrl":"https://arxiv.org/html/2605.05967","noteUrl":"/review?paper=review-neurips-2026-150958","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","interpretation_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150959","title":"Few-Shot Truly Benign DPO Attack for Jailbreaking LLMs","authors":["Yoon Sangyeon","Wonje Jeung","Yoonjun Cho","Dongjae Jeon","Albert No"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Few-Shot Truly Benign DPO Attack for Jailbreaking LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10998","noteUrl":"/review?paper=review-neurips-2026-150959","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150961","title":"SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass","authors":["Chen Qian","Xinran Yu","Xinan Wang","Danyang Li","Guoxuan Chi","Zheng Yang","Qiang Ma","Xin Miao"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.03134v1","noteUrl":"/review?paper=review-neurips-2026-150961","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150963","title":"TIDE: Asymmetric Neural Circuits for Stabilized Temporal Inhibitory-Excitatory Dynamics","authors":["Alexander Kyuroson","Denis Kleyko","Marcus Liwicki"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"TIDE: Asymmetric Neural Circuits for Stabilized Temporal Inhibitory-Excitatory Dynamics","plainExplanation":"Stability must be checked for the actual normalized, heterogeneous-step dynamics. Memory gating, lateral iteration, Perron estimation and matrix-normality claims each need their stated local correction or additional condition.","primaryUrl":"https://arxiv.org/html/2605.19403","noteUrl":"/review?paper=review-neurips-2026-150963","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-150966","title":"OdysSim: Building Foundation Models for Human Behavior Simulation","authors":["Xuhui Zhou","Weiwei Sun","Weihua Du","Jiarui Liu","Haojia Sun","Qianou Ma","Sherry Wu","Yiming Yang","Maarten Sap"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"OdysSim: Building Foundation Models for Human Behavior Simulation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.14199","noteUrl":"/review?paper=review-neurips-2026-150966","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-150967","title":"Tight Generalization Bounds for Noiseless Inverse Optimization","authors":["Sayedpouria Fatemi","Hoomaan Maskan","Suvrit Sra","Peyman Mohajerin Esfahani"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Tight Generalization Bounds for Noiseless Inverse Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08866","noteUrl":"/review?paper=review-neurips-2026-150967","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150968","title":"Action Images: End-to-End Policy Learning via Multiview Video Generation","authors":["Haoyu Zhen","Zixian Gao","Qiao Sun","yilin zhao","Yuncong Yang","Yilun Du","Pengsheng Guo","Tsun-Hsuan Johnson Wang","Yi-Ling Qiao","Chuang Gan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The exact open-gripper encoding gives an empty decoder set","plainExplanation":"For open label \\(g=1\\), every background blue pixel is exactly \\(0.25\\), while the Gaussian-region pixels exceed it. No pixel is strictly below \\(0.25\\), so \\(\\Omega\\) is empty and division by \\(|\\Omega|\\) is undefined even for an exact encoder output. This is not a generation-noise effect. The printed exact decoder has a boundary gap. The inclusive change alone does not prove robustness, and no actual code or empirical failure is inferred.","primaryUrl":"https://arxiv.org/html/2604.06168","noteUrl":"/review?paper=review-neurips-2026-150968","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150970","title":"Affine Tracing: A New Paradigm for Probabilistic Linear Solvers","authors":["Disha Hegde","Marvin Pförtner","Jon Cockayne"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Affine Tracing: A New Paradigm for Probabilistic Linear Solvers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10566","noteUrl":"/review?paper=review-neurips-2026-150970","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150971","title":"Follow the Mean: Reference-Guided Flow Matching","authors":["Pedro Curvo","Maksim Zhdanov","Floor Eijkelboom","Jan-Willem van de Meent"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Follow the Mean: Reference-Guided Flow Matching","plainExplanation":"No issue was found in the checked claims. The former unequal-covariance counterexample is excluded by the revised hypothesis. Under the common covariance, the arithmetic interpolation of posterior means is exact.","primaryUrl":"https://arxiv.org/html/2605.10302","noteUrl":"/review?paper=review-neurips-2026-150971","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150972","title":"ARTIS: Agentic Risk-Aware Test-Time Scaling via Iterative Simulation","authors":["Xingshan Zeng","Lingzhi Wang","Weiwen Liu","Liangyou Li","Yasheng Wang","Lifeng Shang","Xin Jiang","Qun Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ARTIS: Agentic Risk-Aware Test-Time Scaling via Iterative Simulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.01709","noteUrl":"/review?paper=review-neurips-2026-150972","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150973","title":"Vortex: Efficient and Programmable Sparse Attention Serving","authors":["Zhuoming Chen","Xinrui Zhong","Qilong Feng","Ranajoy Sadhukhan","Yang Zhou","Michael Shieh","Zhihao Jia","Beidi Chen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Vortex: Efficient and Programmable Sparse Attention Serving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06453","noteUrl":"/review?paper=review-neurips-2026-150973","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150974","title":"Probabilistic Tiny Recursive Model","authors":["Amin Sghaier","Ali Parviz","Alexia Jolicoeur-Martineau"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Probabilistic Tiny Recursive Model","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19943","noteUrl":"/review?paper=review-neurips-2026-150974","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150976","title":"Language-Assisted Image Clustering Guided by Discriminative Relational Signals and Adaptive Semantic Centers","authors":["Jun Ma","Xu Zhang","Zhengxing Jiao","Yaxin Hou","Hui LIU","Junhui Hou","Yuheng Jia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The defined vector contains logits rather than probabilities","plainExplanation":"The defined vector contains logits rather than probabilities. The definition and interpretation disagree; this does not make logit matching itself invalid.","primaryUrl":"https://arxiv.org/html/2603.24275","noteUrl":"/review?paper=review-neurips-2026-150976","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150977","title":"Flow-DPPO: Divergence Proximal Policy Optimization for Flow Matching Models","authors":["Bowen Ping","Xiangxin Zhou","Penghui Qi","Minnan Luo","Liefeng Bo","Tianyu Pang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The pseudocode omits the variance normalization retained by the equations","plainExplanation":"For equal-covariance Gaussians, the old-policy expected log ratio is \\(\\|\\mu_{\\mathrm{old}}-\\mu\\|^2/(2\\sigma(t)^2)\\). Algorithm 1 line 14 instead sets the divergence to the numerator alone. These are monotone equivalents at a fixed variance only if the threshold is converted accordingly; with timestep-dependent variance, one unchanged threshold does not define the same mask. This is a local normalization inconsistency. Appendix C.4 explicitly says the full normalization is retained; actual repository execution and empirical outcomes are unknown.","primaryUrl":"https://arxiv.org/html/2606.11025","noteUrl":"/review?paper=review-neurips-2026-150977","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150980","title":"Algorithmic Impact Reveals the Hidden Structure of Alignment","authors":["Zachary Wojtowicz","Michelle Si","Finale Doshi-Velez","Ariel Procaccia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Algorithmic Impact Reveals the Hidden Structure of Alignment","plainExplanation":"No issue was found in the checked claims. The proposed counterexample concerns a misspecified best-fit projection. It does not satisfy the exact recovery equations under which the source states its uniqueness result.","primaryUrl":"https://arxiv.org/html/2608.24046v1","noteUrl":"/review?paper=review-neurips-2026-150980","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150981","title":"Behavior Pack Optimization for Video MLLM Post-Training","authors":["Zhaolu Kang","Shiyu Liu","Tailong Luo","Wei Zhang","Yingjie He","Guangyuan Dong","Siheng Wang","Liang He","Lei Wei","Jiaqi Su","Shuang Chen","Guansu Wang","Haoyu Ji","Qishi Zhan","Kaiyue Zhou"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Behavior Pack Optimization for Video MLLM Post-Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03141","noteUrl":"/review?paper=review-neurips-2026-150981","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150984","title":"PLANING: A Loosely Coupled Triangle-Gaussian Framework for Streaming 3D Reconstruction","authors":["Changjian Jiang","Kerui Ren","Xudong Li","Kaiwen Song","Guanghao Li","Linning Xu","Tao Lu","Junting Dong","Yu Zhang","Yu Feng","Bo Dai","Mulin Yu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The listed local vertices contradict the defined triangle frame","plainExplanation":"Every finite MLP output leaves the product zero. Its normalization is \\(0/0\\), so it is not a unit quaternion. Clamping the denominator returns zero, still not a valid unit quaternion. Interpreting zero as axis-angle identity conflicts with the stated four-component quaternion normalization. The printed initialization and rotation formula are incompatible. A simple identity seed fixes validity but does not by itself restore all component expressivity; actual fallbacks and NaNs are unknown.","primaryUrl":"https://arxiv.org/html/2601.22046","noteUrl":"/review?paper=review-neurips-2026-150984","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150985","title":"GARDO: Reinforcing Diffusion Models without Reward Hacking","authors":["Haoran He","Yuxiao YE","Jie Liu","Jiajun Liang","Zhiyong Wang","Ziyang Yuan","Xintao Wang","Hangyu Mao","Meng Wang","Pengfei Wan","Ling Pan"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"GARDO: Reinforcing Diffusion Models without Reward Hacking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.24138","noteUrl":"/review?paper=review-neurips-2026-150985","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150988","title":"Spectrum-Adaptive Generalization Bounds for Trained Deep Transformers","authors":["Mana Sakai","Masaaki Imaizumi"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Spectrum-Adaptive Generalization Bounds for Trained Deep Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07297","noteUrl":"/review?paper=review-neurips-2026-150988","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150989","title":"TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback","authors":["Lei pang","Jun Luo","ruinan Jin"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2508.02833","noteUrl":"/review?paper=review-neurips-2026-150989","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150990","title":"Learning to Evolve Scenes: Reasoning about Human Activities with Scene Graphs","authors":["Francesca Pistilli","Simone Alberto Peirone","Giuseppe Averta"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Minimizing the contrastive ratio pushes paired similarities apart","plainExplanation":"Minimizing the contrastive ratio pushes paired similarities apart. The written optimization direction is wrong; the reported training implementation is not determined by this display.","primaryUrl":"https://arxiv.org/html/2607.02425","noteUrl":"/review?paper=review-neurips-2026-150990","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-150991","title":"MIND: Monge Inception Distance for Generative Models Evaluation","authors":["Quentin Berthet","Clement CREPY","Romuald Elie","Klaus Greff","Michael E Sander","Yu-Han Wu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MIND: Monge Inception Distance for Generative Models Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06797","noteUrl":"/review?paper=review-neurips-2026-150991","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150996","title":"AnaDiffusion: Anatomically Compositional Latent Diffusion for Controllable 3D Brain MRI Generation","authors":["Tracy Han","Lulin Liu","Bangya Liu","Yuanhao Cai","Nuo Chen","Jade Wang","Ziqian Xie","Chenyu You","Shuiwang Ji","Degui Zhi","Zhiwen Fan"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"AnaDiffusion: Anatomically Compositional Latent Diffusion for Controllable 3D Brain MRI Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.23014","noteUrl":"/review?paper=review-neurips-2026-150996","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-150998","title":"MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs","authors":["Ruiyi Yang","Zechen Li","Hao Xue","Imran Razzak","Flora Salim"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs","plainExplanation":"Older mastery peaks need age-dependent decay. Information monotonicity also requires recovery of previous graph attributes, rather than retention of protected nodes alone.","primaryUrl":"https://arxiv.org/html/2605.10064","noteUrl":"/review?paper=review-neurips-2026-150998","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151001","title":"Credit Assignment with Resets in Language Model Reasoning","authors":["Ankur Samanta","Akshayaa Magesh","Ayush Jain","Youliang Yu","Daniel Jiang","Kavosh Asadi","Kaveh Hassani","Paul Sajda","Jalaj Bhandari","Yonathan Efroni"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Keeping successes before a discarded failure biases the base group","plainExplanation":"With success probability \\(p\\), the first retained base sample succeeds if the first draw succeeds, or if it fails and a fresh top-up draw succeeds. Its success probability is \\(p+(1-p)p\\), equal to \\(0.75\\) at \\(p=0.5\\), rather than \\(0.5\\). Fresh top-ups do not remove conditioning of the already retained successes. The iid base-group claim is false for the printed procedure. Resetting as a general technique is not rejected.","primaryUrl":"https://arxiv.org/html/2605.25507","noteUrl":"/review?paper=review-neurips-2026-151001","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151002","title":"A Scalable Nonparametric Continuous-Time Survival Model through Numerical Quadrature","authors":["Chaeyeon Lee","Sehwan Kim","Hyungrok Do"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A Scalable Nonparametric Continuous-Time Survival Model through Numerical Quadrature","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.16208","noteUrl":"/review?paper=review-neurips-2026-151002","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151003","title":"Hyperbolic Concept Bottleneck Models","authors":["Daniel Uyterlinde","Swasti S Mishra","Pascal Mettes"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Hyperbolic Concept Bottleneck Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06440","noteUrl":"/review?paper=review-neurips-2026-151003","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151005","title":"ReSET: Accurate Latency-Critical NVFP4 Reasoning via Step-Aware Temperature Scaling","authors":["Sihwa Lee","Janghwan Lee","Donghoon Yoo","Jae Gon Kim","Hanyul Ryu","Soojung Ryu","Jungwook Choi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ReSET: Accurate Latency-Critical NVFP4 Reasoning via Step-Aware Temperature Scaling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.13233","noteUrl":"/review?paper=review-neurips-2026-151005","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151006","title":"Theory Guided and Interpretable Neural Operator Design for Partial Differential Equation Learning","authors":["Zeyuan Song","Zheyu Jiang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Several proof steps need regularity or constant corrections","plainExplanation":"The claimed embedding is false. A bounded continuous derivative of order \\(k\\) does not give an essentially bounded weak derivative of order \\(k+1\\).","primaryUrl":"https://arxiv.org/html/2609.33715","noteUrl":"/review?paper=review-neurips-2026-151006","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151007","title":"Spectral Progressive Diffusion for Efficient Image and Video Generation","authors":["Howard Xiao","Brian Chao","Lior Yariv","Gordon Wetzstein"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Fine-tuning omits part of the aligned inference-time stage interval","plainExplanation":"Equation (6) gives \\(\\operatorname{aligned}(t)-t=(r-1)t(1-t)/(1+(r-1)t)>0\\) for \\(r>1\\) and interior \\(t\\). Stage inference covers \\((t_i,\\operatorname{aligned}(t_{i-1})]\\), while training uses \\((t_i,t_{i-1}]\\). With transition \\(0.5\\) and ratio two, time \\(0.6\\) belongs to high-resolution inference but is assigned low resolution in fine-tuning. Evaluating the already sampled time in equation (13) does not fill the missing interval. The printed fine-tuning recipe does not match the full inference schedule. The alignment algebra, actual implementation and empirical quality are not refuted.","primaryUrl":"https://arxiv.org/html/2605.18736","noteUrl":"/review?paper=review-neurips-2026-151007","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151008","title":"Normalized Architectures are Natively 4-Bit","authors":["Maxim Fishman","Brian Chmiel","Ron Banner","Daniel Soudry","Boris Ginsburg"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Normalized Architectures are Natively 4-Bit","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06067","noteUrl":"/review?paper=review-neurips-2026-151008","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151010","title":"Environment-Robust Representation Learning with Empirical Bayes","authors":["Yuli Slavutsky","Matthew Shen","Bohan Wu","David Blei"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Environment-Robust Representation Learning with Empirical Bayes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05365","noteUrl":"/review?paper=review-neurips-2026-151010","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151011","title":"AI Control for Sandbagging on Fuzzy Tasks","authors":["Mikhail Terekhov","Caglar Gulcehre","Vivek Hebbar","Joe Benton"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AI Control for Sandbagging on Fuzzy Tasks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.08892","noteUrl":"/review?paper=review-neurips-2026-151011","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151012","title":"Meta$^n$: Recursive Self-Improvement through Emergent Depth","authors":["Zae Myung Kim","Young-Jun Lee","Seungyeon Jwa","Dongyeop Kang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Meta$^n$: Recursive Self-Improvement through Emergent Depth","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.24735","noteUrl":"/review?paper=review-neurips-2026-151012","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151013","title":"Predictive but Not Plannable: RC-aux for Latent World Models","authors":["Wenyuan Li","Guang Li","Keisuke Maeda","Takahiro Ogawa","Miki Haseyama"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Predictive but Not Plannable: RC-aux for Latent World Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07278","noteUrl":"/review?paper=review-neurips-2026-151013","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151014","title":"Catch-Only-One: Non-Transferable Examples for Model-Specific Authorization","authors":["Zihan Wang","Ethan Ma","Zhongkui Ma","Shuofeng Liu","Akide Liu","Derui Wang","Minhui Xue","Guangdong Bai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Catch-Only-One: Non-Transferable Examples for Model-Specific Authorization","plainExplanation":"A large change in logits can leave every predicted class unchanged. Accuracy separation needs control of class contrasts and decision margins.","primaryUrl":"https://arxiv.org/html/2510.10982","noteUrl":"/review?paper=review-neurips-2026-151014","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","unsupported_claim"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151016","title":"Soft Forward-Backward Representations for Zero-shot Reinforcement Learning with General Utilities","authors":["Marco Bagatella","Thomas Rupf","Georg Martius","Andreas Krause"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Soft Forward-Backward Representations for Zero-shot Reinforcement Learning with General Utilities","plainExplanation":"Soft policy extraction uses the wrong reward scale, and Appendix C pairs a Gaussian path with an incompatible velocity. Each issue has an explicit formula correction.","primaryUrl":"https://arxiv.org/html/2602.06769","noteUrl":"/review?paper=review-neurips-2026-151016","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151017","title":"PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies","authors":["Yufei Guo","Yinan Wu","Haoran Duan","guiguang ding","Jungong Han"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.15285","noteUrl":"/review?paper=review-neurips-2026-151017","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151022","title":"MUX: Continuous Reasoning via Multiplexed Tokens","authors":["Ayhan Suleymanzade","Halil Alperen Gozeten","Michael Bronstein","Ismail Ilkan Ceylan","Jinwoo Kim"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"MUX: Continuous Reasoning via Multiplexed Tokens","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.18264","noteUrl":"/review?paper=review-neurips-2026-151022","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151023","title":"When a Zero-Shooter Cheats: Improving Age Estimation via Activation Steering","authors":["Erik Imgrund","Pia Hanfeld","Klim Kireev","Konrad Rieck"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Absolute shortcut and surrogate errors need not add","plainExplanation":"Absolute shortcut and surrogate errors need not add. The causal magnitude identification is unsupported without those premises; qualitative empirical shortcuts are not refuted.","primaryUrl":"https://arxiv.org/pdf/2605.17658v1","noteUrl":"/review?paper=review-neurips-2026-151023","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151024","title":"Teacher-Aware Evolution of Heuristic Programs from Learned Optimization Policies","authors":["Minyu Chen","Song Qin","Ling-I Wu","Jianxin Xue","Guoqiang Li"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Teacher-Aware Evolution of Heuristic Programs from Learned Optimization Policies","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10634","noteUrl":"/review?paper=review-neurips-2026-151024","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151026","title":"Few-Step Cofolding with All-Atom Flow Maps","authors":["Gianluca Scarpellini","Ron Shprints","Peter Holderrieth","Juno Nam","Pranav Murugan","Rafael Gomez-Bombarelli","Tommi Jaakkola","Maruan Al-Shedivat","Nicholas Boffi","Joey Bose"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The average-velocity integral has the opposite orientation to the map","plainExplanation":"Only \\(x_{\\sigma_N}\\) is initialized, but the loop starts at index zero and first reads \\(x_{\\sigma_0}\\), which was never produced; a subsequent expression divides by \\(\\sigma_0=0\\). Produced entries are tagged \\(m+1\\), and the next index is set to one beyond that tag, skipping levels while moving toward higher noise. No produced entry receives terminal tag zero. The FK branch also fails to insert its candidates into the global set used by the return step, so ordinary initialization does not provide its required terminal candidates. These are displayed-pseudocode defects. A literal reimplementation is underdefined; the uncached repository and reported search results are not independently refuted.","primaryUrl":"https://arxiv.org/html/2606.08375","noteUrl":"/review?paper=review-neurips-2026-151026","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151028","title":"The Value of Being Wrong: Self-Mined Visual In-Context Learning from Errors","authors":["Shu Zhao","TAN YU"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"The Value of Being Wrong: Self-Mined Visual In-Context Learning from Errors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16977","noteUrl":"/review?paper=review-neurips-2026-151028","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151029","title":"An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence","authors":["Qizhen (Irene) Zhang","Ankush Garg","Jakob Foerster","Niladri S. Chatterji","Kshitiz Malik","Mike Lewis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.02400","noteUrl":"/review?paper=review-neurips-2026-151029","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151030","title":"FAME: Forecasting Academic Impact via Continuous-Time Manifold Evolution","authors":["Jianrong Ding","Jianyuan Zhong","Zhengyan Shi","Qiang Xu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The selection threshold mixes raw and normalized impact scales","plainExplanation":"The selection threshold mixes raw and normalized impact scales. The selection equation does not implement its raw-impact criterion; implementation behavior is not established.","primaryUrl":"https://arxiv.org/html/2605.07208","noteUrl":"/review?paper=review-neurips-2026-151030","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151031","title":"NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis","authors":["Ramil Khafizov","Ilya Statsenko","Ruslan Rakhimov","Artem Komarichev","Peter Wonka","Evgeny Burnaev"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04722","noteUrl":"/review?paper=review-neurips-2026-151031","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151032","title":"Correpondence Alignment For Improved Virtual Try-On","authors":["Jiyoung Kim","Youngjin Shin","Siyoon Jin","Dahyun Chung","Jisu Nam","Tongmin Kim","JongJae Park","Hyeonwoo Kang","Seungryong Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A sliced attention block is not a normalized garment-coordinate distribution","plainExplanation":"A sliced attention block is not a normalized garment-coordinate distribution. Eq. (18) needs normalization. Eq. (20) is valid on a full row; an analogous entropy criticism applies only if it is instead evaluated on the garment slice.","primaryUrl":"https://arxiv.org/html/2602.17636v1","noteUrl":"/review?paper=review-neurips-2026-151032","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151034","title":"FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models","authors":["Fan Mo","Han Yuxuan","Geng Zhang","Wangbo Zhao","Yang You"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Differentiating the balancing ratio cancels the intended subnet signal","plainExplanation":"Differentiating the balancing ratio cancels the intended subnet signal. The written objective does not give the claimed gradient reweighting; no conclusion is drawn about whether code already detaches it.","primaryUrl":"https://arxiv.org/html/2606.27866","noteUrl":"/review?paper=review-neurips-2026-151034","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151037","title":"Best Arm Identification in Generalized Linear Bandits via Hybrid Feedback","authors":["Qirun Zeng","Xuchuang Wang","Jiayi Shen","Xutong Liu","Fang Kong","Jinhang Zuo"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The positive-definite design domain need not attain its infimum","plainExplanation":"The positive-definite design domain need not attain its infimum. The claimed PD optimum need not exist; spanning does not make the open positive-definite feasible subset compact.","primaryUrl":"https://arxiv.org/html/2605.05745","noteUrl":"/review?paper=review-neurips-2026-151037","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151038","title":"TreeGraft: Adaptive Multi-Drafter Grafting for Tree-Based Speculative Decoding","authors":["Jiaming Fan","Daming Cao","Canchen Huang","Jiale Fu","Jin Zhang","Junjie Gao","Kai Yang","Xiangzhong Luo","Xu Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TreeGraft: Adaptive Multi-Drafter Grafting for Tree-Based Speculative Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.26112","noteUrl":"/review?paper=review-neurips-2026-151038","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151039","title":"Propagation of Chaos in Contextual Flow Maps","authors":["Chen","Zhengjiang Lin","Kaizhao Liu","Philippe Rigollet"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Propagation of Chaos in Contextual Flow Maps","plainExplanation":"Strong contraction can keep fixed-time errors small without making the largest error over infinitely many fresh batches small. The counterexample preserves the stated regularity and admissible parameter conditions.","primaryUrl":"https://arxiv.org/html/2605.16747","noteUrl":"/review?paper=review-neurips-2026-151039","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-151041","title":"GradShield: Alignment Preserving Finetuning","authors":["Zhanhao Hu","Xiao Huang","Patrick Mendoza","Emad A Alghamdi","Basel Alomair","Raluca Popa","David Wagner"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Differentiability alone does not give the quadratic influence remainder","plainExplanation":"Take one datum with loss \\(L_a(\\theta)=\\theta\\), one dummy with zero loss, initial parameter zero, bounded smooth update \\(g(v)=v/\\sqrt{1+v^2}\\), and \\(S(\\theta)=\\theta+|\\theta|^{4/3}\\). Both data orders put the active update at zero. Actual leave-one-out FIHS is \\(\\eta/\\sqrt2-(\\eta/\\sqrt2)^{4/3}\\), while the printed leading term is \\(\\eta/\\sqrt2\\). For fixed \\(n=2\\), the remainder divided by \\(n\\eta^2\\) diverges. A \\(C^1\\) score permits this example. The theorem’s stated generality is too broad. The actual smooth LLM proxy and the fixed-horizon mechanism are not shown to fail.","primaryUrl":"https://arxiv.org/html/2605.14194","noteUrl":"/review?paper=review-neurips-2026-151041","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151046","title":"Tracing Agentic Failure from the Flow of Success","authors":["Samuel (Min-Hsuan) Yeh","Yiwen Zhu","Shaleen Deep","Sharon Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Pooling dependent trajectory steps does not supply the claimed conformal rank","plainExplanation":"Use \\(21\\) iid successful trajectories, each with all \\(100\\) scores equal to one uniform random value \\(U_i\\), and an independent identical test trajectory. The \\(2100\\) pooled scores repeat each trajectory order statistic one hundred times. Their \\(0.8\\) quantile is \\(U_{(17)}\\), giving marginal test-step false-positive rate \\(\\mathbb E[1-U_{(17)}]=5/22>0.2\\). Even a pooled finite-sample rank correction remains at the same repeated block. The claimed distribution-free step-FPR control does not follow from iid trajectories alone. Empirical detector performance is not invalidated.","primaryUrl":"https://arxiv.org/html/2607.12747","noteUrl":"/review?paper=review-neurips-2026-151046","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151048","title":"Cat-DPO: Category-Adaptive Safety Alignment","authors":["Tiankai Yang","Yi Nian","Xinyuan Li","Ruiyao Xu","Henry P Zou","Kaize Ding","Xiyang Hu","Yan Liu","Yue Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The safety proxy and two reductions need narrower claims","plainExplanation":"The preference margin and the on-policy probability of unsafe behavior are different quantities. Their signs relative to the same safety budget need not agree without a calibration assumption.","primaryUrl":"https://arxiv.org/html/2604.17299","noteUrl":"/review?paper=review-neurips-2026-151048","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151049","title":"Optimal algorithmic complexity of inference in quantum kernel methods","authors":["Elies Gil-Fuster","Seongwook Shin","Sofiene Jerbi","Jens Eisert","Maximilian J Kramer"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Optimal algorithmic complexity of inference in quantum kernel methods","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2604.15214","noteUrl":"/review?paper=review-neurips-2026-151049","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151051","title":"Improving Generative Adversarial Networks with Self-Distillation","authors":["Antoni Nowinowski","Krzysztof Krawiec"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Improving Generative Adversarial Networks with Self-Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08577","noteUrl":"/review?paper=review-neurips-2026-151051","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151052","title":"Maximin Robust Bayesian Experimental Design","authors":["Hany Abdulsamad","Sahel Iqbal","Christian Andersson Naesseth","Takuo Matsubara","Adrien Corenflos"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Maximin Robust Bayesian Experimental Design","plainExplanation":"Positive estimator bias prevents the claimed purely quadratic log-MGF bound. The repair keeps the bias term and controls fluctuations around the estimator’s mean.","primaryUrl":"https://arxiv.org/html/2603.14094","noteUrl":"/review?paper=review-neurips-2026-151052","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151053","title":"Sharp Convergence and Sample Complexity of Policy Mirror Descent for Average-Reward MDPs","authors":["Enes Arda","Atilla Eryilmaz"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Sharp Convergence and Sample Complexity of Policy Mirror Descent for Average-Reward MDPs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04117","noteUrl":"/review?paper=review-neurips-2026-151053","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151054","title":"Classification Fields: Arbitrarily Fine Recursive Hierarchical Clustering From Few Examples","authors":["Yicen Li","Ruiyang Hong","Anastasis Kratsios","Haitz Sáez de Ocáriz Borde","Paul D. McNicholas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Classification Fields: Arbitrarily Fine Recursive Hierarchical Clustering From Few Examples","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07119","noteUrl":"/review?paper=review-neurips-2026-151054","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151058","title":"Active Context Selection Improves Simple Regret in Contextual Bandits","authors":["Mohammad Shahverdikondori","Jalal Etesami","Negar Kiyavash"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The least-sampled arm is defined through the family it must itself determine","plainExplanation":"The least-sampled arm is defined through the family it must itself determine. The given hard-family selection is circular; this does not disprove the minimax theorem by another construction.","primaryUrl":"https://arxiv.org/html/2605.20040","noteUrl":"/review?paper=review-neurips-2026-151058","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151060","title":"Automatic Textbook Formalization","authors":["Fabian Gloeckle","Ahmad Rammal","Charles Arnal","Remi Munos","Vivien Cabannes","Gabriel Synnaeve","Amaury Hayat"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Automatic Textbook Formalization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.03071","noteUrl":"/review?paper=review-neurips-2026-151060","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151062","title":"Colour me shocked: Exact Molecular Hessians from MLIPs in O(N) time using sparse differentiation!","authors":["Luca Thiede","Andreas Burger","Alan Aspuru-Guzik"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Colour me shocked: Exact Molecular Hessians from MLIPs in O(N) time using sparse differentiation!","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.24720v1","noteUrl":"/review?paper=review-neurips-2026-151062","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151063","title":"Information Loss and Disparate Effects in Network Embeddings","authors":["Gabriel Chuang","Augustin Chaintreau"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The PSD multiplier differentiates the other diagonal entry","plainExplanation":"The determinant constraint is \\(K_2^2-K_1K_3\\leq0\\). Its derivative in \\(K_3\\) is \\(-K_1\\), whereas its derivative in \\(K_1\\) is \\(-K_3\\). The second stationarity equation repeats \\(K_3\\) where \\(K_1\\) is required. The receipt constructs a boundary optimum \\((K_1,K_2,K_3)=(1,2,4)\\) with valid interior target probabilities that satisfies corrected KKT but leaves printed residual \\(-0.003\\). Strict convexity on the PSD cone confirms it is a true optimum. This is a local KKT formula defect, not failure of the convex PSD optimization formulation or an empirical result.","primaryUrl":"https://arxiv.org/html/2509.12396","noteUrl":"/review?paper=review-neurips-2026-151063","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151066","title":"Fisher Decorator: Refining Flow Policy via  A Local Transport Map","authors":["Xiaoyuan Cheng","Haoyu Wang","Wenxuan Yuan","Ziyan Wang","Zonghao Chen","Li Zeng","Zhuo Sun"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"General small transports have second-order Jacobian contributions to KL","plainExplanation":"Theorem 3.1 gives a Fisher-only second-order KL expansion for a small differentiable transport \\(T(a)=a+\\delta(a)\\), without a near-rigid derivative hypothesis. For standard Gaussian base and \\(T(a)=(1+h)a\\), exact KL is \\(\\tfrac12[(1+h)^2-1-2\\log(1+h)]=h^2+O(h^3)\\).","primaryUrl":"https://arxiv.org/html/2604.17919","noteUrl":"/review?paper=review-neurips-2026-151066","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151068","title":"Do Pathology Foundation Models Encode Disease Progression? A Pseudotime Analysis of Visual Representations","authors":["Pritika Vig","Renchin Wu","William Lotter"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The spectral series includes an unwanted time-zero contribution","plainExplanation":"The spectral series includes an unwanted time-zero contribution. The printed spectral approximation has a local offset error; spectral truncation does not explain it.","primaryUrl":"https://arxiv.org/html/2601.21334","noteUrl":"/review?paper=review-neurips-2026-151068","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151072","title":"Distributed Online Convex Optimization with Compressed Communication: Optimal Regret and Applications","authors":["Sifan Yang","Danyue Li","Lijun Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"An expected communication delay is not a samplewise information barrier","plainExplanation":"An expected communication delay is not a samplewise information barrier. The block-independence equality is false; this does not itself settle the lower-bound rate.","primaryUrl":"https://arxiv.org/html/2604.09276","noteUrl":"/review?paper=review-neurips-2026-151072","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151073","title":"Synthesis Through Simulation: Generating Coherent Enterprise Data via Scalable Agent-System Interaction","authors":["Yipeng Li","Ashutosh Hathidara","Jane Lo","Harshavardhan Abichandani","Gunraj Singh","Atin Ghosh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Synthesis Through Simulation: Generating Coherent Enterprise Data via Scalable Agent-System Interaction","plainExplanation":"Two graphs can require both edge deletions and edge insertions. Their sum can exceed either graph's edge count, even after optimizing the vertex correspondence.","primaryUrl":"https://arxiv.org/html/2610.10549","noteUrl":"/review?paper=review-neurips-2026-151073","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151074","title":"Exploring MLLM-Diffusion Information Transfer with MetaCanvas","authors":["Han Lin","Xichen Pan","Ziqi Huang","Ji Hou","Jialiang Wang","Weifeng Chen","Zecheng He","Felix Juefei-Xu","Junzhe Sun","Zhipeng Fan","Ali Thabet","Mohit Bansal","Chu Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Exploring MLLM-Diffusion Information Transfer with MetaCanvas","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.11464","noteUrl":"/review?paper=review-neurips-2026-151074","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151075","title":"DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation","authors":["Jusuk Lee","Seungjae Lee","Jonghun Shin","Hoseong Jung","Sungha Kim","Daesol Cho","H. Jin Kim","Jia-Bin Huang","Furong Huang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30350","noteUrl":"/review?paper=review-neurips-2026-151075","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151077","title":"Attention-Based Soft Answer Sets","authors":["Wael AbdAlmageed"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Attention-Based Soft Answer Sets","plainExplanation":"The product operator prevents two positions from sharing positive mass on the same symbol. It does not force a position to concentrate its mass on one symbol when the group has fewer positions than available symbols.","primaryUrl":"https://arxiv.org/html/2603.18436v1","noteUrl":"/review?paper=review-neurips-2026-151077","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151080","title":"3D Point Splatting for mmWave Radar Novel View Synthesis","authors":["Adnan Armouti","Yixuan Gao","Rajalakshmi Nandakumar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The azimuth inverse has the wrong sine coefficient","plainExplanation":"The azimuth inverse has the wrong sine coefficient. The inverse-grid display is wrong; which mapping was used in execution remains undetermined.","primaryUrl":"https://arxiv.org/html/2609.11894","noteUrl":"/review?paper=review-neurips-2026-151080","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151084","title":"Reconstructing the Vocal Tract with Differentiable Acoustic Simulation","authors":["Eric Chen","Jin Woo Lee","Vincent Sitzmann"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Reconstructing the Vocal Tract with Differentiable Acoustic Simulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36737","noteUrl":"/review?paper=review-neurips-2026-151084","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151085","title":"BridgeTwist: Twisting Schrödinger Bridges for Training-Free Conditional Sampling","authors":["Mengyu Li","Qianqian Qu","Jun Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BridgeTwist: Twisting Schrödinger Bridges for Training-Free Conditional Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.18658","noteUrl":"/review?paper=review-neurips-2026-151085","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151087","title":"Within-Model vs Between-Prompt Variability in Large Language Models for Creative Tasks","authors":["Jennifer Haase","Jana Gonnermann-Müller","Paul H. P. Hanel","Nicolas Leins","Thomas Kosch","Jan Mendling","Sebastian Pokutta"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Within-Model vs Between-Prompt Variability in Large Language Models for Creative Tasks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.21339","noteUrl":"/review?paper=review-neurips-2026-151087","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151088","title":"Flowette: Flow Matching with Graphette Priors for Graph Generation","authors":["Asiri Wijesinghe","Sevvandi Kandanaarachchi","Daniel M Steinberg","Cheng Soon Ong"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Flowette: Flow Matching with Graphette Priors for Graph Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.23566","noteUrl":"/review?paper=review-neurips-2026-151088","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151090","title":"Norm Enforcement for AI Agents: Robustly Shaping Behavior in Multi-Agent Systems","authors":["Yaowen Ye","Jacob Steinhardt"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Norm Enforcement for AI Agents: Robustly Shaping Behavior in Multi-Agent Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.09766","noteUrl":"/review?paper=review-neurips-2026-151090","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151093","title":"Cost-Aware Learning","authors":["Clara Mohri","Amir Globerson","Haim Kaplan","Tomer Koren","Yishay Mansour"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Cost-Aware Learning","plainExplanation":"The density order is correct for the fractional covering problem. An integral item can overshoot the required coverage by an arbitrarily large amount, so rounding plain cost-sorted greedy has no factor-two guarantee.","primaryUrl":"https://arxiv.org/html/2604.28020","noteUrl":"/review?paper=review-neurips-2026-151093","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151094","title":"Feature Recovery for Object Understanding Under Physical Transformation","authors":["Aditi Tiwari","Sofia Stoica","Savya Khosla","David Forsyth","Heng Ji"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Feature Recovery for Object Understanding Under Physical Transformation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.12078v1","noteUrl":"/review?paper=review-neurips-2026-151094","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151097","title":"DARLING: Detection Augmented Reinforcement Learning with Non-Stationary Guarantees","authors":["Argyrios Gerogiannis","Yu-Han Huang","Venugopal Veeravalli"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Linear successor features need not identify transition distributions","plainExplanation":"The detection argument infers a changed successor-feature expectation from any transition change in a linear MDP; the appendix separately assumes injectivity. Let successor states be \\((-1,0,1/2,1)\\), actions \\(a\\in\\{-1,1\\}\\), and \\(\\phi(x,a)=(1,ax)/\\sqrt2\\).","primaryUrl":"https://arxiv.org/html/2604.16684","noteUrl":"/review?paper=review-neurips-2026-151097","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151098","title":"PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining","authors":["Zelin Jia","Zhao Zhang","Zhicong Tang","Yuhui Yuan","Shixia Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.22789","noteUrl":"/review?paper=review-neurips-2026-151098","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151099","title":"Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learning","authors":["Katarzyna Filus","Kamil Faber","Roberto Corizzo","Christopher Kanan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Positive F1 does not establish retained concept information","plainExplanation":"Positive F1 does not establish retained concept information. The F1-zero diagnostic inference fails; other alignment and probe results are not overturned.","primaryUrl":"https://arxiv.org/html/2605.16374","noteUrl":"/review?paper=review-neurips-2026-151099","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151101","title":"Learning from Trials and Errors: Reflective Test-Time Planning for Embodied LLMs","authors":["Yining Hong","Huang Huang","Manling Li","Fei-Fei Li","Leonidas Guibas","Jiajun Wu","Yejin Choi"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Learning from Trials and Errors: Reflective Test-Time Planning for Embodied LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.21198","noteUrl":"/review?paper=review-neurips-2026-151101","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151102","title":"Trajectory Planning without Trajectory Data: A Manifold-Guided Approach","authors":["Silong Yong","Anji Liu","Cunxi Dai","Carl Busart","Guanya Shi","Yilun Du","Katia Sycara","Yaqi Xie"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The deterministic transition kernel omits the current state","plainExplanation":"The deterministic state-transition kernel omits the current state from its center. It sends a stationary nonzero state to zero rather than leaving it fixed. This is a local transition-kernel typo. It does not establish that the constructed ODE or implemented trajectories omit their current state.","primaryUrl":"https://arxiv.org/html/2610.08863","noteUrl":"/review?paper=review-neurips-2026-151102","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151103","title":"Sample complexity of stochastic optimization with integer variables","authors":["Hongyu Cheng","Yinghao Zheng","Marco Molinaro","Amitabh Basu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Sample complexity of stochastic optimization with integer variables","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07239","noteUrl":"/review?paper=review-neurips-2026-151103","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151105","title":"Leveraging Data Symmetries to Select an Optimal Subset of Training Data under Label Noise","authors":["Kumar Shubham","Pavan Karjol","Kiran M K","Prathosh AP"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Unit-ball volume alone does not determine the distribution constant’s dimensional decay","plainExplanation":"Sorting changes support regularity as well as density. The missing-geometry gap affects the universal feasible-k claim. The corner calculation is an upper bound, not an assertion about an optimal global regularity constant.","primaryUrl":"https://arxiv.org/html/2605.01874","noteUrl":"/review?paper=review-neurips-2026-151105","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151108","title":"Block Optimism for Nonstationary Bandits with Latent Linear Dynamics","authors":["Taehyun Hwang","Hyun-jun Choi","Heesang Ann","Min-hwan Oh"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The confidence radius does not account for the proof’s split failure budget","plainExplanation":"The confidence radius does not account for the proof’s split failure budget. The displayed confidence proof/radius matching has a budget gap; this fixed-factor repair preserves the asymptotic rate and is not a universal confidence-theorem counterexample.","primaryUrl":"https://arxiv.org/html/2610.00911","noteUrl":"/review?paper=review-neurips-2026-151108","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151109","title":"Global linear convergence of entropy-regularized softmax policy gradient beyond tabular MDPs","authors":["Ziyue Chen","David Siska","Lukasz Szpruch"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The explicit rate constant can require an unstated bounded concentrability density","plainExplanation":"The explicit rate constant can require an unstated bounded concentrability density. The stated explicit finite constants are not delivered under all hypotheses. The linear-rate conclusion is not refuted, because the direct-measure repair remains available.","primaryUrl":"https://arxiv.org/html/2605.24939","noteUrl":"/review?paper=review-neurips-2026-151109","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151111","title":"Outcome-Based RL Provably Leads Transformers to Reason, but Only With the Right Data","authors":["Yuval Ran-Milo","‪Yotam Alexander‬‏","Shahar Mendel","Nadav Cohen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Local no-jump renormalization is not global survival conditioning","plainExplanation":"Definition 17 removes long jumps and renormalizes each row, then identifies absorption probabilities with conditioning on no long jump anywhere before absorption. Let \\(K(x,y)\\) be the original short-step kernel and \\(q(x)\\) the probability of surviving to absorption without a long jump.","primaryUrl":"https://arxiv.org/html/2601.15158","noteUrl":"/review?paper=review-neurips-2026-151111","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151112","title":"Discrete Diffusion Models Exploit Asymmetry to Solve Lookahead Planning Tasks","authors":["Itamar Trainin","Shauli Ravfogel","Omri Abend","Amir Feder"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The serialized path length has a local counting error","plainExplanation":"The sequence-length expression uses the total graph-vertex count for the answer suffix. That suffix has only the number of vertices in a single path.","primaryUrl":"https://arxiv.org/html/2602.19980","noteUrl":"/review?paper=review-neurips-2026-151112","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151113","title":"Neural Harmonic Measure Operator","authors":["Jinjin He","Sinan Wang","Yuchen Sun","Bo Zhu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neural Harmonic Measure Operator","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35752","noteUrl":"/review?paper=review-neurips-2026-151113","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151114","title":"Probing for Representation Manifolds in Superposition","authors":["Alexander Modell"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Probing for Representation Manifolds in Superposition","plainExplanation":"Conditional expectation minimizes an unconstrained squared error in the target function. The unit-variance and sequential orthogonality constraints change that minimizer, even when the regression weights are fitted jointly.","primaryUrl":"https://arxiv.org/html/2605.18537","noteUrl":"/review?paper=review-neurips-2026-151114","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151116","title":"Dynamical Adapter Fusion: Constructing A Global Adapter for Pre-Trained Model-based Class-Incremental Learning","authors":["RuiQi Liu","Boyu Diao","Zijia An","Runjie Shao","Feixiang Ge","Zhulin An","Fei Wang","Yongjun Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Clipping does not define constant-Fisher zero quotients","plainExplanation":"Clipping does not define constant-Fisher zero quotients. The printed stable recipe is incomplete on allowed inputs; this does not establish that observed runs encountered them.","primaryUrl":"https://arxiv.org/html/2601.21341","noteUrl":"/review?paper=review-neurips-2026-151116","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151120","title":"Policy Optimization in Hybrid Discrete-Continuous Action Spaces via Mixed Gradients","authors":["Matias Alvo","Daniel Russo","Yashodhan Kanoria"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Policy Optimization in Hybrid Discrete-Continuous Action Spaces via Mixed Gradients","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14297","noteUrl":"/review?paper=review-neurips-2026-151120","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151123","title":"BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation","authors":["Haiquan Wen","Yiwei He","Zhenglin Huang","Tianxiao Li","Zihan Yu","Xingru Huang","Lu Qi","Baoyuan Wu","Xiangtai Li","Guangliang Cheng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.12620","noteUrl":"/review?paper=review-neurips-2026-151123","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151127","title":"Multi-agent Collaboration with State Management","authors":["Mengyang Liu","Taozhi Chen","Zhenhua Xu","Xue Jiang","Yihong Dong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Multi-agent Collaboration with State Management","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20563","noteUrl":"/review?paper=review-neurips-2026-151127","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151128","title":"CanViT: Toward Active-Vision Foundation Models","authors":["Yohaï-Eliel BERREBY","Sabrina Du","Audrey Durand","B. S Krishna"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CanViT: Toward Active-Vision Foundation Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.22570","noteUrl":"/review?paper=review-neurips-2026-151128","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151130","title":"TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models","authors":["Kai Liu","Anqi Li","Junxian Li","Zhixin Wang","Zhikai Chen","Renjing Pei","Yulun Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models","plainExplanation":"The full distance matrix requires quadratic pairwise work, and the greedy recurrence makes one candidate-wide update per retained token. The local fixed-window density score alone can be linear; that does not make the complete selector linear.","primaryUrl":"https://arxiv.org/html/2610.17530v1","noteUrl":"/review?paper=review-neurips-2026-151130","collection":"neurips-2026","severity":"repairable","issueTypes":["complexity_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151132","title":"Learned Lagrangian Models of PDEs via Euler–Lagrange Residual Minimization","authors":["Lyra Zhornyak","Eric Forgoston","M. Ani Hsieh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Zero collocation residual does not imply exact symplecticity with quadrature error","plainExplanation":"Zero collocation residual does not imply exact symplecticity with quadrature error. The universal exact-zero-residual implication fails. Near-symplecticity and the claimed fourth-order behavior are not refuted.","primaryUrl":"https://arxiv.org/html/2605.07157","noteUrl":"/review?paper=review-neurips-2026-151132","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151133","title":"Cascaded Sparse Autoencoders LearnMulti-Level Visual Concepts in Multimodal LLMs","authors":["Yusong Zhao","Hengyi Wang","Tanuja Ganu","Akshay Nambi","Hao Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Cascaded Sparse Autoencoders LearnMulti-Level Visual Concepts in Multimodal LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.16193v3","noteUrl":"/review?paper=review-neurips-2026-151133","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151134","title":"DetailAnywhere: Fashion Detail Generation via Cross-Modal Feature Alignment Distillation","authors":["Zijun Li","Yimin Zhou","Jia Sun","Honglie Wang","Pengcheng Wei","junlong wu","Yongrui Heng","Jiyuan Wang","Huan Ouyang","Boheng Zhang","Huaiqing Wang","Dewen Fan","Qianqian Gan","Fan Yang","Tingting Gao"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"DetailAnywhere: Fashion Detail Generation via Cross-Modal Feature Alignment Distillation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.02220","noteUrl":"/review?paper=review-neurips-2026-151134","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151135","title":"WebSpline: Structure-Informed Splines for Real-Time 3D Gaussians from Monocular Videos","authors":["Jongmin Park","Jeonghwan Yun","Minh-Quan V Bui","Munchurl Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"WebSpline: Structure-Informed Splines for Real-Time 3D Gaussians from Monocular Videos","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02096","noteUrl":"/review?paper=review-neurips-2026-151135","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151137","title":"ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs","authors":["Chayne Thrash","Ali Abbasi","Soheil Kolouri"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The range calculation is valid; rotation merging has an architectural condition","plainExplanation":"If several query heads share one value projection, distinct value rotations cannot all be absorbed into that single shared matrix. The merge description needs to specify the relevant sharing structure.","primaryUrl":"https://arxiv.org/html/2605.10793","noteUrl":"/review?paper=review-neurips-2026-151137","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["missing_assumption"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151140","title":"BSTabDiff: Block-Subunit Diffusion Priors for High-Dimensional Tabular Data Generation","authors":["Al Zadid Sultan Bin Habib","Md Younus Ahamed","Prashnna K Gyawali","Gianfranco Doretto","Donald A Adjeroh"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"BSTabDiff: Block-Subunit Diffusion Priors for High-Dimensional Tabular Data Generation","plainExplanation":"Marginal transformation, block identification and mask factorization require different assumptions. The stated general model does not supply them; the released fixed-rate mask sampler is a valid restricted case.","primaryUrl":"https://arxiv.org/html/2606.09257","noteUrl":"/review?paper=review-neurips-2026-151140","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-151143","title":"Flow-Direct: Feedback-Efficient and Reusable Guidance for Flow Models via Non-Parametric Guidance Field","authors":["Kim Yong Tan","Yueming LYU","Ivor Tsang","Yew Soon Ong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Flow-Direct: Feedback-Efficient and Reusable Guidance for Flow Models via Non-Parametric Guidance Field","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16348","noteUrl":"/review?paper=review-neurips-2026-151143","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151146","title":"The Sampling Complexity of Condorcet Winner Identification in Dueling Bandits","authors":["El Mehdi Saad","Victor Thuot","Nicolas Verzelen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The printed hard-instance blocks do not have the claimed signs and antisymmetry","plainExplanation":"The printed hard-instance blocks do not have the claimed signs and antisymmetry. The printed instance is not admissible as asserted; a corrected lower bound requires rechecking rather than being disproved by these local defects.","primaryUrl":"https://arxiv.org/html/2603.15189","noteUrl":"/review?paper=review-neurips-2026-151146","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151149","title":"Why Do DiT Editors Drift? Plug-and-Play Low Frequency Alignment in VAE Latent Space","authors":["Xiaoce Wang","Sifan Zhou","Kaifei Wang","Leli Xu","Xuerui Qiu","Tao He","Ming Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Why Do DiT Editors Drift? Plug-and-Play Low Frequency Alignment in VAE Latent Space","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08250","noteUrl":"/review?paper=review-neurips-2026-151149","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151150","title":"Agent Explorative Policy Optimization for Agentic Multimodal Reasoning","authors":["Minki Kang","Shizhe Diao","Ryo Hachiuma","Sung Ju Hwang","Pavlo Molchanov","Frank Wang","Byung-Kwan Lee"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Success coverage does not equal a positive GRPO advantage","plainExplanation":"Success coverage includes all-success groups with zero centered advantages, so it does not equal the probability of a positive GRPO reward signal. Proposition 1’s binomial coverage comparison remains valid.","primaryUrl":"https://arxiv.org/html/2605.28774","noteUrl":"/review?paper=review-neurips-2026-151150","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151151","title":"Monotone Inclusion Approach to Weakly Monotone Discrete-Time Finite-Horizon Mean-Field Games","authors":["Ugur Aydin","Tamer Basar","Naci Saldi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Monotone Inclusion Approach to Weakly Monotone Discrete-Time Finite-Horizon Mean-Field Games","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.15079","noteUrl":"/review?paper=review-neurips-2026-151151","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151153","title":"SOAR: Regression-based LiDAR Relocalization for UAVs","authors":["Hengyu Mu","Jianshi Wu","Yuxin Guo","XianLian Lin","Qingyong Hu","Sheng Ao","Chenglu Wen","Cheng Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SOAR: Regression-based LiDAR Relocalization for UAVs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.13267","noteUrl":"/review?paper=review-neurips-2026-151153","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151155","title":"Optimizing the Envy Cycle Elimination Algorithm","authors":["Karen Frilya Celine","Warut Suksompong"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Optimizing the Envy Cycle Elimination Algorithm","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02233","noteUrl":"/review?paper=review-neurips-2026-151155","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151159","title":"ASAP: Attention Sink Anchored Pruning","authors":["Jaehyuk Lee","Hanyoung Kim","Yanggee Kim","Donghun Lee"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Equal-width distance bins are not invariant to rank-preserving transforms","plainExplanation":"The printed pruning algorithm can divide by a zero quota and return more tokens than its budget. Its equal-width distance bins also depend on magnitudes, contradicting the rank-only justification.","primaryUrl":"https://arxiv.org/html/2605.22372","noteUrl":"/review?paper=review-neurips-2026-151159","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151160","title":"Defining Operational Conditions for Safety-Critical AI-Based Systems from Data","authors":["Johann M Christensen","Elena Hoemann","Frank Köster","Sven Hallerbach"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Defining Operational Conditions for Safety-Critical AI-Based Systems from Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.22118","noteUrl":"/review?paper=review-neurips-2026-151160","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151162","title":"Frequency Domain Reservoir Computing","authors":["Klaus Schertler","Xiomara Runge","Andrea Ceni","David Kappel","Claudio Gallicchio"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stated ESP condition is not necessary under leakage","plainExplanation":"Leakage and negative recurrent weights invalidate the proposed necessary ESP condition. Separately, frequency mixing can leave the Hermitian subspace required to apply the packing isometry and ridge-equivalence argument.","primaryUrl":"https://arxiv.org/html/2606.24969","noteUrl":"/review?paper=review-neurips-2026-151162","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151163","title":"Multilingual Safety Alignment via Self-Distillation","authors":["Ruiyang Qin","Qingzhuo Wang","Dongrui Liu","Qiang Li","Zhihua Wei","Wen Shen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Multilingual Safety Alignment via Self-Distillation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.02971","noteUrl":"/review?paper=review-neurips-2026-151163","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151165","title":"Efficient Serving for Dynamic Agent Workflows with Prediction-based KV-Cache Management","authors":["Haoyu Zheng","Fangcheng Fu","Jia Wu","Binhang Yuan","Yongqiang Zhang","Hao Wang","Yuanyuan Zhu","Xiao Yan","Jiawei Jiang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Efficient Serving for Dynamic Agent Workflows with Prediction-based KV-Cache Management","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06472","noteUrl":"/review?paper=review-neurips-2026-151165","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151176","title":"COPRA: Conditional Parameter Adaptation with Reinforcement Learning for Video Anomaly Detection","authors":["Darryl Cherian Jacob","Xinyu Liu","Kai Wang","Pan He"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"COPRA: Conditional Parameter Adaptation with Reinforcement Learning for Video Anomaly Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15325","noteUrl":"/review?paper=review-neurips-2026-151176","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151180","title":"MMDiff: Multimodal Model Diffing for Feature Discovery and Control","authors":["Lachin Naghashyar","Hunar Batra","Ashkan Khakzar","Philip Torr","Ronald Clark","Christian Schroeder de Witt","Constantin Venhoff"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MMDiff: Multimodal Model Diffing for Feature Discovery and Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.09928","noteUrl":"/review?paper=review-neurips-2026-151180","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151181","title":"Toward the Goldilocks Blind Compression of Quantum States","authors":["Hyunho Cha","Chae-Yeun Park","Jungwoo Lee"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Toward the Goldilocks Blind Compression of Quantum States","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.01258","noteUrl":"/review?paper=review-neurips-2026-151181","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151184","title":"High-dimensional Gaussian Graphical Model Testing for Long-Memory Time Series","authors":["Percy S. Zhai","Ping-Shou Zhong","Wei Biao Wu"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"High-dimensional Gaussian Graphical Model Testing for Long-Memory Time Series","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30565","noteUrl":"/review?paper=review-neurips-2026-151184","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151188","title":"LookThere! Sparse Vision by Reinforced Selection","authors":["Sreehari Rammohan","Yousef Yassin","Anthony Fuller","Junfeng Wen","Evan Shelhamer","Carl Vondrick"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LookThere! Sparse Vision by Reinforced Selection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.04698","noteUrl":"/review?paper=review-neurips-2026-151188","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151189","title":"SePO: Self-Evolving Prompt Agent for System Prompt Optimization","authors":["Wangcheng Tao","Han Wu","Weng-Fai Wong"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"SePO: Self-Evolving Prompt Agent for System Prompt Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04465","noteUrl":"/review?paper=review-neurips-2026-151189","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151194","title":"Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection","authors":["Jianqiao Cui","haoxian zhu","Yanghao Zhang","Hanbiao Meng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.16049","noteUrl":"/review?paper=review-neurips-2026-151194","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151195","title":"Selective Rollout: Mid-Trajectory Termination for Multi-Sample Agent RL","authors":["Zhiyuan Zhai","Xin Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A ninety-percent termination gate requires all eight trajectories","plainExplanation":"A ninety-percent termination gate requires all eight trajectories. This concerns the formal threshold’s integer requirement, rather than an accusation about a measured termination percentage.","primaryUrl":"https://arxiv.org/html/2605.05802","noteUrl":"/review?paper=review-neurips-2026-151195","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151196","title":"TaskGround: Structured Executable Task Inference for Full-Scene Household Reasoning","authors":["ZhiYuan Feng","Yu Deng","Ruichuan An","Zhenhua Liu","Qixiu Li","Keming Wu","Zhiying Du","Weijie Wang","Haoxiao Wang","Shuang Chen","Sicheng Xu","Yaobo Liang","Jiaolong Yang","Baining Guo"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"TaskGround: Structured Executable Task Inference for Full-Scene Household Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18109","noteUrl":"/review?paper=review-neurips-2026-151196","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151197","title":"Estimating Model-Level Membership Inference Vulnerability Without Reference Models","authors":["Euodia Dodd","Natasa Krco","Igor Shilov","Matthew R Wicker","Yves-Alexandre de Montjoye"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The nonmember average is a false-positive rate rather than a true-negative rate","plainExplanation":"Equation (2) returns one for predicted membership, while Equation (7) averages that indicator over nonmembers and labels it TNR. On the nonmember set, an output of one is a false positive.","primaryUrl":"https://arxiv.org/html/2510.19773","noteUrl":"/review?paper=review-neurips-2026-151197","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151204","title":"OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping","authors":["Mohammed Ehab","Aymane El Gadarri","Vivek Farias","Adam Jozefiak","Ciamac C Moallemi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.11089","noteUrl":"/review?paper=review-neurips-2026-151204","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151205","title":"Paleoinspired Vision: From Exploring Colour Vision Evolution to Inspiring Camera Design","authors":["Yijie Lu","Zhimin Zong","Junjie Zhang","Shenghan Su","Lin Gu","Ziteng Cui","Zeyun Zhao","Yan Pu","Jing Lu","Daisuke Kojima","Ruogu Fang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Paleoinspired Vision: From Exploring Colour Vision Evolution to Inspiring Camera Design","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2412.19439","noteUrl":"/review?paper=review-neurips-2026-151205","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151206","title":"ES-Merging: Biological MLLM Merging via Embedding Space Signals","authors":["Wonbin Lee","Dongki Kim","Sung Ju Hwang"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"ES-Merging: Biological MLLM Merging via Embedding Space Signals","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.14405","noteUrl":"/review?paper=review-neurips-2026-151206","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151208","title":"Low-Rank Adaptation for Critic Learning in Off-Policy Reinforcement Learning","authors":["Yuan Zhuang","Yuexin Bian","Sihong He","Jie Feng","Qing Su","Songyang Han","Jonathan Petit","Shihao Ji","Yuanyuan Shi","Fei Miao"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Uniform reflected actions do not produce a stationary centre peak","plainExplanation":"The described uniform-action reflected walk does not have the asserted stationary centre peak. Standard holding reflection gives a uniform law; bouncing reflection gives equal masses across all interior states. The explicit stationary-law description is incorrect under either standard reflection convention. The behavior/target distribution shift can remain; measured replay distributions and critic results are not determined by this correction.","primaryUrl":"https://arxiv.org/html/2604.18978","noteUrl":"/review?paper=review-neurips-2026-151208","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151210","title":"Local linear convergence of gradient methods for overparameterized Gaussian mixtures","authors":["Jingxing Wang","Vasileios Charisopoulos","Maryam Fazel"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The auxiliary bound controls tangent motion by an arbitrarily small coefficient","plainExplanation":"The auxiliary bound controls tangent motion by an arbitrarily small coefficient. The printed auxiliary bound is false; a corrected normal-distance statement may preserve the convergence result.","primaryUrl":"https://arxiv.org/html/2605.30936","noteUrl":"/review?paper=review-neurips-2026-151210","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151212","title":"Provably Efficient Regularized Online RLHF with Generalized Bilinear Preferences","authors":["Junghyun Lee","Minju Hong","Kwang-Sung Jun","Chulhee Yun","Se-Young Yun"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The curvature hypothesis does not cover every estimator in the theorem","plainExplanation":"The generic-link dual-gap theorem uses curvature outside its stated domain and admits a counterexample; the logistic case has the required global control. The coverage lemma needs a vectorization transpose, and the auxiliary exploration proof needs a larger logarithmic budget.","primaryUrl":"https://arxiv.org/html/2602.23116","noteUrl":"/review?paper=review-neurips-2026-151212","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151213","title":"Learning When Visual Context Matters for Mouse Behavior Analysis","authors":["Weihan Li","Jingyang Ke","Qinheng Pu","Yule Wang","Chengrui Li","Anqi Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning When Visual Context Matters for Mouse Behavior Analysis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.07279","noteUrl":"/review?paper=review-neurips-2026-151213","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151215","title":"DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech","authors":["Junwon Moon","Yejin Lee","Seungbeom Kim","Hoseong Ahn","Sewoong Park","Heeseung Kim","Kyuhong Shim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.04140","noteUrl":"/review?paper=review-neurips-2026-151215","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151217","title":"Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators in Transformers","authors":["Yihong Chen","Zhouchen Lin","Quanming Yao"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators in Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.17771","noteUrl":"/review?paper=review-neurips-2026-151217","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151218","title":"GlyphAnchor: Enhancing Visual Text Rendering via Position-Anchored Glyph Priors","authors":["Qiang Xiang","Shuang Sun","Binglei Li","Yibo Chen","Xu Tang","Yao Hu","Junping Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GlyphAnchor: Enhancing Visual Text Rendering via Position-Anchored Glyph Priors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.02349","noteUrl":"/review?paper=review-neurips-2026-151218","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151219","title":"Edit the Bits, Diff the Codes: Bitwise Residual Editing for Visual Autoregressive Models","authors":["Shengqiang Zhang","Ruotong Liao","Volker Tresp","Barbara Plank","Hinrich Schuetze"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Edit the Bits, Diff the Codes: Bitwise Residual Editing for Visual Autoregressive Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.13558","noteUrl":"/review?paper=review-neurips-2026-151219","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151221","title":"Probabilistic Data-Driven Modelling of Astrophysical Transients: The Neural Process Family for Ultrafast and Class-Agnostic Light Curve Reconstruction","authors":["Siddharth Chaini","federica bianco","Ashish Mahabal"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Probabilistic Data-Driven Modelling of Astrophysical Transients: The Neural Process Family for Ultrafast and Class-Agnostic Light Curve Reconstruction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.27527","noteUrl":"/review?paper=review-neurips-2026-151221","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151223","title":"UpSafe℃: Upcycling for Controllable Safety in Large Language Models","authors":["Yuhao Sun","Lingyun Yu","Zhuoer Xu","Kun Yang","shiwen cui","Yongdong Zhang","Hongtao Xie"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Temperature scaling can reverse the claimed monotone routing preference","plainExplanation":"The routing bias changes log odds affinely, rather than producing a proportional probability difference. The combined bias-and-temperature rule can also reduce safety-expert probability as the control parameter increases.","primaryUrl":"https://arxiv.org/html/2510.02194","noteUrl":"/review?paper=review-neurips-2026-151223","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151226","title":"Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures","authors":["Mark Muchane","Sean M Richardson","Kiho Park","Victor Veitch"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The algorithm multiplies the feature matrices in the wrong order for its penalty","plainExplanation":"The algorithm multiplies the feature matrices in the wrong order for its penalty. The algorithm and feature-biorthogonality definition disagree; no implementation result is inferred.","primaryUrl":"https://arxiv.org/html/2506.01197","noteUrl":"/review?paper=review-neurips-2026-151226","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151227","title":"Intend, Reflect, Refine: An Adaptive Multimodal Reflection Framework for Autonomous Driving","authors":["Zisheng Chen","Yuping Qiu","Jianhua Han","Tao Tang","Xiuwei Chen","Likui Zhang","Yingcong Chen","Hang Xu","Xiaodan Liang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Intend, Reflect, Refine: An Adaptive Multimodal Reflection Framework for Autonomous Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.22913","noteUrl":"/review?paper=review-neurips-2026-151227","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151229","title":"Fast 4D Mesh Generation by Spatio-Temporal Attention Chains","authors":["Dvir Samuel","Yuval Atzmon","Gal Chechik","Yoni Kasten"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The temporal confidence factor cancels from the normalized attention","plainExplanation":"The temporal confidence factor cancels from the normalized attention. The formula does not implement the claimed confidence weighting.","primaryUrl":"https://arxiv.org/html/2605.19786","noteUrl":"/review?paper=review-neurips-2026-151229","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151232","title":"TALON: Confidence-Aware Speculative Decoding with Adaptive Token Trees","authors":["Tianyu Liu","Qitan Lv","Yuhao Shen","Jun Zhang","Xiao Sun","Xiaoyan Sun"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"TALON: Confidence-Aware Speculative Decoding with Adaptive Token Trees","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.07353","noteUrl":"/review?paper=review-neurips-2026-151232","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151234","title":"MissPath-FM: Flow Matching with Structured Priors for Partially Observed Time Series","authors":["Genpei Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Coverage calibration error does not identify posterior variance collapse","plainExplanation":"Coverage calibration error does not identify posterior variance collapse. The collapse diagnosis is not implied by ACE alone; the measured calibration error itself is not refuted.","primaryUrl":"https://arxiv.org/html/2610.17715","noteUrl":"/review?paper=review-neurips-2026-151234","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151235","title":"A Kernel Nonconformity Score for Multivariate Conformal Prediction","authors":["Louis Meyer","Wenkai Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Reusing calibration residuals to fit the score breaks finite-sample coverage","plainExplanation":"Calibration residuals participate in fitting the score map, while the test residual does not. Exchangeable residuals therefore need not produce exchangeable fitted scores. The conformal rank argument requires symmetry at the score level and cannot be applied directly to this arrangement. The contradiction concerns the finite-sample marginal coverage claim for Algorithm 1 as written. The separate asymptotic conditional clause has additional assumptions and is not disproved by this construction; the separate-fit repair does not establish that clause.","primaryUrl":"https://arxiv.org/html/2604.21595","noteUrl":"/review?paper=review-neurips-2026-151235","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151237","title":"Learning Weakly Communicating Average-Reward CMDPs: Strong Duality and Improved Regret","authors":["Kihyun Yu","Beomhan Baek","Dabeen Lee"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Learning Weakly Communicating Average-Reward CMDPs: Strong Duality and Improved Regret","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11586","noteUrl":"/review?paper=review-neurips-2026-151237","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151238","title":"Sponsored Questions and How to Auction Them","authors":["Kshipra Bhawalkar","Alexandros Psomas","Di Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The proposed equilibrium leaves a profitable strict bid reduction","plainExplanation":"The proposed equilibrium leaves a profitable strict bid reduction. The displayed profile is not a Nash equilibrium; the welfare lower-bound theorem may admit the local profile repair.","primaryUrl":"https://arxiv.org/html/2512.03975","noteUrl":"/review?paper=review-neurips-2026-151238","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151239","title":"Almost Sure Convergence Rates of Stochastic Approximation and Reinforcement Learning via a Poisson-Moreau Drift","authors":["Xinyu Liu","Zixuan Xie","Shangtong Zhang"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Almost Sure Convergence Rates of Stochastic Approximation and Reinforcement Learning via a Poisson-Moreau Drift","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07104","noteUrl":"/review?paper=review-neurips-2026-151239","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151244","title":"AgentWeave: Efficient Distributed Agent Serving via Flow Decomposition","authors":["Kaibin Guo","Pengtu Li","Zicong Hong","Zhiyuan Fang","Wuhui Chen","Rachid Guerraoui","Anne-marie Kermarrec"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The scalar scheduler objective does not guarantee nonincreasing tail latency","plainExplanation":"The scalar scheduler objective does not guarantee nonincreasing tail latency. The no-tail-increase guarantee is unsupported; measured P90 improvements can still hold.","primaryUrl":"https://arxiv.org/html/2610.14804","noteUrl":"/review?paper=review-neurips-2026-151244","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151245","title":"Just Ramp-Up: Debiasing Regression-based Estimator for A/B Tests under Network Interference","authors":["Qianyi Chen","Bo Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The unnormalized interference quadratic has variance of order n","plainExplanation":"Lemma C.1 claims \\(\\operatorname{Var}((z-w\\mathbf1)^\\top Bz)=O(1/n)\\) under complete randomization and the stated regular-interference conditions. On an undirected cycle let \\(B=A/2\\) and exactly \\(d=n/2\\) units be treated.","primaryUrl":"https://arxiv.org/html/2410.12740","noteUrl":"/review?paper=review-neurips-2026-151245","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151246","title":"Compute-optimal data scaling for neural surrogates via multi-fidelity training","authors":["Paul Setinek","Felix Koehler","Fabian Paischer","Nils Thuerey","Johannes Brandstetter"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Compute-optimal data scaling for neural surrogates via multi-fidelity training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.17221","noteUrl":"/review?paper=review-neurips-2026-151246","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151250","title":"Truth as a Compression Artifact in Language Model Training","authors":["Konstantin Krestnikov"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Truth as a Compression Artifact in Language Model Training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.11749","noteUrl":"/review?paper=review-neurips-2026-151250","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151251","title":"SASA: Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability","authors":["Seyed Arshan Dalili","Mehrdad Mahdavi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The quantitative population risk gain can exceed the maximum possible gain","plainExplanation":"The population risk-gain lower bound can exceed the maximum possible gain. Factor scaling contradicts the claimed uniqueness up to orthogonal rotations, and the finite-sample missing-direction threshold fails even under balanced routing.","primaryUrl":"https://arxiv.org/html/2606.06333","noteUrl":"/review?paper=review-neurips-2026-151251","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151252","title":"Self-Organized Conformal Prediction: Reducing Regional Coverage Gaps with Unsupervised Group Discovery","authors":["Louis Berthier","Ahmed Shokry","Maxime Moreaud","Guillaume Ramelet","Aymeric Dieuleveut"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Self-Organized Conformal Prediction: Reducing Regional Coverage Gaps with Unsupervised Group Discovery","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.29403","noteUrl":"/review?paper=review-neurips-2026-151252","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151253","title":"Multistage Defer Trees for Hybrid Interpretability: If at First You Can't Succeed, Tree Again","authors":["Zakk Heile","Hayden McTavish","Margo Seltzer","Cynthia Rudin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Multistage Defer Trees for Hybrid Interpretability: If at First You Can't Succeed, Tree Again","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30995","noteUrl":"/review?paper=review-neurips-2026-151253","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151256","title":"AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces","authors":["Sungho Park","Wonjoong Kim","Rongyuan Tan","Jue Zhang","WOOK SHIN HAN","Pengfei Gao","Chanyoung Park","Qingwei Lin","Saravan Rajmohan","Dongmei Zhang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.23041","noteUrl":"/review?paper=review-neurips-2026-151256","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151258","title":"Belief Engine: Configurable Stance Dynamics for Multi-Agent LLM Deliberation","authors":["Joshua C Yang","Maurice Flechtner","Damian Dailisan","Michiel Bakker"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Belief Engine: Configurable Stance Dynamics for Multi-Agent LLM Deliberation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15343v1","noteUrl":"/review?paper=review-neurips-2026-151258","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151260","title":"FARE: Forensic Acceptance Region Estimation for Catching Bait-and-Switch Image Generators","authors":["Kai Yao","Marc Juarez"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FARE: Forensic Acceptance Region Estimation for Catching Bait-and-Switch Image Generators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30982","noteUrl":"/review?paper=review-neurips-2026-151260","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151263","title":"BASTION: Budget-Aware Speculative Decoding with Tree-structured Block Diffusion Drafting","authors":["Soowon Oh","Nam Cao","Yujin Kim","Hojung Jung","Huzama Ahmad","Sangmin Bae","Se-Young Yun"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"BASTION: Budget-Aware Speculative Decoding with Tree-structured Block Diffusion Drafting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29727","noteUrl":"/review?paper=review-neurips-2026-151263","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151264","title":"SkillOpt: Executive Strategy for Self-Evolving Agent Skills","authors":["Yifan Yang","Ziyang Gong","weiquan Huang","Qihao Yang","Ziwei Zhou","Zisu Huang","Yan Li","Xuemei Gao","Qi Dai","Bei Liu","Kai Qiu","Yuqing Yang","Dongdong Chen","Xue Yang","Chong Luo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SkillOpt: Executive Strategy for Self-Evolving Agent Skills","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23904","noteUrl":"/review?paper=review-neurips-2026-151264","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151266","title":"MedFlowSeg: Flow Matching for Medical Image Segmentation with Frequency-Aware Attention","authors":["Zhi Chen","Runze Hu","Le Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"MedFlowSeg: Flow Matching for Medical Image Segmentation with Frequency-Aware Attention","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.19675v2","noteUrl":"/review?paper=review-neurips-2026-151266","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151267","title":"The Key to Going Linear: Analysis-Driven Transformer Linearization","authors":["Anna Kuzina","Paul Whatmough","Babak Ehteshami Bejnordi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Diagonal gates cannot be moved past the key contractions","plainExplanation":"The key-product remainder includes interactions between gates, and concentration of key directions does not by itself make that product a centering projection. Diagonal forget gates must remain interleaved with the key contractions. The omitted uniform attention weights also contribute a value mean that need not vanish with prefix length. A separate interval correction makes the displayed cache partition agree with its stated budget. These findings concern the mathematical approximation and specification, not trained-model performance.","primaryUrl":"https://arxiv.org/html/2607.07706","noteUrl":"/review?paper=review-neurips-2026-151267","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement","formula_error","incorrect_bound","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151269","title":"Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning","authors":["Sudip Bhujel","Shanghao Shi","Ruiquan Huang","Ning Zhang","Yang Xiao"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30258","noteUrl":"/review?paper=review-neurips-2026-151269","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151271","title":"Adam under Generalized Smoothness with Second-Moment-Type Stochastic Gradients","authors":["ruinan Jin","Difei Cheng","Ling Chen","Jun Luo","Hao Zhou","Youzhi Zhang"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Adam under Generalized Smoothness with Second-Moment-Type Stochastic Gradients","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.37787","noteUrl":"/review?paper=review-neurips-2026-151271","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151272","title":"Indications of Belief-Guided Agency and Meta-Cognitive Monitoring in Large Language Models","authors":["Noam Steinmetz Yalon","Ariel Goldstein","Liad Mudrik","Mor Geva"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Indications of Belief-Guided Agency and Meta-Cognitive Monitoring in Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.02467","noteUrl":"/review?paper=review-neurips-2026-151272","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151273","title":"Cross Flow: One-Step Generation Across Latent and Pixel Spaces","authors":["Xiyuan Wang","Xiao Zhang","Yang Li","Ruoxi Jiang","Zhao Zhong","Liefeng Bo","Muhan Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A derivative-only loss needs a generative boundary condition","plainExplanation":"A derivative-only loss needs a generative boundary condition. The unrestricted branch is underidentified; the default \\(\\phi=r\\) experiments and the boundary-qualified appendix construction are not refuted.","primaryUrl":"https://arxiv.org/html/2606.19970","noteUrl":"/review?paper=review-neurips-2026-151273","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151277","title":"Missing data and cluster graphs: cluster-level missingness vs variable-level missingness","authors":["Willow J. Scott","Eugenio Valdano","Charles Assaad"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Missing data and cluster graphs: cluster-level missingness vs variable-level missingness","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20943","noteUrl":"/review?paper=review-neurips-2026-151277","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151278","title":"Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers","authors":["Jan Dubiński","Jan Betley","Anna Sztyber-Betley","Daniel Tan","Owain Evans"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.25891","noteUrl":"/review?paper=review-neurips-2026-151278","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151279","title":"World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video","authors":["Liyuan Zhu","Shengyu Huang","Amrita Mazumdar","Tianye Li","Zan Gojcic","Gordon Wetzstein","Iro Armeni","Shalini De Mello","Alex Trevithick"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.01202","noteUrl":"/review?paper=review-neurips-2026-151279","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151280","title":"Stochastic Matching Bandits with Rare Optimization Updates","authors":["Jung-hun Kim","Min-hwan Oh"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The batch-budget schedule and baseline revenue objective need separate corrections","plainExplanation":"Algorithm 2 can exceed its supplied epoch budget because its first scale does not depend on that budget. The separate baseline UCB objective omits heterogeneous reward weights. Both assertions remain in v3; the main batched revenue formulas already include the weights.","primaryUrl":"https://arxiv.org/html/2509.04194","noteUrl":"/review?paper=review-neurips-2026-151280","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","formula_error"],"classifications":["major"]},{"slug":"review-neurips-2026-151283","title":"Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies","authors":["Xinchen Jin","Aditya Chatterjee","Pranav Kumar","Rohan Paleja"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17204","noteUrl":"/review?paper=review-neurips-2026-151283","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151285","title":"Measure Less, Know More: Self-Supervised Test-Time Feature Acquisition","authors":["Eeshaan Jain","Linus Bleistein","Bart Deplancke","Charlotte Bunne"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Likelihood-ratio contraction requires the fractional-moment regime","plainExplanation":"Likelihood-ratio contraction requires the fractional-moment regime. The moment regime and inherited proof need correction; no empirical failure follows from this theoretical gap.","primaryUrl":"https://arxiv.org/html/2610.03454","noteUrl":"/review?paper=review-neurips-2026-151285","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151286","title":"Disentangled Representation Learning via Flow Matching","authors":["Jinjin Chi","Taoping Liu","Mengtao Yin","Ximing Li","Yongcheng Jing","Jialie Shen","Leszek Rutkowski","Dacheng Tao"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Disentangled Representation Learning via Flow Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.05214","noteUrl":"/review?paper=review-neurips-2026-151286","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151288","title":"Decentralized Ranking Aggregation via Gossip: Convergence and Robustness","authors":["Kerrian Le Caillec","Anna van Elst","Igor Colin","Stephan Clémençon"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Integer vote replacement invalidates the unrounded breakdown bound","plainExplanation":"The empirical breakdown bound overlooks integer vote replacement, and the diameter lower bound fails on rankings already at consensus. Arbitrary Byzantine vectors are not clipped by Borda’s identity map, while the honest-mean stability argument uses a different reference from the theorem’s original target.","primaryUrl":"https://arxiv.org/html/2602.22847","noteUrl":"/review?paper=review-neurips-2026-151288","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151290","title":"VICO: Visual Environments Co-Evolving for Vision-Language Model Reasoning","authors":["Meng Lu","Ligeng Zhu","Olivia Xiao","Yuchen Zhuang","Zihan Wang","Kuncheng Wu","Bangya Liu","Yu Wang","Charles Fleming","Wenqi Shi","Xuan Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Reward variance is not generally proportional to policy-gradient second moment","plainExplanation":"Reward variance is not generally proportional to policy-gradient second moment. The general proportionality/maximizer justification fails; intermediate difficulty can remain a conditional heuristic.","primaryUrl":"https://arxiv.org/html/2610.10782","noteUrl":"/review?paper=review-neurips-2026-151290","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151295","title":"Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation","authors":["Qianhao Yuan","Jie Lou","XingYu Li","Hongyu Lin","Le Sun","Xianpei Han","Yaojie Lu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18740","noteUrl":"/review?paper=review-neurips-2026-151295","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151296","title":"CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening","authors":["Loka Li","Jin Tian","Kun Zhang"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening","plainExplanation":"The same observations and sparse interaction mask admit different mixtures of the latent concepts. The stated assumptions therefore do not identify each concept individually.","primaryUrl":"https://arxiv.org/html/2610.07340","noteUrl":"/review?paper=review-neurips-2026-151296","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","lemma_error","missing_assumption"],"classifications":["critical"]},{"slug":"review-neurips-2026-151299","title":"AlloGen: Conformation-Selective Binder Design with Differential State Scoring","authors":["Hanqun Cao","Aastha Pal","Sumi Kimura","Yesol Kim","Jingjie Zhang","Pheng-Ann Heng","Pranam Chatterjee"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AlloGen: Conformation-Selective Binder Design with Differential State Scoring","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05474v1","noteUrl":"/review?paper=review-neurips-2026-151299","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151300","title":"NashDreamer: Model-Based Reinforcement Learning for Zero-Sum Imperfect-Information Games","authors":["Tomáš Holeček","Viliam Lisy"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"NashDreamer: Model-Based Reinforcement Learning for Zero-Sum Imperfect-Information Games","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.01549","noteUrl":"/review?paper=review-neurips-2026-151300","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151301","title":"SEAR: Sample Efficient Action Chunking Reinforcement Learning","authors":["C F Maximilian Nagy","Onur Celik","Emiliyan Gospodinov","Florian Seligmann","Weiran Liao","Aryan Kaushik","Gerhard Neumann"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Per-action entropy does not restore feedback lost inside an open-loop chunk","plainExplanation":"The chunk Q-function duplicates a boundary reward, although the later target already uses the correct sum. The critic loss indexes a different set of horizons from its targets and actor. Per-action entropy also cannot restore feedback unavailable within an open-loop chunk.","primaryUrl":"https://arxiv.org/html/2603.01891","noteUrl":"/review?paper=review-neurips-2026-151301","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151303","title":"BEACON: Cross-Domain Co-Training of Generative Robot Policies via Best-Effort Adaptation","authors":["Antong Zhang","Han Qi","Heng Yang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The surrogate qualification is corrected; one boundary convention remains","plainExplanation":"The remaining issue is the undefined inverse at the boundary, not the old incompatible projection set. The factorization loses identifiability there, so a convention is needed.","primaryUrl":"https://arxiv.org/html/2605.08571","noteUrl":"/review?paper=review-neurips-2026-151303","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151304","title":"Efficient and Transferable Agentic Knowledge Graph RAG via Reinforcement Learning","authors":["Junhong Lin","Shicheng Liu","Jinyeop Song","Song Wang","Julian Shun","Yada Zhu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Efficient and Transferable Agentic Knowledge Graph RAG via Reinforcement Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2509.26383","noteUrl":"/review?paper=review-neurips-2026-151304","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151305","title":"Test-Time Compute Games","authors":["Ander Artola Velasco","Dimitrios Rontogiannis","Stratis Tsirtsis","Manuel Rodriguez"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The equilibrium and convergence claims must include abstention","plainExplanation":"An allowed market can have no participating users. In that case every provider receives zero utility at every action profile, so every profile is a Nash equilibrium and the dominant provider serves no queries.","primaryUrl":"https://arxiv.org/html/2601.21839","noteUrl":"/review?paper=review-neurips-2026-151305","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151306","title":"Semantic Routing: Exploring Multi-Layer LLM Feature Weighting for Diffusion Transformers","authors":["Bozhou Li","Yushuo Guan","Haolin Li","Bohan Zeng","Yiyan Ji","Yue Ding","Pengfei Wan","Kun Gai","Yuanxing Zhang","Wentao Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Semantic Routing: Exploring Multi-Layer LLM Feature Weighting for Diffusion Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.03510","noteUrl":"/review?paper=review-neurips-2026-151306","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151307","title":"Rank-Transformed Dissimilarity Profiles for High-Dimensional Classification","authors":["Xiangbo Mo","Hao Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The shifted fourth-moment formula needs transposed mixing matrices","plainExplanation":"The general shifted fourth-moment formula needs transposed mixing matrices. The central-limit proof’s retained blocks also share latent coordinates, invalidating the asserted independence without a wider gap or a dependent-limit argument.","primaryUrl":"https://arxiv.org/html/2306.15199","noteUrl":"/review?paper=review-neurips-2026-151307","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151309","title":"Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps","authors":["Yanke Zhou","Yiduo Li","Hanlin Tang","Maohua Li","Kan Liu","Tao Lan","Lin Qu","Yuan Yao","Xiaoxing Ma"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16928","noteUrl":"/review?paper=review-neurips-2026-151309","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151310","title":"Balancing Frequencies and Pixels in Flow Matching","authors":["Lucas Degeorge","Paul Couairon","Arijit Ghosh","Alexei Efros","David Picard","Vicky Kalogeiton"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The spectral loss does not give equal total weight to each octave","plainExplanation":"Equal per-bin residuals give octave totals proportional to their bin counts, contradicting equal total octave weighting. The complex Fourier residual also remains phase-sensitive when amplitudes agree.","primaryUrl":"https://arxiv.org/html/2609.02748","noteUrl":"/review?paper=review-neurips-2026-151310","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151312","title":"Reliable Abstention under Adversarial Injections: Lower Bounds and New Upper Bounds","authors":["Ezra Edelman","Surbhi Goel"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Reliable Abstention under Adversarial Injections: Lower Bounds and New Upper Bounds","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.20111","noteUrl":"/review?paper=review-neurips-2026-151312","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151313","title":"PermuQuant: Lowering Per-Group Quantization Error by Reordering Channels for Diffusion Models","authors":["Yongsen Cheng","Kai Liu","Kaiwen Tao","Junxian Li","Zhixin Wang","Zhikai Chen","Renjing Pei","Yulun Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PermuQuant: Lowering Per-Group Quantization Error by Reordering Channels for Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09503","noteUrl":"/review?paper=review-neurips-2026-151313","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151315","title":"COMPOSE: Composing Future Theorems from Citations and Formal Structure","authors":["David Busbib","Michael Werman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The ideal-power example merges two different monotonicity lemmas","plainExplanation":"The ideal-power illustration combines two distinct monotonicity lemmas into a false disjunction. The following bullets already state the correct lemmas separately.","primaryUrl":"https://arxiv.org/html/2605.30333","noteUrl":"/review?paper=review-neurips-2026-151315","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151318","title":"Understanding Generalization through Decision Pattern Shift","authors":["Huiqi Deng","Yibo Li","Quanshi Zhang","Peng Zhang","Hongbin Pei","Xia Hu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The cosine average counts ordered pairs with an unordered denominator","plainExplanation":"The cosine average counts ordered pairs with an unordered denominator. This is a literal normalization error; an intended unordered-sum convention would supply the local notation repair.","primaryUrl":"https://arxiv.org/html/2605.13148","noteUrl":"/review?paper=review-neurips-2026-151318","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151319","title":"Utility-Constrained Policy Optimization","authors":["Mehrdad Moghimi","Bernardo Avila Pires"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The general budget representation and actor specification need qualification","plainExplanation":"Two initial budgets can reach the same physical state and current stock while imposing different optimal allocations of expected cost across histories. A policy that sees only the current stock cannot distinguish them.","primaryUrl":"https://arxiv.org/html/2606.14029","noteUrl":"/review?paper=review-neurips-2026-151319","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151320","title":"Arena-T2I Hard: Benchmarking and Improving Faithfulness with Dependency-Aware Checklist Rewards","authors":["Yuanhao Ban","Tong Xie","Sohyun An","Yunqi Hong","Evan Frick","I-Hung Hsu","Wei-Lin Chiang","Ion Stoica","Cho-Jui Hsieh"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Arena-T2I Hard: Benchmarking and Improving Faithfulness with Dependency-Aware Checklist Rewards","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.31711","noteUrl":"/review?paper=review-neurips-2026-151320","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151322","title":"GATE-AD: Graph Attention Network Encoding for Few-Shot Industrial Visual Anomaly Detection","authors":["ANGELOS PSYRRIS","Yannis Panagakis","Maria Vakalopoulou","Georgios Th. Papadopoulos"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"GATE-AD: Graph Attention Network Encoding for Few-Shot Industrial Visual Anomaly Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.15300","noteUrl":"/review?paper=review-neurips-2026-151322","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151323","title":"Online Budget Allocation with Censored Semi-Bandit Feedback","authors":["François Bachoc","Nicolò Cesa-Bianchi","Tom Cesari","Roberto Colomboni"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Online Budget Allocation with Censored Semi-Bandit Feedback","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2508.05844","noteUrl":"/review?paper=review-neurips-2026-151323","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151324","title":"Neurosymbolic Object-Centric Learning with Distant Supervision","authors":["Stefano Colamonaco","David Debot","Giuseppe Marra"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neurosymbolic Object-Centric Learning with Distant Supervision","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2506.16129","noteUrl":"/review?paper=review-neurips-2026-151324","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151325","title":"Agent Meltdowns: The Road to Hell Is Paved with Helpful Agents","authors":["Rishi Jha","Harold Triedman","Vitaly Shmatikov","Arkaprabha Bhattacharya"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Agent Meltdowns: The Road to Hell Is Paved with Helpful Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.19149","noteUrl":"/review?paper=review-neurips-2026-151325","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151327","title":"Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems","authors":["Lang Feng","Longtao Zheng","Shuo He","Fuxiang Zhang","Bo An"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The score upper bound cannot imply gradient second-moment inflation","plainExplanation":"An upper bound on score moments cannot establish unconditional gradient-moment inflation. Agent-wise reward normalization also leaves score–advantage dependence uncontrolled, so it does not yield a bound from score statistics alone.","primaryUrl":"https://arxiv.org/html/2602.08847","noteUrl":"/review?paper=review-neurips-2026-151327","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151328","title":"UniFlow-Audio: Unified Flow Matching for Audio Generation from Omni-Modalities","authors":["Xuenan Xu","Jiahao Mei","Ye Tao","Zihao Zheng","Zeyu Xie","Yaoyun Zhang","Haohe Liu","Yuning Wu","Ming Yan","Mengyue Wu","Wen Wu","Chao Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"UniFlow-Audio: Unified Flow Matching for Audio Generation from Omni-Modalities","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.24391","noteUrl":"/review?paper=review-neurips-2026-151328","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151329","title":"MiniPIC: Flexible Position-Independent Caching in <100LOC","authors":["Nathan Ordonez","Thomas Parnell"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FIFO admission does not establish completion dependencies","plainExplanation":"FIFO admission does not establish completion dependencies. FIFO plus a general concurrency cap is insufficient for the scheduling guarantee; this is not a diagnosis of a pinned scheduler implementation.","primaryUrl":"https://arxiv.org/html/2606.13126v1","noteUrl":"/review?paper=review-neurips-2026-151329","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151332","title":"To Call or Not to Call: Diagnosing Intrinsic Over-Calling Bias in LLM Agents","authors":["Wei Shi","Ziheng Peng","Sihang Li","Xiting Wang","Xiang Wang","Mengnan Du","Na Zou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"To Call or Not to Call: Diagnosing Intrinsic Over-Calling Bias in LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18882","noteUrl":"/review?paper=review-neurips-2026-151332","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151333","title":"Few-Step Diffusion Language Models via Trajectory Self-Distillation","authors":["Tunyu Zhang","Xinxi Zhang","Ligong Han","Haizhou Shi","Xiaoxiao He","Zhuowei Li","Hao Wang","Kai Xu","Akash Srivastava","Chengzhi Mao","Hao Wang","Vladimir Pavlovic","Dimitris Metaxas"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A deterministic terminal prior does not determine the model conditional","plainExplanation":"A deterministic terminal prior fixes neither the model’s endpoint conditional nor its total correlation. The data-constant identity holds for the true endpoint posterior or a model constrained to match it, rather than every model.","primaryUrl":"https://arxiv.org/html/2602.12262","noteUrl":"/review?paper=review-neurips-2026-151333","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-151335","title":"Think-with-Rubrics: From External Evaluator to Internal Reasoning Guidance","authors":["Jiachen Yu","Zhihao Xu","junjie wang","Yujiu Yang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Think-with-Rubrics: From External Evaluator to Internal Reasoning Guidance","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07461","noteUrl":"/review?paper=review-neurips-2026-151335","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151336","title":"SurvivalPFN: Amortizing Survival Prediction via In-Context Bayesian Inference","authors":["Shi-ang Qi","Vahid Balazadeh Meresht","Michael Cooper","Russell Greiner","Rahul Krishnan"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Batch event extrema couple censoring to event times","plainExplanation":"Batch event extrema couple censoring to event times. The event-extrema branch violates the construction premise; the independent fixed-horizon branch remains valid.","primaryUrl":"https://arxiv.org/html/2605.15488","noteUrl":"/review?paper=review-neurips-2026-151336","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151337","title":"When Symbol Names Should Not Matter: A Logistic Theory of Fresh-Symbol Classification","authors":["Wenjie Guan","Jelena Bradic"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The softmax extension misses the change in effective binary ridge","plainExplanation":"The softmax extension misses a change in effective ridge, and an optional chaos-quantile bound omits variance terms. The head-pair moment calculation, value-projection expectation, fresh-representative Gram argument and listed collision count each need local corrections. These findings preserve the binary result formulated with exact quantiles and the intended limiting kernel.","primaryUrl":"https://arxiv.org/html/2605.07120","noteUrl":"/review?paper=review-neurips-2026-151337","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-151339","title":"All Roads Lead to Rome: Flow-driven Multi-Anchor Exploration for Open-Environment Active 3D Mapping","authors":["Yang Li","Aming WU","Zihao Zhang","Ziju Han","Sijia Zhang","Yahong Han"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"All Roads Lead to Rome: Flow-driven Multi-Anchor Exploration for Open-Environment Active 3D Mapping","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36889","noteUrl":"/review?paper=review-neurips-2026-151339","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151342","title":"Emergent Low-Rank Training Dynamics in MLPs with Smooth Activations","authors":["Alec Xu","Can Yaras","Matthew Asato","Qing Qu","Laura Balzano"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The subspace intersection has a lower dimension bound rather than an equality","plainExplanation":"The residual nullspace depends on the Gaussian variables generating the masks. The independence-based proof step is unsupported; no claim is made that the final bound cannot have another proof.","primaryUrl":"https://arxiv.org/html/2602.06208","noteUrl":"/review?paper=review-neurips-2026-151342","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151345","title":"Replicable Constrained Bandits","authors":["Matteo Bollini","Gianmarco Genalti","Francesco Emanuele Stradi","Matteo Castiglioni","Alberto Marchesi"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The safety mixture uses safe cost where it needs safe slack","plainExplanation":"The safety mixture uses the safe strategy’s cost where it needs its slack, allowing threshold violations. A separate probability-bound step also requires an impossible margin inequality; correcting its constant does not resolve the distinct random-set conditioning question.","primaryUrl":"https://arxiv.org/html/2602.14580","noteUrl":"/review?paper=review-neurips-2026-151345","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151346","title":"When and What to Prune?  Stage-Aware Visual Token Pruning for Efficient VLA","authors":["TIANJUN SHI","Haotian Xiong","Ziyu Gong","Qi Lu","Li Li"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"When and What to Prune?  Stage-Aware Visual Token Pruning for Efficient VLA","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.05273","noteUrl":"/review?paper=review-neurips-2026-151346","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151348","title":"Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration","authors":["Davide Evangelista","Elena Morotti","Francesco Pivi","Maurizio Gabbrielli"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The probability-flow ODE has an extra noise-scale factor","plainExplanation":"The probability-flow ODE has an extra noise-scale factor. This local ODE typo does not invalidate the subsequent update that already uses the corrected identity.","primaryUrl":"https://arxiv.org/html/2605.12573","noteUrl":"/review?paper=review-neurips-2026-151348","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151350","title":"Progressive Risk Estimation for Accident Anticipation","authors":["Samet Hicsonmez","Eray Çakar","Nermin Samet","Fatma Guney"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Progressive Risk Estimation for Accident Anticipation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.32811","noteUrl":"/review?paper=review-neurips-2026-151350","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151352","title":"Kernel-Gradient Drifting Models","authors":["Maria Esteban-Casadevall","Jorge Carrasco-Pollo","Max Welling","Jan-Willem van de Meent","Erik Bekkers","Floor Eijkelboom"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Kernel-Gradient Drifting Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10727","noteUrl":"/review?paper=review-neurips-2026-151352","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151357","title":"Simple yet Effective Semi-supervised Knowledge Distillation from Vision-Language Models via Dual-Head Optimization","authors":["Seongjae Kang","Dong Bok Lee","Hyungjoon Jang","Sung Ju Hwang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Temperature-one KL control does not justify arbitrary-temperature accuracy","plainExplanation":"The temperature-matching lemma still transfers a KL bound through sharpening without a valid bound on the scaled distributions. The conditional interpolation result is preserved when the required head errors are assumed directly. A separate KL-direction typo remains in the training pseudocode.","primaryUrl":"https://arxiv.org/html/2505.07675","noteUrl":"/review?paper=review-neurips-2026-151357","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","incorrect_bound","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-151363","title":"Distribution Matching Distillation without Fake Score Network","authors":["YoungJoong Kim","Deokyeong Lee","Jaesik Park"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Distribution Matching Distillation without Fake Score Network","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19256","noteUrl":"/review?paper=review-neurips-2026-151363","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151364","title":"Diversity Combining for Multi-Path LLM Reasoning","authors":["Guangsheng Yu","Litianyi Zhang","Qin Wang","Xu Wang","Mingyuan Li","Shaoxiong Ji","Ren Ping Liu","Massimo Piccardi"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Diversity Combining for Multi-Path LLM Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38829","noteUrl":"/review?paper=review-neurips-2026-151364","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151365","title":"ERIS: Enhancing Privacy and Scalability in Federated Learning via Federated Shard Aggregation","authors":["Dario Fenoglio","Pasquale Polverino","Jacopo Quizi","Martin Gjoreski","Akash Dhasade","Marc Langheinrich"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The mini-batch variance constant omits the finite-population factor","plainExplanation":"Corollary C.6 gives \\(\\Gamma_2=(m-b)G^2/(mb)\\) for unbiased uniform mini-batches of size \\(b\\) from \\(m\\) bounded per-sample gradients. Without replacement, the exact mean variance is \\((m-b)\\sigma^2/[b(m-1)]\\), where \\(\\sigma^2=m^{-1}\\sum_s\\|g_s-\\overline g\\|^2\\leq G^2\\).","primaryUrl":"https://arxiv.org/html/2602.08617v3","noteUrl":"/review?paper=review-neurips-2026-151365","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151366","title":"SMOG: Scalable Meta-Learning for Multi-Objective Bayesian Optimization","authors":["Leonard Papenmeier","Petru Tighineanu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"SMOG: Scalable Meta-Learning for Multi-Objective Bayesian Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.22131","noteUrl":"/review?paper=review-neurips-2026-151366","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151368","title":"Simple Extensions of Single-Objective Acquisition Functions and Hedge Strategies for Multi-Objective Bayesian Optimization","authors":["Haris Moazam Sheikh"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Simple Extensions of Single-Objective Acquisition Functions and Hedge Strategies for Multi-Objective Bayesian Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31940","noteUrl":"/review?paper=review-neurips-2026-151368","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151370","title":"The Power of Second Order Methods for Sequence Preconditioning","authors":["Annie Marsden","Elad Hazan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Power of Second Order Methods for Sequence Preconditioning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08390","noteUrl":"/review?paper=review-neurips-2026-151370","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151371","title":"Inference-Time Refinement Closes the Synthetic-Real Gap in Tabular Diffusion","authors":["Eugenio Lomurno","Filippo Balzarini","Francesco Benelle","Francesca Pia Panaccione","Matteo Matteucci"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Zero Chamfer cost identifies support rather than multiplicity","plainExplanation":"Zero Chamfer cost identifies represented support rather than multiset multiplicity. Exact copying of the reference batch is a global minimum, so the bidirectional objective does not structurally prevent it.","primaryUrl":"https://arxiv.org/html/2605.06261","noteUrl":"/review?paper=review-neurips-2026-151371","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151375","title":"$\\texttt{RNAGenScape}$: property-guided, optimized generation of mRNA sequences with manifold Langevin dynamics","authors":["Danqi Liao","Chen Liu","Xingzhi Sun","Dié Tang","Haochen Wang","Scott E Youlten","Srikar Gopinath","Haejeong Lee","Ethan C Strayer","Antonio J Giraldez","Smita Krishnaswamy"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"$\\texttt{RNAGenScape}$: property-guided, optimized generation of mRNA sequences with manifold Langevin dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.24736","noteUrl":"/review?paper=review-neurips-2026-151375","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151378","title":"Not All Tasks Quantize Equally: Fisher-Guided Quantization for Visual Geometry Transformer","authors":["Yipu Zhang","Jintao Cheng","Weilun Feng","Jiehao Luo","Chuanguang Yang","Zhulin An","Yongjun Xu","Wei Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Mean stationarity does not justify dropping correlated quantization error","plainExplanation":"Quantization error depends on the activation and can be correlated with its loss gradient. A zero mean gradient alone does not make their product have zero mean.","primaryUrl":"https://arxiv.org/html/2605.15828","noteUrl":"/review?paper=review-neurips-2026-151378","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151379","title":"Pessimistic Latent Task-aware Optimization for Robust Offline Meta-Reinforcement Learning","authors":["Chang-Hoon Jeong","Kisung Shin","Hanwool Sul","Byoung-Tak Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Two proof steps need stronger justification","plainExplanation":"An upper bound on a positive quantity cannot replace it in the right side of a lower bound. Even if a chosen example attains the value bound, discarding the factor \\(1/(1+\\gamma)\\) requires another argument.","primaryUrl":"https://arxiv.org/html/2610.15428","noteUrl":"/review?paper=review-neurips-2026-151379","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151381","title":"Towards Unified Dynamic Face Landmark Detection","authors":["Sebastian Regalado Lozano","Varshanth R Rao","Ruowei Jiang","Parham Aarabi","Igor Gilitschenski"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A source-local offset cannot supply missing landmark spacing","plainExplanation":"Equation (8) adds a start offset to a source-local landmark index to obtain a normalized unified-template position. Adding an offset preserves adjacent index increments.","primaryUrl":"https://arxiv.org/html/2608.10346","noteUrl":"/review?paper=review-neurips-2026-151381","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151382","title":"How Far Are VLMs from Privacy Awareness in the Physical World? An Empirical Study","authors":["Junran Wang","Xinjie Shen","Zehao Jin","Pan Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"How Far Are VLMs from Privacy Awareness in the Physical World? An Empirical Study","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.05340","noteUrl":"/review?paper=review-neurips-2026-151382","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151384","title":"Step-wise Rubric Rewards for LLM Reasoning","authors":["Weichu Xie","Haozhe Zhao","Wenpu Liu","Yongfu Zhu","Liang Chen","Minghao Ye","Zirong Chen","Yuqi Xu","Shuai Dong","Ziyue Wang","Xinbo Xu","Kean Shi","Ruoyu Wu","Xiaoying Zhang","Wenqi Shao","Baobao Chang","Nan Duan","Jiaqi Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Subtracting a baseline does not justify random standard-deviation scaling","plainExplanation":"The outcome advantage includes a format term despite its accuracy-only description. The unbiasedness argument also treats random standard-deviation scaling as a subtractive baseline, leaving a bias that need not vanish with group size.","primaryUrl":"https://arxiv.org/html/2605.17291","noteUrl":"/review?paper=review-neurips-2026-151384","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-151385","title":"ParaVT: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning","authors":["Zuhao Yang","Kaichen Zhang","Sudong Wang","Keming Wu","Zhongyu Yang","Bo Li","Xiaojuan Qi","Shijian Lu","Xingxuan Li","Lidong Bing"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"ParaVT: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20342","noteUrl":"/review?paper=review-neurips-2026-151385","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151386","title":"Counterfactual Rollout Replay: Forkable Environments as Free Process Rewards for Software Engineering Agents","authors":["Yuanhao li","Hongbo Wang","Xuhong Chen","Yiming Cao","Xunzhu Tang"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Counterfactual Rollout Replay: Forkable Environments as Free Process Rewards for Software Engineering Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33875","noteUrl":"/review?paper=review-neurips-2026-151386","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151387","title":"Phaedra: Learning High-Fidelity Discrete Tokenization for the Physical Sciences","authors":["Levi Lingsch","Georgios Kissas","Johannes Jakubik","Siddhartha Mishra"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Phaedra: Learning High-Fidelity Discrete Tokenization for the Physical Sciences","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.03915","noteUrl":"/review?paper=review-neurips-2026-151387","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151388","title":"How to make the most of your masked language model for protein engineering","authors":["Calvin McCarter","Nick Bhattacharya","Sebastian Ober","Hunter Elliott"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The scalarization signs oppose the stated objective preferences","plainExplanation":"The search prefers lower \\(S=\\log\\sum_i e^{c_if_i}\\) but interprets negative coefficients as minimizing raw objectives and gives positive weight to success probability. Its derivative is \\(\\partial S/\\partial f_i=c_i\\operatorname{softmax}_i\\).","primaryUrl":"https://arxiv.org/html/2603.10302","noteUrl":"/review?paper=review-neurips-2026-151388","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151389","title":"Mechanism-Aware Ensemble Conditioning for Data-Limited Emulation of Extreme Events","authors":["Isabella Thiel","Juan M. Bello-Rivas","Yannis Kevrekidis","Themis Sapsis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Area distributions cannot identify localization or temporal persistence","plainExplanation":"Exceedance area measures spatial extent but discards position. Its reported distribution and frequency summaries also discard temporal order, so they cannot establish localization or persistence.","primaryUrl":"https://arxiv.org/html/2609.30746","noteUrl":"/review?paper=review-neurips-2026-151389","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151396","title":"NASDAQ: Normalized Observation Space Dynamics-Augmented Q-Learning","authors":["Xinwei Liu","Junyuan Liang","Zicong Hong","Jianting Zhang","Wuhui Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"NASDAQ: Normalized Observation Space Dynamics-Augmented Q-Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.21297","noteUrl":"/review?paper=review-neurips-2026-151396","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151397","title":"NestRL: A Nested Training Regime for Mutual Adaptation in Human–AI Teaming","authors":["Upasana Biswas","Durgesh Kalwar","Subbarao Kambhampati","Sarath Sreedharan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"NestRL: A Nested Training Regime for Mutual Adaptation in Human–AI Teaming","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.17737","noteUrl":"/review?paper=review-neurips-2026-151397","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151398","title":"LLM-Auction: Generative Auction towards LLM-Native Advertising","authors":["Chujie Zhao","Qun Hu","Shiping Song","Dagui Chen","Han Zhu","Jian Xu","Bo Zheng"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"LLM-Auction: Generative Auction towards LLM-Native Advertising","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.10551","noteUrl":"/review?paper=review-neurips-2026-151398","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151399","title":"DashAttention: Differentiable and Adaptive Sparse Hierarchical Attention","authors":["Yuxiang Huang","Nuno Gonçalves","Federico Alvetreti","Lei Li","Xu Han","Edoardo Ponti","André Martins","Marcos Treviso"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Routing support changes can make the attention output discontinuous","plainExplanation":"Support-dependent log centering can make the attention output jump when a chunk enters the routing support, contrary to the global differentiability claim. A separate explanation of the negative-KL masking limit has its sign reversed; the masking formula itself is correct.","primaryUrl":"https://arxiv.org/html/2605.18753","noteUrl":"/review?paper=review-neurips-2026-151399","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-151400","title":"PhysTacGen: Physics-Aware Visual-Tactile Sensor Image Generation","authors":["Guo Tang","Yongtao Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PhysTacGen: Physics-Aware Visual-Tactile Sensor Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.08068v1","noteUrl":"/review?paper=review-neurips-2026-151400","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151401","title":"Automatic Constraint Policy Optimization based on Continuous Constraint Interpolation Framework for Offline Reinforcement Learning","authors":["Xinchen Han","Qiuyang Fang","Hossam Afifi","Michel Marot"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The auxiliary suboptimality gap can be negative","plainExplanation":"The parametric policy projection loses a state-dependent normalizer, and the printed SAC objective cancels its own policy dependence. The shaped-advantage functional also need not be maximized by the claimed reference policy, so its suboptimality gap can be negative. A separate denominator regularization step does not upper-bound the original ratio.","primaryUrl":"https://arxiv.org/html/2601.23010","noteUrl":"/review?paper=review-neurips-2026-151401","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151402","title":"EgoTac: In-the-wild Tactile Prediction from Egocentric Vision","authors":["Wenkang Zhang","Chengbo Yuan","Zicheng Zhang","Zhengxue Cheng","Yang Gao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EgoTac: In-the-wild Tactile Prediction from Egocentric Vision","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.15060","noteUrl":"/review?paper=review-neurips-2026-151402","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151403","title":"World Models as Adversaries: Multi-Agent Self-Play Fine-Tuning for Robust Motion Planning","authors":["Tong Nie","Yuewen Mei","Junlin He","Yihong Tang","Jian Sun","Wei Ma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"World Models as Adversaries: Multi-Agent Self-Play Fine-Tuning for Robust Motion Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.10630","noteUrl":"/review?paper=review-neurips-2026-151403","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151405","title":"A Unified Uncertainty Representation for Graph Neural Networks via Doubly-Spectral Stochastic Expansion","authors":["Fred Xu","Thomas Markovich","florence regol","Yizhou Sun"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"A Unified Uncertainty Representation for Graph Neural Networks via Doubly-Spectral Stochastic Expansion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35703v1","noteUrl":"/review?paper=review-neurips-2026-151405","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151406","title":"When and How to Canonize: a Generalization Perspective","authors":["Yonatan Sverdlov","Benjy Friedmann","Snir Hordan","Nadav Dym"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"When and How to Canonize: a Generalization Perspective","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11008","noteUrl":"/review?paper=review-neurips-2026-151406","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151407","title":"One Step is Enough: Multi-Agent Reinforcement Learning Based on One-Step Policy Optimization for Order Dispatch on Ride-Sharing Platforms","authors":["Zijian Zhao","Sen Li"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"One Step is Enough: Multi-Agent Reinforcement Learning Based on One-Step Policy Optimization for Order Dispatch on Ride-Sharing Platforms","plainExplanation":"Uniformity in time does not supply uniformity over initial states of arbitrarily small probability.","primaryUrl":"https://arxiv.org/html/2507.15351","noteUrl":"/review?paper=review-neurips-2026-151407","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","mathematical_error","notation","reporting_inconsistency"],"classifications":["major"]},{"slug":"review-neurips-2026-151409","title":"GNNs Meet Sequence Models Along the Shortest-Path: an Expressive Method for Link Prediction","authors":["Francesco Ferrini","Veronica Lachi","Antonio Longa","Bruno Lepri","Andrea Passerini"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"GNNs Meet Sequence Models Along the Shortest-Path: an Expressive Method for Link Prediction","plainExplanation":"Shortest paths between adjacent vertices can omit common-neighbor information that NCN distinguishes, contradicting the universal expressiveness claim. The preliminary automorphism argument also needs permutations to preserve node features.","primaryUrl":"https://arxiv.org/html/2507.07138","noteUrl":"/review?paper=review-neurips-2026-151409","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","missing_assumption"],"classifications":["critical"]},{"slug":"review-neurips-2026-151410","title":"Toward Optimal Regret in Robust Pricing: Decoupling Corruption and Time","authors":["Kalana K Kalupahana","Francesco Emanuele Stradi","Matteo Castiglioni","Alberto Marchesi"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Toward Optimal Regret in Robust Pricing: Decoupling Corruption and Time","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08290","noteUrl":"/review?paper=review-neurips-2026-151410","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151412","title":"Building Transformation Layers for Riemannian Neural Networks","authors":["Ziheng Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The exponential-map construction needs an inverse-branch restriction","plainExplanation":"The global transformation-layer solution needs a normal-neighborhood restriction so that the logarithm inverts the exponential. Two SPD formulas also need local corrections: log-Cholesky scalar multiplication is not an ordinary matrix power, and the output trace normalization must use the output dimension.","primaryUrl":"https://arxiv.org/html/2609.35436","noteUrl":"/review?paper=review-neurips-2026-151412","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151413","title":"MonoPhysics: Estimating Geometry, Appearance, and Physical Parameters from Monocular Videos","authors":["Daniel Rho","Jun M Choi","Matthew Thornton","Biswadip Dey","Roni Sengupta"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A volume rule that only shrinks cannot enforce nodewise equality","plainExplanation":"The method claims \\(\\sum_iV_iw_{i\\ell}\\) equals cell volume at every node, while its initialization repeatedly multiplies particle volume by corrections capped at one. For unit cell size and one particle with quadratic B-spline weights, any initial volume at most one gives node volumes below one.","primaryUrl":"https://arxiv.org/html/2605.30320","noteUrl":"/review?paper=review-neurips-2026-151413","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151414","title":"CellMSA: Context Modeling for Single-Cell Representation Learning","authors":["Suyuan Zhao","Minghao Liu","Yizhen Luo","Zaiqing Nie"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"CellMSA: Context Modeling for Single-Cell Representation Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38908","noteUrl":"/review?paper=review-neurips-2026-151414","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151416","title":"Online Allocation with Unknown Shared Supply","authors":["Tzeh Y Neoh","XianJun, Davin Choo","Mengchu Yue","Milind Tambe"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Known total demands permit the allocation guarantee ruled out by Proposition 8","plainExplanation":"The first impossibility claim in Proposition 8 is contradicted by a feasible allocation rule using known total demands. Separately, Algorithm 2’s pivot clipping forces the opposite inequality to the one claimed in Lemma 17; the later theorem requires a corrected partition argument.","primaryUrl":"https://arxiv.org/html/2605.07080","noteUrl":"/review?paper=review-neurips-2026-151416","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-151418","title":"Hierarchical Denoising For Multi-Step Visual Reasoning","authors":["Zezhong Qian","Xiaowei Chi","Chak Wing Mak","Tianze Zhou","Ruibin Yuan","Yuhan Rui","Hengzhe Sun","Zhuoqun Wu","Yuming Li","Siyuan Qian","Sirui Han","Shanghang Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The literal exponent and ceiling rule give a different schedule","plainExplanation":"The literal exponent and ceiling rule give a different schedule. The displayed formula/parameter pair is inconsistent; an unreported exponent precision prevents inferring an incorrect actual run budget.","primaryUrl":"https://arxiv.org/html/2607.15278","noteUrl":"/review?paper=review-neurips-2026-151418","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151419","title":"When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction","authors":["Vardhan Dongre","Joseph Hsieh","Viet Lai","Seunghyun Yoon","Trung Bui","Dilek Hakkani-Tur"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12922","noteUrl":"/review?paper=review-neurips-2026-151419","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151420","title":"ModelLens: Finding the Best for Your Task from Myriads of Models","authors":["Rui Cai","Wenjie Mo","Xiaofei Wen","Qiyao Ma","Wenhui Zhu","Xiwen Chen","Muhao Chen","Zhe Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ModelLens: Finding the Best for Your Task from Myriads of Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07075","noteUrl":"/review?paper=review-neurips-2026-151420","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151422","title":"Progressive Residual Warmup for Language Model Pretraining","authors":["Tianhao Chen","Xin Xu","Lu Yin","Hao CHEN","Yang Wang","Shizhe Diao","Can Yang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Progressive Residual Warmup for Language Model Pretraining","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.05369","noteUrl":"/review?paper=review-neurips-2026-151422","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151423","title":"Video-Zero: Self-Evolution Video Understanding","authors":["ruixu zhang","Deyi Ji","Lanyun Zhu","Xuanyi Liu","Yuxin Meng","Ruihang Chu","Yujiu Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Video-Zero: Self-Evolution Video Understanding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14733","noteUrl":"/review?paper=review-neurips-2026-151423","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151424","title":"LatentUM: Unleashing the Potential of Interleaved Cross-Modal Reasoning via a Latent-Space Unified Model","authors":["Jiachun Jin","Zetong Zhou","Xiao Yang","Hao Zhang","Pengfei Liu","Jun Zhu","Zhijie Deng"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"LatentUM: Unleashing the Potential of Interleaved Cross-Modal Reasoning via a Latent-Space Unified Model","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.02097","noteUrl":"/review?paper=review-neurips-2026-151424","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151426","title":"LLM-Based Multi-Agent Blackboard System for Information Discovery in Data Science","authors":["Alireza Salemi","Mihir Parmar","Palash Goyal","Yiwen Song","Jinsung Yoon","Hamed Zamani","Tomas Pfister","Hamid Palangi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LLM-Based Multi-Agent Blackboard System for Information Discovery in Data Science","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.01285","noteUrl":"/review?paper=review-neurips-2026-151426","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151429","title":"D-DOIT: Training-free Adaptation of Discrete Diffusion via Doob's h-Transform","authors":["Jieke Wu","Qijie Zhu","Weimin Wu","Zeqi Ye","Minshuo Chen","Han Liu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"D-DOIT: Training-free Adaptation of Discrete Diffusion via Doob's h-Transform","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04938","noteUrl":"/review?paper=review-neurips-2026-151429","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151432","title":"Systematic Hazard Sampling: Minimal-Variance Inference for Discrete Diffusion and Flow Models","authors":["Seunghwan Jang","Wonje Jeung","SooJean Han"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Systematic Hazard Sampling: Minimal-Variance Inference for Discrete Diffusion and Flow Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.02799","noteUrl":"/review?paper=review-neurips-2026-151432","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151434","title":"StaRPO: Stability-Augmented Reinforcement Policy Optimization","authors":["Jinghan Zhang","Fengran Mo","Tharindu Cyril Weerasooriya","Ruimin Dai","Xiaoyan Han","Yanjie Fu","Dakuo Wang","Kunpeng Liu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Two local mathematical definitions need correction","plainExplanation":"There are \\(K-1\\) defined transitions and only \\(K-2\\) adjacent pairs of them.","primaryUrl":"https://arxiv.org/html/2604.08905","noteUrl":"/review?paper=review-neurips-2026-151434","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_ambiguity","formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151435","title":"VESTA: Visual Exploration with Statistical Tool Agents","authors":["William Rudman","Abhishek Divekar","Kanishk Jain","Sebastian Joseph","Stella Offner","Matt Lease","Kyle Mahowald","Greg Durrett","Junyi Jessy Li"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"VESTA: Visual Exploration with Statistical Tool Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.00384","noteUrl":"/review?paper=review-neurips-2026-151435","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151438","title":"The Pareto Frontier of Randomized Learning-Augmented Online Bidding","authors":["Mathis Degryse","Imrane Zaakour","Christoph Dürr","Spyros Angelopoulos"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Pareto Frontier of Randomized Learning-Augmented Online Bidding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06106","noteUrl":"/review?paper=review-neurips-2026-151438","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151440","title":"GazeFlow: From Human Gaze Behavior to Generative Egocentric Gaze Prediction","authors":["Sheng Zhao","Weikai Lin","Yuhao Zhu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Regression on an encoded video recovers the encoded conditional velocity","plainExplanation":"Regression on an encoded video recovers the encoded conditional velocity. The full-video population recovery needs an additional premise; the encoded regression identity remains valid.","primaryUrl":"https://arxiv.org/html/2609.38519","noteUrl":"/review?paper=review-neurips-2026-151440","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151446","title":"SkillOS: Learning Skill Curation for Self-Evolving Agents","authors":["Siru Ouyang","Jun Yan","Yanfei Chen","Rujun Han","Zifeng Wang","Bhavana Dalvi Mishra","Rui Meng","Chun-Liang Li","Yizhu Jiao","Kaiwen Zha","Maohao Shen","Vishy Tirumalashetty","George Lee","Jiawei Han","Tomas Pfister","Chen-Yu Lee"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SkillOS: Learning Skill Curation for Self-Evolving Agents","plainExplanation":"The reported group sizes are all greater than one, so the denominator \\(|G|-1\\) in the task-outcome reward is nonzero in those configurations. No unjustified proof step, bound, or false invariance was identified in the checked statements.","primaryUrl":"https://arxiv.org/html/2605.06614","noteUrl":"/review?paper=review-neurips-2026-151446","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151448","title":"A Principled Self-Referenced Early Stopping Approach for Deep Image Prior","authors":["Chaoyan Huang","Cheng-Han Huang","Ismail Alkhouri","Rongrong Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mask independence does not make a mask-trained predictor’s holdout risk unbiased","plainExplanation":"The random-mask argument treats a predictor trained on the retained pixels as though its errors were independent of the mask. A coordinatewise interpolator has zero held-out clean risk and positive full clean risk, contradicting the claimed unbiasedness.","primaryUrl":"https://arxiv.org/html/2605.25299","noteUrl":"/review?paper=review-neurips-2026-151448","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151449","title":"Algorithm for Contextual Queueing Bandits with Rate-Optimal Queue Length Regret","authors":["Seoungbin Bae","Dabeen Lee"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Algorithm for Contextual Queueing Bandits with Rate-Optimal Queue Length Regret","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.09668","noteUrl":"/review?paper=review-neurips-2026-151449","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151450","title":"DSSP: Diffusion State Space Policy  with Hierarchical Full-History Conditioning","authors":["Zhiyuan Guan","Jianshu Hu","Han Fang","Yunpeng Jiang","Yize Huang","Shujia Li","Xiao Li","Yutong Ban"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The denoising Bayes risk must condition on the noisy action too","plainExplanation":"Equation (9) predicts the clean action using \\(x_\\tau,\\tau\\) as well as \\(X\\), so its Bayes risk is \\(\\mathbb E\\operatorname{tr}\\operatorname{Cov}(x_0\\mid x_\\tau,\\tau,X)\\). With constant \\(X\\), a fair binary action \\(A=\\pm1\\) has unconditional variance one. Observing \\(Y=A+\\epsilon\\) with standard Gaussian noise gives posterior mean \\(\\tanh Y\\) and strictly smaller MSE. Omitting that noisy observation changes the optimum. Nested history classes can still give weak Bayes-risk dominance. The printed equality, rather than that general dominance principle, is wrong.","primaryUrl":"https://arxiv.org/html/2605.14598v2","noteUrl":"/review?paper=review-neurips-2026-151450","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151451","title":"DICE: Decoupling Capability from Intervention Necessity in LLM Tutoring","authors":["Sayantan Pal","Kaiyi Ji","Rohini Srihari"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A difficulty-only explanation need not predict a monotone trend","plainExplanation":"The difficulty-ladder argument assumes monotonicity that does not follow from a difficulty-only explanation. The weighted training loss also uses different supervision normalizations in Eq. (2) and Algorithm 1 while retaining the same KL coefficient, changing the objective.","primaryUrl":"https://arxiv.org/html/2610.04825","noteUrl":"/review?paper=review-neurips-2026-151451","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151460","title":"Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection","authors":["Xiaona Zhou","Tianjiao (Joey) Yu","Muntasir Wahed","Constantin Brif","Ismini Lourentzou"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.30344","noteUrl":"/review?paper=review-neurips-2026-151460","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151461","title":"GeoPair: Geometry-Preserving Cross-Layer Factorization for Training-Free Transformer Compression","authors":["Baher Mohammad","Ammar Ali","Stamatios Lefkimmiatis"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Regularized linear algebra is not an exact solve of the displayed unregularized objective","plainExplanation":"The hard-thresholding step can increase its quadratic objective. The general descent guarantee fails; this does not diagnose convergence of a particular trained implementation.","primaryUrl":"https://arxiv.org/html/2609.25963","noteUrl":"/review?paper=review-neurips-2026-151461","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151464","title":"Structuring Open-Ended NAS: Semi-Automated Design Knowledge Structuring with LLMs for Efficient Neural Architecture Search","authors":["Yuiko Sakuma","Masakazu Yoshimura","Marcel Gröpl","Zitang Sun","Junji Otsuka","Atsushi Irie","Takeshi Ohashi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Structuring Open-Ended NAS: Semi-Automated Design Knowledge Structuring with LLMs for Efficient Neural Architecture Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19247","noteUrl":"/review?paper=review-neurips-2026-151464","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151465","title":"Diffusion Guidance Is a Controllable Policy Improvement Operator","authors":["Kevin Frans","Seohong Park","Pieter Abbeel","Sergey Levine"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Diffusion Guidance Is a Controllable Policy Improvement Operator","plainExplanation":"Stronger fixed-reference guidance can lower discounted return, and the claimed Gibbs optimizer solves a local surrogate rather than the full control objective. The sampling pseudocode also misprints a step size that is correct in the author sampler.","primaryUrl":"https://arxiv.org/html/2505.23458","noteUrl":"/review?paper=review-neurips-2026-151465","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","formula_error","invalid_inference"],"classifications":["critical"]},{"slug":"review-neurips-2026-151468","title":"Dynamic Cross-Modal Prompt Generation for Multimodal Continual Instruction Tuning","authors":["Tao Hu","Da-Wei Zhou"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Dynamic Cross-Modal Prompt Generation for Multimodal Continual Instruction Tuning","plainExplanation":"No issue was found in the checked claims. The earlier two-dimensional full-rank counterexample violates the new dimension premise. Under the revised premise, the cap cannot prevent the specified threshold from being reached.","primaryUrl":"https://arxiv.org/html/2605.10765","noteUrl":"/review?paper=review-neurips-2026-151468","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151471","title":"Towards Explainable Industrial Anomaly Detection via Knowledge-Guided Latent Reasoning","authors":["Peng Chen","Chao Huang","Yunkang Cao","Chengliang Liu","Wei Wang","wenqiang wang","Mingbo Yang","Li Shen","Wenqi Ren","XIAOCHUN CAO"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Towards Explainable Industrial Anomaly Detection via Knowledge-Guided Latent Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.09850","noteUrl":"/review?paper=review-neurips-2026-151471","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151472","title":"Accelerating Video Inverse Problem Solvers with Autoregressive Diffusion Models","authors":["Taesung Kwon","Jonghyun Park","Hyungjin Chung","Jong Chul Ye"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Prefix error must be related to the cache norm used by the stability bound","plainExplanation":"The recursion measures a latent-prefix mismatch \\(\\delta_n\\), but the velocity Lipschitz assumption controls a separately defined cache/context norm. Take prefix mismatch one, cache map \\(c=100\\,\\mathrm{prefix}\\), velocity \\(v=c\\), and one step of size \\(0.1\\) from identical current states.","primaryUrl":"https://arxiv.org/html/2605.20624","noteUrl":"/review?paper=review-neurips-2026-151472","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151473","title":"Understanding Gradient Orthogonalization for Deep Learning via Non-Euclidean Trust-Region Optimization","authors":["Dmitry Kovalev"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Understanding Gradient Orthogonalization for Deep Learning via Non-Euclidean Trust-Region Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2503.12645","noteUrl":"/review?paper=review-neurips-2026-151473","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151474","title":"ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation","authors":["Yu Zhang","YIDI SHAO","Wenqi Ouyang","Yushi LAN","Zhexin Liang","Chengrui Wu","Xudong XU","Xingang Pan"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27852","noteUrl":"/review?paper=review-neurips-2026-151474","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151477","title":"CitySTAR: Agent-Driven Structured and Topology-Aware Reasoning for Open-Vocabulary Urban 3D Grounding","authors":["Shuai Zhang","Hongye Hou","qinghe liu","Zhuoxiao Li","Dongli Wu","Jing OU","Yuan Liu","Wufan Zhao"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"CitySTAR: Agent-Driven Structured and Topology-Aware Reasoning for Open-Vocabulary Urban 3D Grounding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.19911v1","noteUrl":"/review?paper=review-neurips-2026-151477","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151478","title":"Data-Adaptive Mahalanobis Metric Learning for Cross-Head Attention in Transformers","authors":["Zhenke Duan","Xin Li","Jiqun Pan","Xiaofei Dong","Hanwen Ning","Xinyuan Song"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The printed generator dimensions give a much larger break-even depth","plainExplanation":"The expressivity theorem assumes full-dimensional effective matrices that differ from the fixed-slice architecture used for the complexity claim. The parameter comparison also understates the dense generator cost, and the full-layer complexity omits dense value/output work when head count varies.","primaryUrl":"https://arxiv.org/html/2610.13261","noteUrl":"/review?paper=review-neurips-2026-151478","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151480","title":"MicroWorld: Empowering Multimodal Large Language Models to Bridge the Microscopic Domain Gap with Multimodal Attribute Graph","authors":["Manyu Li","Ruian He","Chenxi Ma","Weimin Tan","Bo Yan"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"MicroWorld: Empowering Multimodal Large Language Models to Bridge the Microscopic Domain Gap with Multimodal Attribute Graph","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10120","noteUrl":"/review?paper=review-neurips-2026-151480","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151481","title":"Can In-Context Learning Support Intrinsic Curiosity?","authors":["Eric Elmoznino","Sangnie Bhardwaj","Johannes von Oswald","Rajai Nasser","Blaise Aguera y Arcas","João Sacramento","Rif A. Saurous","Guillaume Lajoie"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"An observational future action can reveal a hidden environment parameter","plainExplanation":"An observational future action can reveal a hidden environment parameter. The observational posterior claim is false; an explicitly interventional formulation could retain the intended prior invariance.","primaryUrl":"https://arxiv.org/html/2606.19476","noteUrl":"/review?paper=review-neurips-2026-151481","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151483","title":"Consistency-Preserving Concept Erasure via Unsafe–Safe Pairing and Directional Fisher-weighted Adaptation","authors":["Yongwoo Kim","Sungmin Cha","Hyunsoo Kim","Jaewon Lee","Donghyun Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Column normalization requires independent column projectors","plainExplanation":"Column normalization requires independent column projectors. The whole-matrix derivative display is wrong; a shorthand columnwise implementation could already use the correct operation.","primaryUrl":"https://arxiv.org/html/2602.05339","noteUrl":"/review?paper=review-neurips-2026-151483","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151484","title":"Parabolic Position Encoding: Vision-Centric, Principled, Extrapolatable, General","authors":["Christoffer Koo Ohrstrom","Rafael I Cabral Muchacho","Yifei Dong","Filippos Moumtzidellis","Ronja Güldenring","Florian T. Pokorny","Lazaros Nalpantidis"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Parabolic Position Encoding: Vision-Centric, Principled, Extrapolatable, General","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.01418","noteUrl":"/review?paper=review-neurips-2026-151484","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151486","title":"Dependency-Guided Parallel Decoding in Discrete Diffusion Language Models","authors":["Liran Ringel","Ameen A Ali","Yaniv Romano"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The common-joint total-variation dependency is symmetric","plainExplanation":"The common-joint total-variation dependency is symmetric. The mathematical common-joint claim is wrong; asymmetry in practical estimated dependencies remains possible.","primaryUrl":"https://arxiv.org/html/2604.02560","noteUrl":"/review?paper=review-neurips-2026-151486","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151487","title":"RAO-Nav: Probing Omni-Language Models for Zero-shot Semantic Audio-Visual Navigation","authors":["Qilang Ye","Meng Liu","Yu Zhou"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"RAO-Nav: Probing Omni-Language Models for Zero-shot Semantic Audio-Visual Navigation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.32224","noteUrl":"/review?paper=review-neurips-2026-151487","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151488","title":"MC^2: Monte Carlo Correction for Fast Elliptic PDE Solving","authors":["Ethan Hsu","Ivan Ge","Hong M Yam"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stated biharmonic boundary data do not identify a unique solution","plainExplanation":"The stated biharmonic boundary data do not identify a unique solution. The stated dense-field boundary-value input is underdetermined. A manufactured reference and correction task can still be well defined; ordinary Laplace Dirichlet uniqueness is unaffected.","primaryUrl":"https://arxiv.org/html/2605.09288","noteUrl":"/review?paper=review-neurips-2026-151488","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151489","title":"Code2World: A GUI World Model via Renderable Code Generation","authors":["Yuhao Zheng","Li&amp;#x27;an Zhong","Yi Wang","Rui Dai","Kaikui Liu","Xiangxiang Chu","Linyuan Lü","Philip Torr","Kevin Qinghong Lin"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Code2World: A GUI World Model via Renderable Code Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09856","noteUrl":"/review?paper=review-neurips-2026-151489","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151491","title":"Structured Unitary Tensor Network Representations for Circuit-Efficient Quantum Data Encoding","authors":["Guang Lin","Toshihisa Tanaka","Qibin Zhao"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Structured Unitary Tensor Network Representations for Circuit-Efficient Quantum Data Encoding","plainExplanation":"No issue was found in the checked claims. The former sign reversal has been corrected. The surrounding Euler decomposition also uses independent angles.","primaryUrl":"https://arxiv.org/html/2602.16266","noteUrl":"/review?paper=review-neurips-2026-151491","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151496","title":"Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment","authors":["Youngjoon Jang","Liliane Momeni","Zifan Jiang","Joon Son Chung","Gul Varol","Andrew Zisserman"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.08040","noteUrl":"/review?paper=review-neurips-2026-151496","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151497","title":"Efficiently Representing Algorithms With Chain-of-Thought Transformers","authors":["Yanhong Li","Anej Svete","Ashish Sabharwal","Will Merrill"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The binary reconstruction recurrence is off by a factor of two","plainExplanation":"The binary reconstruction recurrence is off by a factor of two. The local recurrence is wrong; its constant-factor repair preserves asymptotic simulation resources.","primaryUrl":"https://arxiv.org/html/2606.19697","noteUrl":"/review?paper=review-neurips-2026-151497","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151498","title":"Deep Learning-based Algebraic Reynolds Stress Closures for RANS simulations of Turbulent Flows","authors":["Daniel Dehtyriov","Jonathan F. MacArt","Justin Sirignano"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Convergent fixed-point iteration does not imply nonsingular residual Jacobians","plainExplanation":"The appendix infers invertibility of \\(I-DG\\), and validity of a Neumann/implicit differentiation argument, from convergence of the forward solver. For \\(G(x)=x-x^3\\), every initial point in \\((-1,1)\\) converges to zero, since the magnitude strictly decreases away from zero.","primaryUrl":"https://arxiv.org/html/2605.26358","noteUrl":"/review?paper=review-neurips-2026-151498","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151500","title":"ACER: Towards Generalizable Protein-ligand Co-folding","authors":["Nopsinth Vithayapalert","Francesca Grisoni"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"ACER: Towards Generalizable Protein-ligand Co-folding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://www.biorxiv.org/content/10.64898/2026.06.02.728568v2.full.pdf","noteUrl":"/review?paper=review-neurips-2026-151500","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151501","title":"From Cortical Synchronous Rhythm to Brain Inspired Learning Mechanism: An Oscillatory Spiking Neural Network with Time-Delayed Coordination","authors":["Tingting Dan","Guorong Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The full oscillator dynamics include terms missing from the Lyapunov derivative","plainExplanation":"The potential decreases along the frozen, drift-free gradient subsystem, but the full oscillator dynamics also contain intrinsic-frequency drift and possible explicit time dependence. Those missing derivative terms can make the potential increase, contradicting the global Lyapunov claim.","primaryUrl":"https://arxiv.org/html/2605.01656","noteUrl":"/review?paper=review-neurips-2026-151501","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151502","title":"Chaining 2-FWL GNNs for Combinatorial Graph Alignment","authors":["Marc Lelarge"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Chaining 2-FWL GNNs for Combinatorial Graph Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.03086","noteUrl":"/review?paper=review-neurips-2026-151502","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151503","title":"SkiP: When to Skip and When to Refine for Efficient Robot Manipulation","authors":["Mingtong Dai","Guanqi Peng","Yongjie Bai","Feng yan","chunjie chen","Lingbo Liu","Liang Lin","Xinyu Wu"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The terminal masked-loss sample has a zero denominator","plainExplanation":"The terminal masked-loss sample has a zero denominator. The loss has an admitted boundary-definition gap; this is not a verified training crash.","primaryUrl":"https://arxiv.org/html/2605.15536","noteUrl":"/review?paper=review-neurips-2026-151503","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151504","title":"Rethinking Long-Video Efficiency: A Joint Allocation Perspective on Frames, Pixels, and Front-End Latency","authors":["Sixun Dong","Wei Li","Andong Deng","Qi Qian","Victor Zhu","Zhengping Ji","Chen Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Constant frame similarities make the quality normalization undefined","plainExplanation":"Constant frame similarities make the quality normalization undefined. The mathematical selector omits a boundary case; no empirical incidence is claimed.","primaryUrl":"https://arxiv.org/html/2610.04318","noteUrl":"/review?paper=review-neurips-2026-151504","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151509","title":"Structure-aware Reinforcement Learning for Protein Directed Evolution","authors":["Zikun Nie","Suyuan Zhao","Yizhen Luo","Siqi Fan","Zaiqing Nie"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Anchored norm preservation does not imply pairwise Lipschitz continuity","plainExplanation":"Anchored norm preservation does not imply pairwise Lipschitz continuity. The radial statement can hold under the paper’s anchored definition, but its pairwise Lipschitz interpretation does not follow.","primaryUrl":"https://arxiv.org/html/2609.39048","noteUrl":"/review?paper=review-neurips-2026-151509","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151510","title":"The SEC Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data","authors":["Nick Bettencourt","Xiaowei Ding","Kay Giesecke"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"The SEC Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.18192","noteUrl":"/review?paper=review-neurips-2026-151510","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151511","title":"EasyLens: A Training-Free Plug-and-Play Subtle-Lesion Representation Amplifier for Medical Vision-Language Models","authors":["QIWEI ZENG","Hao Wang","Jinghao Lin","Shuchang Ye","Yuezhe Yang","Yige Peng","Haoyuan Che","Jinman Kim","Lei Bi"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"EasyLens: A Training-Free Plug-and-Play Subtle-Lesion Representation Amplifier for Medical Vision-Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06379","noteUrl":"/review?paper=review-neurips-2026-151511","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151513","title":"The Attacker in the Mirror: Breaking Self-Consistency in Safety via Anchored Bipolicy Self-Play","authors":["Gabriele La Malfa","Emanuele La Malfa","Saar Cohen","Jie Zhang","Michael Luck","Michael Wooldridge","Elizabeth Black"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The reward-range bound misses a factor under the printed KL convention","plainExplanation":"The attacker and defender objectives are exact negatives under their shared parameter-dependent law, so the claimed nonzero sum of total gradients does not follow. A separate KL-to-value bound uses an insufficient constant for rewards in $[-1,1]$ and the paper’s half-relative-entropy convention.","primaryUrl":"https://arxiv.org/html/2605.08427","noteUrl":"/review?paper=review-neurips-2026-151513","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151516","title":"Outlier-robust Diffusion Posterior Sampling for Bayesian Inverse Problems","authors":["Yiming Yang","Xiaoyuan Cheng","Yi He","Kaiyu Li","Wenxuan Yuan","Zhuo Sun"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The conditional covariance identity fails even for a deterministic prior","plainExplanation":"The robustness condition needs uniform rather than pointwise boundedness, and the posterior covariance/Jacobian identities fail on a deterministic prior. Figure 4 also reverses the likelihood-loss sign and omits the sampled Gaussian innovation required by its reverse-SDE interpretation.","primaryUrl":"https://arxiv.org/html/2602.02045","noteUrl":"/review?paper=review-neurips-2026-151516","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151517","title":"Mirror descent actor-critic methods for entropy regularised MDPs in general spaces: stability and convergence","authors":["Denis Zorba","David Siska","Lukasz Szpruch"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A local proof repair preserves the stability theorem","plainExplanation":"The displayed argument uses the looser bound with twice the centered-log-density norm and an incorrect square expansion. Its resulting contraction coefficient is not justified by all step sizes allowed in the theorem. Centering supplies a sharper inequality that repairs the argument.","primaryUrl":"https://arxiv.org/html/2602.10838","noteUrl":"/review?paper=review-neurips-2026-151517","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["proof_gap"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151519","title":"SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance","authors":["Xinyue Zeng","Jiawei zhang","Yujun Yan","Dawei Zhou"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.30192","noteUrl":"/review?paper=review-neurips-2026-151519","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151520","title":"DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging","authors":["Krishna Khadka","Yu Lei","Raghu Kacker","D. R Kuhn"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The one-pass search does not guarantee 1-minimality","plainExplanation":"An additive linear head can make a previously necessary unit redundant after another deletion, so a single pass need not return a 1-minimal set. A necessary unit can also have zero target-logit drop, making its normalized importance undefined.","primaryUrl":"https://arxiv.org/html/2602.19274","noteUrl":"/review?paper=review-neurips-2026-151520","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","definition_error","false_statement","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-151522","title":"Submodular Benchmark Selection","authors":["Alexander Smola"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Submodular Benchmark Selection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.02209","noteUrl":"/review?paper=review-neurips-2026-151522","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151525","title":"NoTVLA: Semantics-Preserving Robot Adaptation via Narrative Action Interfaces","authors":["Zheng Huang","Mingyu Liu","Xiaoyi Lin","Muzhi Zhu","Ye Lin","Canyu Zhao","Zongze Du","Xiaoman Li","Hao Zhong","Yiduo Jia","Hao Chen","Chunhua Shen"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"NoTVLA: Semantics-Preserving Robot Adaptation via Narrative Action Interfaces","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.03895","noteUrl":"/review?paper=review-neurips-2026-151525","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151528","title":"CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting","authors":["Pengpeng Yu","Runqing Jiang","Qi Zhang","Dingquan Li","Jing Wang","Yulan Guo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25563","noteUrl":"/review?paper=review-neurips-2026-151528","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151529","title":"REFLEX: Reflective Evolution from LLM Experience","authors":["Pan Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"REFLEX: Reflective Evolution from LLM Experience","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.16496","noteUrl":"/review?paper=review-neurips-2026-151529","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151530","title":"Online Differentially Private Consistent Clustering","authors":["Edith Cohen","Vadym Doroshenko","Badih Ghazi","Pritish Kamath","Alexander Knop","Ravi Kumar","Ethan Leeman","Pasin Manurangsi","Adam Sealfon","Marika Swanberg"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Online Differentially Private Consistent Clustering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.27896","noteUrl":"/review?paper=review-neurips-2026-151530","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151532","title":"On the Construction and Implications of Low-Loss Valleys in LoRA-based Bayesian Inference","authors":["Daniel Dold","Emanuel Sommer","Julius Kobialka","Oliver Dürr","David Rügamer"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"On the Construction and Implications of Low-Loss Valleys in LoRA-based Bayesian Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29580","noteUrl":"/review?paper=review-neurips-2026-151532","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151533","title":"Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision","authors":["Yihao Xue","Allan Zhang","Jianhao Huang","Amit Sahai","Baharan Mirzasoleiman"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.00927","noteUrl":"/review?paper=review-neurips-2026-151533","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151535","title":"D$^2$Quant: Accurate Low-bit Post-Training Weight Quantization for LLMs","authors":["Xianglong Yan","chengzhu bao","Zhiteng Li","Tianao Zhang","Shaoqiu Zhang","Ruobing Xie","Xingwu Sun","Yulun Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The SNR convention does not yield the stated relative MSE reduction","plainExplanation":"The explanation links mean-subtraction MSE reduction to \\(SNR/(1+SNR)\\), while its SNR definition uses \\(|\\mu|/\\mathrm{variance}\\). For deviations with mean \\(\\mu\\) and variance \\(v\\), subtracting the mean changes MSE from \\(\\mu^2+v\\) to \\(v\\), giving relative reduction \\(\\mu^2/(\\mu^2+v)\\).","primaryUrl":"https://arxiv.org/html/2602.02546","noteUrl":"/review?paper=review-neurips-2026-151535","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151537","title":"Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images","authors":["Zijun Gao","Chunbin Gu","Xiangde Luo","Jinxi Xiang","Pheng-Ann Heng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36429","noteUrl":"/review?paper=review-neurips-2026-151537","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151538","title":"Higher-Order Action Supervision Makes A Strong Policy Class","authors":["Peng Cheng","Yunxian Hou","Zhi Zhou","Qian Zhang","Chang Huang","Xianyuan Zhan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Higher-Order Action Supervision Makes A Strong Policy Class","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.11175","noteUrl":"/review?paper=review-neurips-2026-151538","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151540","title":"Dual Dimensionality for Local and Global Attention","authors":["Zhiyuan Wang","Xuan Luo","Sirui Zeng","Xifeng Yan"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Dual Dimensionality for Local and Global Attention","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.18587v1","noteUrl":"/review?paper=review-neurips-2026-151540","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151541","title":"One-Shot Private Confidence Regions via Resampling","authors":["Po-Ling Loh","Debepsita Mukherjee","Shourya Pandey","Purnamrita Sarkar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The clipping-radius example has an extra noise-scale factor","plainExplanation":"The example’s Euclidean clipping bound contains an extra noise-scale factor. Coordinate clipping gives a radius linear in \\(\\sigma_{\\rm proxy}\\), not quadratic in it. This is a scale typo in an example. The general sensitivity theorem is not refuted, but a literal calibration using the incorrect radius would need correction.","primaryUrl":"https://arxiv.org/html/2610.08460","noteUrl":"/review?paper=review-neurips-2026-151541","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151542","title":"FLAG: Flow Policy MaxEnt-RL by Latent Augmented Guidance","authors":["Sungha Kim","Gawon Lee","Jusuk Lee","Jonghae Park","H. Jin Kim","Daesol Cho"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The main KL rate needs the appendix’s tail assumptions","plainExplanation":"Smoothness alone does not control the relative smoothing error in low-density tails. The main statement omits the extra assumptions that its appendix explicitly uses to justify the integrated rate.","primaryUrl":"https://arxiv.org/html/2605.30749","noteUrl":"/review?paper=review-neurips-2026-151542","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["missing_assumption"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151544","title":"Variance-Optimal State Resampling for Reinforcement Learning with Verifiable Rewards","authors":["Yibo Wang","Yuanxin Liu","Zhaoran Wang"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Variance-Optimal State Resampling for Reinforcement Learning with Verifiable Rewards","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.15888","noteUrl":"/review?paper=review-neurips-2026-151544","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151546","title":"Global Convergence of Four-Layer Matrix Factorization under Random Initialization","authors":["Minrui Luo","Weihang Xu","Xiang Gao","Maryam Fazel","Simon Du"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Global Convergence of Four-Layer Matrix Factorization under Random Initialization","plainExplanation":"The arc-to-chord conversion in Eq. (44) reverses the event inclusion: the chord length is smaller than the arc length.","primaryUrl":"https://arxiv.org/html/2511.09925","noteUrl":"/review?paper=review-neurips-2026-151546","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151553","title":"SetAD: Semi-Supervised Anomaly Learning in Contextual Sets","authors":["Jianling Gao","Chongyang Tao","Xuelian Lin","Junfeng Liu","Shuai Ma"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SetAD: Semi-Supervised Anomaly Learning in Contextual Sets","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.07863","noteUrl":"/review?paper=review-neurips-2026-151553","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151556","title":"Concepts in Motion: Temporal Concept Bottleneck Model for Interpretable Video Classification","authors":["Patrick Knab","Sascha Marton","Philipp J Schubert","Drago A Guggiana Nilo","Christian Bartelt"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Concepts in Motion: Temporal Concept Bottleneck Model for Interpretable Video Classification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.20899","noteUrl":"/review?paper=review-neurips-2026-151556","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151558","title":"Where Does Warm-Up Come From? Adaptive Scheduling for Norm-Constrained Optimizers","authors":["Artem Riabinin","Andrey Veprikov","Arman Bolatov","Martin Takac","Aleksandr Beznosikov"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Where Does Warm-Up Come From? Adaptive Scheduling for Norm-Constrained Optimizers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.05813","noteUrl":"/review?paper=review-neurips-2026-151558","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151561","title":"Capacity-Constrained Online Convex Optimization with Delayed Feedback","authors":["Alexander Ryabchenko","Idan Attias","Dan Roy"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The first-order regret bound needs a strongly convex regularizer","plainExplanation":"Squared infinity-norm regularization has flat directions, and a legal choice of the stated update can incur linear regret without delays. The explicit counterexample uses the corrected tie rule. Restricting the squared-norm update to inner-product geometry restores the missing stability premise; a general-norm version needs an appropriate strongly convex regularizer. The Euclidean bandit specialization is not refuted.","primaryUrl":"https://arxiv.org/html/2606.11711","noteUrl":"/review?paper=review-neurips-2026-151561","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151563","title":"Diffusion Models Observe Only Gradients: A Geometric Perspective on Score Matching Errors","authors":["Nail B Khelifa","Richard Turner","Ramji Venkataramanan"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Diffusion Models Observe Only Gradients: A Geometric Perspective on Score Matching Errors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06179","noteUrl":"/review?paper=review-neurips-2026-151563","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151565","title":"Auto-Rubric as Reward: From Implicit Preference to Explicit Generative Criteria","authors":["Juanxi Tian","Fengyuan Liu","Jiaming Han","Yilei Jiang","Yongliang Wu","Yesheng Liu","Haodong Li","Furong Xu","Wanhua Li"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Descent on the positive PPO surrogate lowers the rewarded action probability","plainExplanation":"Descent on the positive PPO surrogate lowers the rewarded action probability. The printed optimization sign is wrong; it does not establish the sign used in executable training.","primaryUrl":"https://arxiv.org/html/2605.08354","noteUrl":"/review?paper=review-neurips-2026-151565","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151566","title":"Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers","authors":["Jing Xu","Cunjian Chen","Qiuhong Ke"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.11237","noteUrl":"/review?paper=review-neurips-2026-151566","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151567","title":"Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents","authors":["Hao Wang","Guozhi Wang","Han Xiao","YufengZhou","Yue Pan","Jichao Wang","Ke Xu","yafei wen","XIAOHU RUAN","xiaoxin chen","Honggang Qi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.10674","noteUrl":"/review?paper=review-neurips-2026-151567","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151571","title":"HiLoRA: Adaptive Hierarchical LoRA Routing for Training-Free Domain Generalization","authors":["Ziyi Han","Huanyu Wang","Zeyu Zhang","Xiangxiang Dai","Xutong Liu","John C. S. Lui"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The Gaussian score needs a nonsingular estimator and a consistent coefficient formula","plainExplanation":"A covariance estimated from \\(m\\) centered observations has rank at most \\(m-1\\). Therefore a full nonsingular Gaussian density is not defined when the feature dimension is at least \\(m\\), unless an explicit regularization or dimensionality reduction is supplied.","primaryUrl":"https://arxiv.org/html/2510.12266","noteUrl":"/review?paper=review-neurips-2026-151571","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151573","title":"Online Bayesian Calibration under Gradual and Abrupt System Changes","authors":["Yang Xu","Chiwoo Park"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Online Bayesian Calibration under Gradual and Abrupt System Changes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06612","noteUrl":"/review?paper=review-neurips-2026-151573","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151574","title":"One for All: A Non-Linear Transformer can enable Cross-Domain Generalization for In-Context Reinforcement Learning","authors":["Bowen He","Juncheng Dong","Lin Lin","Xiang Cheng"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A shared value offset does not cancel from discounted TD","plainExplanation":"A shared value offset does not cancel from discounted TD. The explanation of discounted bias cancellation is false; Theorem 4.1 and convergence of the actual internal TD losses are not refuted.","primaryUrl":"https://arxiv.org/html/2605.09727","noteUrl":"/review?paper=review-neurips-2026-151574","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151578","title":"Efficient Neural Field Learning via Adaptive Coverage and Focused Sampling","authors":["Guang Zhao","Xihaier Luo","Huan-Hsin Tseng","Seungjun Lee","Shinjae Yoo","Yihui Ren","Wei Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Efficient Neural Field Learning via Adaptive Coverage and Focused Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02410","noteUrl":"/review?paper=review-neurips-2026-151578","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151579","title":"Optimal Ansatz-free Hamiltonian Learning In Situ","authors":["Taiqi Zhou","Weiyuan Gong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Optimal Ansatz-free Hamiltonian Learning In Situ","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19486","noteUrl":"/review?paper=review-neurips-2026-151579","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151581","title":"Query Lower Bounds for Diffusion Sampling","authors":["Zhiyang Xun","Eric Price"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Query Lower Bounds for Diffusion Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.10857","noteUrl":"/review?paper=review-neurips-2026-151581","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151583","title":"TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs","authors":["Zihan Lin","Xiaohan Wang","Jie Cao","Jiajun Chai","Wei Lin","Guojun Yin","Ran He"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33589","noteUrl":"/review?paper=review-neurips-2026-151583","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151588","title":"Generalized Intention Modeling in Multi-Agent Reinforcement Learning","authors":["Mateusz Odrowaz-Sypniewski","Jasmine Bayrooti","Ajay Shankar","Amanda Prorok"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The entropy argument needs a representation-independent constant and a discrete domain","plainExplanation":"This inequality is valid for discrete Shannon entropy, but it has no general counterpart for differential entropy. Scaling and dimension reduction can reverse the inequalities used in the proof.","primaryUrl":"https://arxiv.org/html/2605.31318","noteUrl":"/review?paper=review-neurips-2026-151588","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151589","title":"Deep Learning as Neural Low-Degree Filtering: A Spectral Theory of Hierarchical Feature Learning","authors":["Yatin Dandi","Matteo Vilucchio","Luca Arnaboldi","Hugo Tabanelli","Florent Krzakala"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A bounded sign teacher is invisible to the second-order recovery matrix","plainExplanation":"The stacked linearized dynamics replace neuron-specific growth rates with one shared spectral filter. Proposition 2 also overstates second-order recovery: the bounded sign teacher has information exponent one, yet its second-order empirical matrix is rotationally invariant and cannot recover the fixed teacher direction.","primaryUrl":"https://arxiv.org/html/2605.13612","noteUrl":"/review?paper=review-neurips-2026-151589","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-151594","title":"WASP: Weakly Aligned Spatiotemporal Pairs for Fetal Brain MRI-Ultrasound Learning","authors":["Francesco Correnti","Gabriele Magrini","Marco Mistretta","Niccolò Biondi","Pietro Pala","Alessandro Ramalli","simona fiori","Andrew Bagdanov","Matteo Lenge"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"WASP: Weakly Aligned Spatiotemporal Pairs for Fetal Brain MRI-Ultrasound Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04601","noteUrl":"/review?paper=review-neurips-2026-151594","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151595","title":"Learning to target with network interference","authors":["Xiaomeng Wang","Hamsa Bastani","Osbert Bastani","Zhimei Ren"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The exploration horizon can exceed the total horizon","plainExplanation":"The exploration horizon can exceed the total horizon. The policy definition and proof need a cap or stronger condition; the rate itself is not disproved.","primaryUrl":"https://arxiv.org/html/2605.27794","noteUrl":"/review?paper=review-neurips-2026-151595","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151597","title":"OpenWebRL: Demystifying Online Multi-turn Reinforcement Learning for Visual Web Agents","authors":["Rui Yang","Qianhui Wu","Yuxi Chen","Hao Bai","Wenlin Yao","Hao Cheng","Baolin Peng","Huan Zhang","Tong Zhang","Jianfeng Gao"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"OpenWebRL: Demystifying Online Multi-turn Reinforcement Learning for Visual Web Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02031","noteUrl":"/review?paper=review-neurips-2026-151597","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151598","title":"Bridging the Simulation-to-Experiment Gap with Adversarial Distribution Alignment","authors":["Kai Nelson","Tobias Kreiman","Sergey Levine","Aditi Krishnapriyan"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Bridging the Simulation-to-Experiment Gap with Adversarial Distribution Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.01169v1","noteUrl":"/review?paper=review-neurips-2026-151598","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151602","title":"Imitation from Observations with Trajectory-Level Generative Embeddings","authors":["Yongtao Qu","Shangzhe Li","Weitong Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The positive distance reward contradicts the stated cross-entropy minimization","plainExplanation":"Equation (2) uses positive log distance as a reward to maximize, while Equation (3) identifies the same expectation with a cross-entropy surrogate to minimize. With one expert embedding, \\(r=\\log(1+d)\\) is strictly increasing in distance.","primaryUrl":"https://arxiv.org/html/2601.00452","noteUrl":"/review?paper=review-neurips-2026-151602","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151603","title":"Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-Tuning","authors":["Cristian McGee","El Houcine Bergou","Aritra Dutta"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02190","noteUrl":"/review?paper=review-neurips-2026-151603","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151608","title":"ForesightFlow: Self-Guided Flow Matching for Improving Vision-Language-Action Models","authors":["Yunpeng Mei","Jiakai He","Hongjie Cao","Chenyu Wang","Xiaowen Zhu","Yihan Zhou","Jiamin Wang","Chenbo Xin","Peng Cheng","Yuxuan Yang","Yijie Wang","Xinhu Zheng","Gao Huang","Jie Chen","Gang Wang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"ForesightFlow: Self-Guided Flow Matching for Improving Vision-Language-Action Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04968v1","noteUrl":"/review?paper=review-neurips-2026-151608","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151609","title":"How Hard Is It for Message-Passing GNNs to Simulate One Weisfeiler-Lehman Color-Refinement Step?","authors":["Guanyu Cui","Yuhe Guo","Zhewei Wei","Hsin-Hao Su"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"How Hard Is It for Message-Passing GNNs to Simulate One Weisfeiler-Lehman Color-Refinement Step?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2410.01308","noteUrl":"/review?paper=review-neurips-2026-151609","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151611","title":"Accelerating Rectified Flow Models  via Trajectory-Aware Caching","authors":["Xiao Liu","Kai Liu","Naiyang Guan","Hongliang Lu","Zhixin Wang","Zhikai Chen","Renjing Pei","Yulun Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Accelerating Rectified Flow Models  via Trajectory-Aware Caching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16789","noteUrl":"/review?paper=review-neurips-2026-151611","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151614","title":"The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL.","authors":["Nicolas Beltran Velez","Felix Friedrich","Xiaofeng Zhang","Reyhane Askari Hemmat","Xiaochuang Han","Adriana Romero-Soriano","Michal Drozdzal"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL.","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19162","noteUrl":"/review?paper=review-neurips-2026-151614","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151616","title":"Meow-Omni 1: A Multimodal Large Language Model for Feline Ethology","authors":["Jucheng Hu","Zhangquan Chen","Yulin Chen","Chengjie Hong","Liang Zhou","Tairan Wang","Sifei Li","Giulio Zhu","Feng Zhou","Yiheng Zeng","Suorong Yang","Dongzhan Zhou"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Meow-Omni 1: A Multimodal Large Language Model for Feline Ethology","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09152","noteUrl":"/review?paper=review-neurips-2026-151616","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151617","title":"Symmetry Guarantees Statistic Recovery in Variational Inference","authors":["Daniel Marks","Dario Paccagnan","Mark van der Wilk"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Symmetry Guarantees Statistic Recovery in Variational Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.18310","noteUrl":"/review?paper=review-neurips-2026-151617","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151619","title":"Structured State-Space Regularization for Generation-Friendly Image Tokenization","authors":["Jinsung Lee","Jaemin Oh","Namhun Kim","Dongwon Kim","Byung-Jun Yoon","Suha Kwak"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The Fourier constant mode prevents strict contraction","plainExplanation":"The update is claimed to have induced norm strictly below one although its Fourier basis includes the constant mode. That mode has Laplacian eigenvalue zero.","primaryUrl":"https://arxiv.org/html/2604.11089","noteUrl":"/review?paper=review-neurips-2026-151619","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151621","title":"Selective Disk Bispectrum: A Complete and Rotation Invariant Image Descriptor","authors":["Adele Lantow","Nina Miolane"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The real radial variance constants omit repeated-factor contributions","plainExplanation":"The radial noise-bias formula uses the wrong coefficient index, and the exact radial variance constants disagree with real Gaussian moments. The claim that only zero-phase entries can be biased also excludes allowed negative real coefficients, whose phase is $\\pi$.","primaryUrl":"https://arxiv.org/html/2511.19706","noteUrl":"/review?paper=review-neurips-2026-151621","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151622","title":"FastDSAC: Unlocking the Potential of Maximum Entropy RL in High-Dimensional Humanoid Control","authors":["JUN XUE","Junze Wang","Shanze Wang","Xinming Zhang","Yanjun Chen","Wei Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mean-one standard-deviation weights do not preserve a variance budget","plainExplanation":"The modulation weights satisfy \\(\\sum_iw_i=N\\) and multiply Gaussian standard deviations, but are interpreted as a zero-sum variance exploration budget. For unit base deviations and weights \\((1.5,0.5)\\), total variance changes from two to \\(1.5^2+0.5^2=2.5\\).","primaryUrl":"https://arxiv.org/html/2603.12612","noteUrl":"/review?paper=review-neurips-2026-151622","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151623","title":"Swift Sampling: Selecting Temporal Surprises via Taylor Series","authors":["Dahye Kim","Bhuvan Sachdeva","Karan Uppal","Naman Gupta","Vineeth N Balasubramanian","Deepti Ghadiyaram"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Swift Sampling: Selecting Temporal Surprises via Taylor Series","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22678","noteUrl":"/review?paper=review-neurips-2026-151623","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151624","title":"CoMet: Context and Multiplicity Decomposition for Multimodal Uncertainty Estimation","authors":["Sanghyuk Chun","William Yang","Amaya Dharmasiri","Olga Russakovsky"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The adaptation formula logs an already logged head and reverses the range","plainExplanation":"The adaptation formula logs an already logged head and reverses the range. The written head targets, adaptation score and ranges do not define one common construction.","primaryUrl":"https://arxiv.org/html/2606.32012","noteUrl":"/review?paper=review-neurips-2026-151624","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151625","title":"NEXUS: Neural Energy Fields for Physically Consistent Contact-Rich 3D Object Dynamics","authors":["Qizhen Ying","Guangming Wang","Yangchen Pan","Victor Prisacariu","Brian Sheil","Yixiong Jing"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A velocity-dependent spring scalar is not a conservative mechanical potential","plainExplanation":"The energy model allows velocity-conditioned structural stiffness but defines force only by the negative position gradient of the resulting scalar. With unit relative mass, take \\(U(x,v)=(1+v^2)x^2/2\\), force \\(\\dot v=-U_x\\) and \\(\\dot x=v\\).","primaryUrl":"https://arxiv.org/html/2606.15015","noteUrl":"/review?paper=review-neurips-2026-151625","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151626","title":"Why Struggle with Continuous Latents? Interpretable Discrete Latent Reasoning via Rendered Compression","authors":["Shuochen Chang","Qingyang Liu","Shaobo Wang","Bingjie Gao","Qianli Ma","Haonan Zhao","Yibo Miao","Yulin Sun","Zelin Peng","Jiangtong Li","Li Niu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Why Struggle with Continuous Latents? Interpretable Discrete Latent Reasoning via Rendered Compression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.29712","noteUrl":"/review?paper=review-neurips-2026-151626","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151628","title":"Correlating Cross-Iteration Noise for DP-SGD using Model Curvature","authors":["Xin Gu","Yingtai Xiao","Guanlin He","Jiamu Bai","Daniel Kifer","Kiwan Maeng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The exact average-gradient formula uses the final-step noise filter","plainExplanation":"The average-gradient analysis applies a final-step noise filter to intermediate iterates, invalidating the claimed exact quadratic formula and a subsequent convergence bound. Loewner Hessian domination also does not imply the stated gradient norm comparison. The pseudocode separately inverts the training-step count and applies the clip norm twice to the noise.","primaryUrl":"https://arxiv.org/html/2510.05416","noteUrl":"/review?paper=review-neurips-2026-151628","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151629","title":"Topological Out-of-Domain Generalization in Dynamical Systems Reconstruction","authors":["Georg Trede","Charlotte Doll","Elias Daniel Weber","Daniel Durstewitz"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Topological Out-of-Domain Generalization in Dynamical Systems Reconstruction","plainExplanation":"A single affine feature can match distinct fixed-point and Jacobian growth rates, and a nonlinear scalar feature can exactly represent some sampled flows. The claimed maximal extrapolation certificate also omits same-order material-derivative terms; a separate positional-bias equation has a local subscript typo.","primaryUrl":"https://arxiv.org/html/2606.22969","noteUrl":"/review?paper=review-neurips-2026-151629","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","notation","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-151631","title":"Predicting Plasticity in Deep Continual Learning: A Theoretical Perspective","authors":["Jiuqi Wang","Jayanth Srinivasa","Claire Chen","Shuze D Liu","Ali Payani","Shangtong Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Predicting Plasticity in Deep Continual Learning: A Theoretical Perspective","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09044","noteUrl":"/review?paper=review-neurips-2026-151631","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151634","title":"Online Min-Max Optimization: From Individual Regrets to Cumulative Saddle Points","authors":["Abhijeet Vyas","Brian Bullins"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Static gap control does not imply convergence of the average action","plainExplanation":"The static gap bound does not imply the claimed convergence of the average action. The dynamic analysis also changes the variation definition by moving a maximum inside a sum. The impossibility statement needs online-algorithm quantifiers and corrected saddle values, while the PL parameter is the strong-convexity constant rather than its reciprocal.","primaryUrl":"https://arxiv.org/html/2602.10565","noteUrl":"/review?paper=review-neurips-2026-151634","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151636","title":"Generating in the Limit with Infinitely Many Hallucinations","authors":["Irene Strauss","Alexandra Butoi","Ryan Cotterell"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Generating in the Limit with Infinitely Many Hallucinations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.28354","noteUrl":"/review?paper=review-neurips-2026-151636","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151637","title":"Dynamically Structured Diffusion Language Model Decoding via Bayesian Inference","authors":["Bian Sun","Kevin Zhai","Mubarak Shah","Zhenyi Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The conditioned schedule posterior is uniform over block permutations","plainExplanation":"The conditioned schedule posterior is uniform over block permutations. The posterior-prioritization rationale is incorrect; deterministic sorting and measured decoding results are not refuted.","primaryUrl":"https://arxiv.org/html/2605.09820","noteUrl":"/review?paper=review-neurips-2026-151637","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151643","title":"ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting","authors":["Luchao Wang","Kaimin Liao","Hua Wang","Qian Ren","Zhi Chen","Yaohua Tang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Opacity sums do not determine exact transmittance ordering","plainExplanation":"The opacity-sum threshold in Appendix A.3 does not guarantee the claimed exact transmittance ordering. A valid small-opacity example reverses it; an exact logarithmic threshold supplies the stronger condition the argument needs.","primaryUrl":"https://arxiv.org/html/2605.08739","noteUrl":"/review?paper=review-neurips-2026-151643","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151644","title":"Geometry-Aware Optimal Transport: Fast Intrinsic Dimension and Wasserstein Distance Estimation","authors":["Ferdinand Genans","Olivier Wintenberger"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The dimension upper bound needs a reciprocal factor","plainExplanation":"Inverting this interval gives an upper factor of \\(1/(1-\\gamma)\\), not \\(1+\\gamma\\). The stronger upper factor in the theorem does not follow from its displayed proof event.","primaryUrl":"https://arxiv.org/html/2602.04335","noteUrl":"/review?paper=review-neurips-2026-151644","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151646","title":"Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs","authors":["Zeyi (Andy) Liu","Michael Zhang","Ilana Greenberg","Adam Alnasser","Lucas Baker","John Sous"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06225","noteUrl":"/review?paper=review-neurips-2026-151646","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151648","title":"Leaner transformers can easily learn to cluster","authors":["Charlotte Park","Kenneth Clarkson","Lior Horesh","Takuya Ito","Parikshit Ram"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The tie case and finite Lipschitz constant need repair","plainExplanation":"A tied query receives an averaged cluster code rather than either binary cluster code. The subsequent exact-code selection then fails to include it in either center update.","primaryUrl":"https://arxiv.org/html/2610.09760","noteUrl":"/review?paper=review-neurips-2026-151648","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151649","title":"SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles","authors":["Yuyao Ge","Yiwei Wang","Yuchen He","Baolong Bi","Lingrui Mei","Jiayu Yao","Lizhe Chen","Shenghua Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.09832","noteUrl":"/review?paper=review-neurips-2026-151649","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151659","title":"TokenCLIP: Token-wise Prompt Learning for Zero-shot Anomaly Detection","authors":["Qihang Zhou","Bin-Bin Gao","Guansong Pang","Xin Wang","Jiming Chen","Shibo He"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Raw transport thresholds eliminate the stated dynamic assignments","plainExplanation":"Theorem 3.1 defines arithmetic cluster centroids $\\mu_p,\\mu_q$ and variances $\\sigma_p^2,\\sigma_q^2$. For positive masses transported from both clusters to a subspace, Eq. (10) uses these moments to lower-bound that column's transport cost. A bias–variance decomposition must use the same weights as the cost being decomposed. Appendix J uses transport weights $T_{ij}$ but substitutes arithmetic cluster moments. Uniform row and column marginals do not make a column's weights uniform within each cluster, so the omitted centered cross term need not vanish.","primaryUrl":"https://arxiv.org/html/2510.21171","noteUrl":"/review?paper=review-neurips-2026-151659","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151662","title":"I Have a Stream: Making Self-Supervised Learning Work on Continuous Video","authors":["Ivan Martinović","Lukas Knobel","Yuki Asano"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"I Have a Stream: Making Self-Supervised Learning Work on Continuous Video","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.40333","noteUrl":"/review?paper=review-neurips-2026-151662","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151663","title":"Beyond Selection: Token Parameterization for Extreme Visual Token Compression","authors":["Rui Zhong","YU LI","Zheyu Yan","Cheng Zhuo"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Beyond Selection: Token Parameterization for Extreme Visual Token Compression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35232","noteUrl":"/review?paper=review-neurips-2026-151663","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151667","title":"RayTun3R: Online Camera Adaptation in 3D Foundation Models","authors":["Daniil Sinitsyn","Nikita Araslanov","Daniel Cremers"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The displayed AbsRel formula lacks relative-depth normalization","plainExplanation":"The displayed AbsRel formula lacks relative-depth normalization. The metric definition is mislabeled; the displayed formula does not determine the actual evaluated benchmark values.","primaryUrl":"https://arxiv.org/html/2607.02711","noteUrl":"/review?paper=review-neurips-2026-151667","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151668","title":"Quantized Reasoning Models Think They Need to Think Longer, but They Do Not","authors":["Sanae Lotfi","Polina Kirichenko","Steven Li","Zechun Liu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Quantized Reasoning Models Think They Need to Think Longer, but They Do Not","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.00206","noteUrl":"/review?paper=review-neurips-2026-151668","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151670","title":"Adaptive Power Iteration Method for Differentially Private PCA","authors":["Ta Duy Nguyen","Alina Ene","Huy Nguyen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Gaussian mechanism calibrates variance as though it were standard deviation","plainExplanation":"The printed algorithm squares its retained Gram matrix and uses a threshold floor inconsistent with the proof. The privacy analysis also confuses variance with standard deviation and does not compose its internal budgets into the headline pair. Further proof corrections concern the Gaussian fourth moment, weighted filtering, initialization dimension, truncation bias, two-sided Bernstein constants, zero-singular SVD completions, a dropped sample-gap hypothesis, an inflated intermediate success probability and a missing factor in the recurrence. These local defects have distinct scopes; they do not jointly establish failure of the final asymptotic rate.","primaryUrl":"https://arxiv.org/html/2602.11454","noteUrl":"/review?paper=review-neurips-2026-151670","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151672","title":"RepoLaunch: Automating Build and Management of Code Repositories across Languages and Platforms","authors":["Kenan Li","Rongzhi Li","Linghao Zhang","Qirui Jin","Liao Zhu","XiaoSong Huang","Geng Zhang","Yikai Zhang","Shilin He","Chengxing Xie","Xin Zhang","Zijian Jin","Bowen Li","Chaoyun Zhang","Yu Kang","Yufan Huang","Elsie Nallipogu","Saravan Rajmohan","Qingwei Lin","Dongmei Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RepoLaunch: Automating Build and Management of Code Repositories across Languages and Platforms","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.05026","noteUrl":"/review?paper=review-neurips-2026-151672","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151673","title":"Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis","authors":["Jingxi Feng","Xudong Chen","Yifan Zhang","Heming Xu","Hongcheng Han","Xijing Wang","Dong Zhang","Shaoyi Du"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The NWJ substitution has the wrong sign on the product-law term","plainExplanation":"Conditional mutual information needs the conditional product reference. The conditional-information justification is invalid; a raw negative cross-entropy objective can still be a trivial valid lower bound.","primaryUrl":"https://arxiv.org/html/2609.37220","noteUrl":"/review?paper=review-neurips-2026-151673","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151674","title":"Privacy Amplification Persists under Unlimited Synthetic Data Release","authors":["Clément Pierquin","Aurélien Bellet","Marc Tommasi","Matthieu Boussard"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Tangential mean shifts can have zero Gram divergence","plainExplanation":"The lower amplification bound compares a Gram-release divergence to Gaussian mean-shift divergence and allows equal norms of the two means. Gram laws depend only on squared mean norm.","primaryUrl":"https://arxiv.org/html/2602.04895","noteUrl":"/review?paper=review-neurips-2026-151674","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","mathematical_error","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151675","title":"MIST: Reliable Streaming Decision Trees for Online Class-Incremental Learning via McDiarmid Bound","authors":["Phu-Hoa Pham","Chi Nguyen Tran","Phu-Quy Nguyen-Lam","Dao S Minh","Trung-Kiet Huynh","Long Tran-Thanh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Finite-sample Gini gain is discontinuous at sampled feature values","plainExplanation":"The proof treats finite-sample Gini gain as Lipschitz in a numerical threshold, although empirical routing creates jumps. Its Gaussian quantile conversion also reverses a finite-displacement inequality. The inheritance variance bound separately uses parent pseudo-count mass instead of the routed mass received by each child.","primaryUrl":"https://arxiv.org/html/2605.11617","noteUrl":"/review?paper=review-neurips-2026-151675","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151676","title":"Learning Where to Simulate: Generative Active Sampling for Online PDE Surrogate Training","authors":["Pierre Cesar","Sofya Dymchenko","Abhishek Purandare","Bruno Raffin"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Learning Where to Simulate: Generative Active Sampling for Online PDE Surrogate Training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.09949","noteUrl":"/review?paper=review-neurips-2026-151676","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151680","title":"Listening to the Wise Few: Query–Key Alignment Unlocks Latent Correct Answers in Large Language Models","authors":["Eduard Tulchinskii","Kristian Kuznetsov","Laida Kushnareva","Anastasia Voznyuk","Andrei Andriiainen","Irina Piontkovskaya","Evgeny Burnaev","Serguei Barannikov"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Listening to the Wise Few: Query–Key Alignment Unlocks Latent Correct Answers in Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2410.02343","noteUrl":"/review?paper=review-neurips-2026-151680","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151683","title":"Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models","authors":["Guangzhi Xiong","Qiao Jin","Sanchit Sinha","Zhiyong Lu","Aidong Zhang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20158","noteUrl":"/review?paper=review-neurips-2026-151683","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151684","title":"Latent Action Reparameterization for Efficient Agent Inference","authors":["Wenhao Huang","Qingwen Zeng","Qiyue Chen","Zijie Guo","Yu Sun","Cheng Yang","Siru Ouyang","Jiri Gesi","Fang Wu","Jiayi Zhang","Huaming Chen","Bang Liu","Robert Tang","Chenglin Wu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Latent Action Reparameterization for Efficient Agent Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18597","noteUrl":"/review?paper=review-neurips-2026-151684","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151685","title":"Training Generalizable Collaborative Agents via Strategic Risk Aversion","authors":["Chengrui Qu","Yizhou Zhang","Nicolas Lanzetti","Eric Mazumdar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The opponent Hessian contains cross-player blocks","plainExplanation":"The Gaussian multi-opponent calculation omits cross-player Hessian blocks, so pairwise precision positivity need not ensure joint integrability. The symmetric robot mean also misses a one-half factor. In the finite-game argument, the distance estimate drops the strong-concavity parameter and consequently understates the derived threshold.","primaryUrl":"https://arxiv.org/html/2602.21515","noteUrl":"/review?paper=review-neurips-2026-151685","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151688","title":"MURPHY: Feedback-Aware GRPO with Retrospective Credit Assignment for Multi-Turn Code Generation","authors":["Chanakya Ekbote","Vijay Lingam","Sujay Sanghavi","Luke Huan","Behrooz O Tehrani","Anoop Deoras","Stefano Soatto"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MURPHY: Feedback-Aware GRPO with Retrospective Credit Assignment for Multi-Turn Code Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.07833","noteUrl":"/review?paper=review-neurips-2026-151688","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151689","title":"Generalized Bayes for Causal Inference","authors":["Emil Javurek","Dennis Frauen","Yuxin Wang","Stefan Feuerriegel"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Generalized Bayes for Causal Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.03035","noteUrl":"/review?paper=review-neurips-2026-151689","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151691","title":"RATS! Patches Talk Through Registers: Emergent Parts in Register Attention Transformers","authors":["Timing Yang","Predrag Neskovic","Jansen Seheult","Wenchao Han","Anand Bhattad","Alan Yuille","Feng Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"RATS! Patches Talk Through Registers: Emergent Parts in Register Attention Transformers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.14701","noteUrl":"/review?paper=review-neurips-2026-151691","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151694","title":"A Hierarchical World Model for Driving","authors":["Sudhanshu Mittal","Arian Mousakhan","Silvio Galesso","Karim Farid","Johannes Dienert","Rajat Sahay","Thomas Brox"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"A Hierarchical World Model for Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.15898","noteUrl":"/review?paper=review-neurips-2026-151694","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151695","title":"Rethinking Language Model Scaling under Transferable Hypersphere Optimization","authors":["Liliang Ren","Yang Liu","yelong shen","Weizhu Chen"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Bounded local Jacobians do not bound depthwise transport products","plainExplanation":"The depth theorem assumes local Jacobians of order one and small steps, then treats downstream residual-Jacobian products as constant-order factors. A product of factors \\(I+\\alpha J\\) can grow like \\((1+\\alpha)^L\\) despite \\(\\|J\\|=O(1)\\).","primaryUrl":"https://arxiv.org/html/2603.28743","noteUrl":"/review?paper=review-neurips-2026-151695","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151697","title":"A Subgoal-driven RL Framework for Improving Long-Horizon Web Agents","authors":["Taiyi Wang","Sian Gooding","Florian Hartmann","Oriana Riva","Edward Grefenstette"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Undiscounted potential differences can change the optimal policy","plainExplanation":"The actual shaping rule uses undiscounted potential differences inside a discounted return. Its path-dependent terms can reverse the original action ranking, and an outcome-trained primary critic does not remove them. The invariance proof in Appendix A.2 uses the different, correctly discounted shaping rule.","primaryUrl":"https://arxiv.org/html/2603.19685v1","noteUrl":"/review?paper=review-neurips-2026-151697","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151698","title":"Toward a Unified Statistical Theory of Unsupervised Pretraining and Supervised Neural Knowledge Graph Learning","authors":["Suqi Liu","Miklos Z. Racz","Jifan Zhang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The substituted complexity term can become negative","plainExplanation":"The end-to-end risk bound substitutes an uncapped architecture bound into $H\\log(en/H)/n$. Actual pseudodimension can satisfy $p\\le n$ on a small finite domain while the architecture bound exceeds $en$, making the advertised complexity contribution—and eventually the risk upper bound—negative.","primaryUrl":"https://arxiv.org/html/2607.26346","noteUrl":"/review?paper=review-neurips-2026-151698","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151699","title":"Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting","authors":["Daniel Holmberg","Joel Oskarsson","Erik Wikingsson","Fredrik Lindsten","Teemu Roos"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15470","noteUrl":"/review?paper=review-neurips-2026-151699","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151700","title":"TiRex-2: Generalizing TiRex to Multivariate Data and Streaming","authors":["Patrick Podest","Marco Pichler","Elias Bürger","Levente Zólyomi","Bernhard Voggenberger","Wilhelm F Berghammer","Daniel Klotz","Sebastian Böck","Günter Klambauer","Sepp Hochreiter"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TiRex-2: Generalizing TiRex to Multivariate Data and Streaming","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.01204","noteUrl":"/review?paper=review-neurips-2026-151700","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151706","title":"Network of Theseus (Like the ship)","authors":["Vighnesh Subramaniam","Colin Conwell","Boris Katz","Andrei Barbu","Brian Cheung"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The softmax temperature limit is reversed","plainExplanation":"The appendix’s repeated-trace expression is not an unbiased HSIC estimator. The neighborhood-overlap limit uses the wrong temperature direction, and separately exact-masked distributions can make the KL alignment loss infinite.","primaryUrl":"https://arxiv.org/html/2512.04198","noteUrl":"/review?paper=review-neurips-2026-151706","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151707","title":"Mean-Field Parallel Decoding for Discrete Diffusion Language Models","authors":["Tamim Zoabi","Ameen A Ali","Liran Ringel","Lior Wolf"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Mean-Field Parallel Decoding for Discrete Diffusion Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.15805v1","noteUrl":"/review?paper=review-neurips-2026-151707","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151709","title":"Learn from your own latents and not from tokens: A sample-complexity theory","authors":["Daniel Korchinski","Alessandro Favero","Matthieu Wyart"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Learn from your own latents and not from tokens: A sample-complexity theory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27734","noteUrl":"/review?paper=review-neurips-2026-151709","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151711","title":"Explanations over Graphs: An Agent Architecture for IT Enterprise Diagnostic Tasks","authors":["Saurabh Jha","Rohan R. Arora","Bhavya Bhavya","Noah Zheutlin","Paulina Toro Isaza","Laura Shwartz","Yu Deng","Daby Sow","Ruchir Puri"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Explanations over Graphs: An Agent Architecture for IT Enterprise Diagnostic Tasks","plainExplanation":"No issue was found in the checked claims. The saved counterexample assumes that an origin explains a symptom whenever a path connects them, so any one reachable origin is a sufficient explanation. The source does not specify that set-cover interpretation of causality.","primaryUrl":"https://arxiv.org/html/2601.17915v2","noteUrl":"/review?paper=review-neurips-2026-151711","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151714","title":"HaM-World: Soft-Hamiltonian World Models with Selective Memory for Planning","authors":["Haoyun Tang","Haodong Cui","Keyao Xu","Zhan-Dong Mei","Kun Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The regret subtraction has the opposite sign from its interpretation","plainExplanation":"The selection-regret definition subtracts best return from selected return, reversing the positive-cost convention used in the table. The GRU comparison also confuses fixed trained parameters with gates that still depend on input and hidden state.","primaryUrl":"https://arxiv.org/html/2605.05951","noteUrl":"/review?paper=review-neurips-2026-151714","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151715","title":"In-Expectation Convergence of Stochastic Gradient Methods under Heavy-Tailed Noise","authors":["Zijian Liu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"In-Expectation Convergence of Stochastic Gradient Methods under Heavy-Tailed Noise","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.00520","noteUrl":"/review?paper=review-neurips-2026-151715","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151717","title":"Two-Fidelity Best-Action Identification for Stochastic Minimax Tree","authors":["Peter Chen","Xi Chen"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Two-Fidelity Best-Action Identification for Stochastic Minimax Tree","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.01708","noteUrl":"/review?paper=review-neurips-2026-151717","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151719","title":"Early Failure Detection and Intervention in Video Diffusion Models","authors":["Byungki Kwon","Sohwi Lim","Nam Hyeon-Woo","Moon Ye-Bin","Tae-Hyun Oh"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Early Failure Detection and Intervention in Video Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.14320","noteUrl":"/review?paper=review-neurips-2026-151719","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151720","title":"EMO: Pretraining Mixture of Experts for Emergent Modularity","authors":["Ryan Wang","Akshita Bhagia","Sewon Min"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EMO: Pretraining Mixture of Experts for Emergent Modularity","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06663","noteUrl":"/review?paper=review-neurips-2026-151720","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151721","title":"Post-ADC Inference: Valid Inference After Active Data Collection","authors":["Shuichi Nishino","Tomohiro Shiraishi","Teruyuki Katsuoka","Ichiro Takeuchi"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Post-ADC Inference: Valid Inference After Active Data Collection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11511","noteUrl":"/review?paper=review-neurips-2026-151721","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151725","title":"The Stability of Online Algorithms in Performative Prediction","authors":["Gabriele Farina","Juan C Perdomo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Stability of Online Algorithms in Performative Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.24207","noteUrl":"/review?paper=review-neurips-2026-151725","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151727","title":"Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement","authors":["Abdul Moeed","Stefan Schrod","Martin Rohbeck","Marc J Bonder","Pavlo Lutsik","Oliver Stegle","Daniel Dimitrov"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"A log-space shift is not pure multiplicative count scaling","plainExplanation":"An additive log1p shift includes an offset after inversion, so it is not exactly equivalent to the stated multiplicative count perturbation. Either intervention remains independently definable.","primaryUrl":"https://arxiv.org/html/2606.08493","noteUrl":"/review?paper=review-neurips-2026-151727","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151728","title":"Geometric Signatures of Reasoning: A Spectral Perspective on Task Hardness","authors":["Aria Masoomi","Mahsa Bazzaz","Adel Javanmard","Vahab Mirrokni"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Maximal effective dimension does not uniquely identify a flat spectrum","plainExplanation":"Flat spectra maximize the threshold-based effective dimension, but need not be its unique maximizers. Equality at the final inspected partial sum does not force a flat spectrum.","primaryUrl":"https://arxiv.org/html/2607.01571","noteUrl":"/review?paper=review-neurips-2026-151728","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151729","title":"Offline Constrained Reinforcement Learning under Partial Data Coverage","authors":["Seokmin Ko","Ambuj Tewari","Kihyuk Hong"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Scaled-simplex interpolation must use scaled vertices","plainExplanation":"The proof extends affine concentration from simplex vertices to \\(\\lambda=B\\sum_i\\alpha_i e_i\\), but its interpolation evaluates unscaled \\(e_i\\). For \\(F(\\lambda)=c+v^\\top\\lambda\\), convex interpolation at that point uses \\(Be_i\\), together with zero if included.","primaryUrl":"https://arxiv.org/html/2505.17506","noteUrl":"/review?paper=review-neurips-2026-151729","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151730","title":"Unlocking the Duality between Flow and Field Matching","authors":["Daniil Shlenskii","Alexander Varlamov","Nazar Buzun","Aleksandr Korotin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Interior uniqueness does not determine an endpoint-conditioned curve","plainExplanation":"Appendix B.1 assumes local spatial Lipschitz regularity only on compact subsets of \\(0<t<T\\), then infers a unique trajectory from singular endpoint conditions. Distinct admissible curves can share both endpoints.","primaryUrl":"https://arxiv.org/html/2602.02261","noteUrl":"/review?paper=review-neurips-2026-151730","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151733","title":"Identifying Structural Biases from Causal Mechanism Shifts","authors":["Praharsh Nanavati","Jilles Vreeken","David Kaltenpoth"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Identifying Structural Biases from Causal Mechanism Shifts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.18834","noteUrl":"/review?paper=review-neurips-2026-151733","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151734","title":"AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation","authors":["Shanghua Gao","Ada Fang","Marinka Zitnik"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.28655","noteUrl":"/review?paper=review-neurips-2026-151734","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151737","title":"Frontier Task Synthesis Via Solution-Centric Evolution","authors":["Yangzhen Wu","Aaron Li","Wenjie Ma","Li Cao","Ziheng Zhou","Mert Cemri","Shu Liu","Yuran Xiu","Chenxiao Yan","Haikun Zhao","Bin Yu","Ion Stoica","Dawn Song"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Frontier Task Synthesis Via Solution-Centric Evolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.01286","noteUrl":"/review?paper=review-neurips-2026-151737","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151739","title":"ComPose: When to Trust Hands for Object Pose Tracking","authors":["Jisu Shin","Junoh Lee","JunGyu Lee","Inhwan Bae","Dohyeon Lee","Hokyun Im","Youngwoon Lee","Hae-Gon Jeon"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Shortest-path SLERP requires endpoint sign alignment","plainExplanation":"The quaternion blend uses angle \\(\\arccos|q_0\\cdot q_1|\\) but combines the original endpoint signs. Absolute-dot angles alone do not choose consistent quaternion representatives.","primaryUrl":"https://arxiv.org/html/2605.23523","noteUrl":"/review?paper=review-neurips-2026-151739","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151741","title":"Beyond Empirical Support: Structured Outlier Generation via Sinkhorn Optimal Transport","authors":["Haixiang Sun","Andrew Liu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Beyond Empirical Support: Structured Outlier Generation via Sinkhorn Optimal Transport","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31470","noteUrl":"/review?paper=review-neurips-2026-151741","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151742","title":"RelAgent: LLM Agents as Data Scientists for Relational Learning","authors":["Xingyue Huang","Louis Tichelman","Jinwoo Kim","Krzysztof Olejniczak","Ismail Ilkan Ceylan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RelAgent: LLM Agents as Data Scientists for Relational Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07840","noteUrl":"/review?paper=review-neurips-2026-151742","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151743","title":"Contrastive Representation Shaping for LLM Unlearning","authors":["Haoran Tang","Rajiv Khanna"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The raw-dot-product result does not imply decreasing cosine","plainExplanation":"The proved raw-dot-product decrease does not imply a cosine decrease after a Euclidean update. A sampled-negative improvement also need not reduce overlap with an independently averaged retain population.","primaryUrl":"https://arxiv.org/html/2601.22028","noteUrl":"/review?paper=review-neurips-2026-151743","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-151744","title":"Extrapolative Weight Averaging Reveals Correctness–Efficiency Frontiers in Code RL","authors":["Kunhao Zheng","Juliette Decugis","Pierre Chambon","Jonas Gehring","Taco Cohen","Benjamin Negrevergne","Gabriel Synnaeve"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strict and inclusive test tiers do not have the same average reward","plainExplanation":"Equation (2) requires success at input lengths strictly below powers of ten, while the appendix tier reward includes the boundary and uses the largest passing tier. A boundary failure distinguishes the rewards.","primaryUrl":"https://arxiv.org/html/2605.28751","noteUrl":"/review?paper=review-neurips-2026-151744","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151745","title":"Glob3R: Global Structure-from-Motion with 3D Foundation Models","authors":["Junyuan Deng","Heng Li","Kejie Qiu","Lingteng Qiu","Rui Peng","Weichao Shen","Weihao Yuan","Siyu Zhu","Zilong Dong","Ping Tan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The warp description switches coordinate direction","plainExplanation":"Equation (5) minimizes residuals \\(X_j-c_i-d_{ij}R_i^\\top v_{ij}\\) with free centers, points and depths, without a noncollapse scale constraint. All centers and points equal with all depths zero make every residual vanish.","primaryUrl":"https://arxiv.org/html/2607.09225","noteUrl":"/review?paper=review-neurips-2026-151745","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151748","title":"Raw-Routed Mixture of Adapters: A Causal Intervention for Routing Collapse in Time Series Foundation Models","authors":["Hung Phan","Thuy T Nguyen","Minh N Dinh","Nhat-Quang Tran"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Joint backbone motion can reverse the router probability derivative","plainExplanation":"Joint backbone motion can reverse the claimed router update direction, and raw routing removes a direct derivative path without eliminating all loss-mediated feedback. Calibrated classifier cross-entropy also retains an excess-risk term and need not equal conditional entropy.","primaryUrl":"https://arxiv.org/html/2609.39445","noteUrl":"/review?paper=review-neurips-2026-151748","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151751","title":"The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning","authors":["Mojtaba Sahraee-Ardakan","Mauricio Delbracio","Peyman Milanfar"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning","plainExplanation":"The general schedule assumptions do not guarantee posterior time concentration. Even when concentration holds for standard schedules, an unbounded-score covariance can remain leading order, noise drift error can stay finite rather than diverge, and a posterior signal mixture can lose the claimed exponential accuracy.","primaryUrl":"https://arxiv.org/html/2602.18428","noteUrl":"/review?paper=review-neurips-2026-151751","collection":"neurips-2026","severity":"critical","issueTypes":["lemma_error"],"classifications":["critical"]},{"slug":"review-neurips-2026-151752","title":"SurgVista: Long-Horizon Surgical World Modeling with Plausible Instrument-Tissue Dynamics","authors":["Wentao Pan","WUYANG LI","Shengyuan Liu","Xinyu Liu","Hengyu Liu","Yixuan Yuan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SurgVista: Long-Horizon Surgical World Modeling with Plausible Instrument-Tissue Dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19889","noteUrl":"/review?paper=review-neurips-2026-151752","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151754","title":"LensVLM: Selective Context Expansion for Compressed Visual Representation of Text","authors":["Roy Xie","Dan Friedman","Donghan Yu","Bowen Pan","Christopher Fifty","Jang-Hyun Kim","Xianzhi Du","Zhe Gan","Vivek Rathod","Bhuwan Dhingra"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"LensVLM: Selective Context Expansion for Compressed Visual Representation of Text","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07019","noteUrl":"/review?paper=review-neurips-2026-151754","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151757","title":"DocPTBench: Benchmarking End-to-End Photographed Document Parsing and Translation","authors":["Yongkun Du","Pinxuan Chen","Xuye Ying","Zhineng Chen"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"DocPTBench: Benchmarking End-to-End Photographed Document Parsing and Translation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2511.18434","noteUrl":"/review?paper=review-neurips-2026-151757","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151760","title":"Native Multi-Dimensional Subquadratic Operators via Input Dependent Long Convolutions","authors":["David Wessels","FARHAD RAMEZANGHORBANI","David W Romero","Alireza Moradzadeh","Olivia Viessmann","Maksim Zhdanov","John St. John","Ken Janik","David Knigge","Yucheng Tang","Erik Bekkers","Saee Paliwal"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Native Multi-Dimensional Subquadratic Operators via Input Dependent Long Convolutions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.19378","noteUrl":"/review?paper=review-neurips-2026-151760","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151764","title":"SIGMA: Semantic-Difference Instruction-Grounding Mask Annotator for Text-Driven Image Manipulation Localization","authors":["Peiyu Zhuang","Jianquan Yang","Haodong Li","Zhuoying Cai","Ruitao Xie","Jishen Zeng","Baoying Chen","Jiwu Huang","XIAOCHUN CAO"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"SIGMA: Semantic-Difference Instruction-Grounding Mask Annotator for Text-Driven Image Manipulation Localization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27924","noteUrl":"/review?paper=review-neurips-2026-151764","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151767","title":"One View Is Enough: In-the-Wild Monocular Pretraining for Novel View Generation","authors":["Adrien RAMANANA RAHARY","Nicolas Dufour","Patrick Perez","David Picard"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"One View Is Enough: In-the-Wild Monocular Pretraining for Novel View Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.23488","noteUrl":"/review?paper=review-neurips-2026-151767","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151768","title":"Randomized Kriging Believer For Parallel Bayesian Optimization With Regret Bounds","authors":["Shuhei Sugiura","Ichiro Takeuchi","Shion Takeno"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Randomized Kriging Believer For Parallel Bayesian Optimization With Regret Bounds","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.01470","noteUrl":"/review?paper=review-neurips-2026-151768","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151771","title":"Learning to Cut: Reinforcement Learning for Benders Decomposition","authors":["Haochen Cai","Xian Yu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The cut-free initial master needs a lower bound on its recourse variables","plainExplanation":"At the first iteration there are no such cuts, so a free \\(\\theta^\\omega\\) can tend to \\(-\\infty\\). Finiteness of the true recourse function does not bound an unconstrained surrogate variable.","primaryUrl":"https://arxiv.org/html/2605.06516","noteUrl":"/review?paper=review-neurips-2026-151771","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","definition_error","formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151774","title":"Pseudo-Labeling for Unsupervised Domain Adaptation with Kernel GLMs","authors":["Nathan Weill","Kaizheng Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The oracle inequality lacks curvature control on the candidate domain","plainExplanation":"The oracle inequality allows candidates outside its curvature-controlled domain and admits a logistic counterexample. The quadratic-form lemma also lacks dependence assumptions needed to turn coordinatewise tail control into its trace-only bound.","primaryUrl":"https://arxiv.org/html/2603.19422","noteUrl":"/review?paper=review-neurips-2026-151774","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151776","title":"Alignment Imprint: Zero-Shot AI-Generated Text Detection via Provable Preference Discrepancy","authors":["Junxi Wu","Kailin Huang","Dongjian Hu","Bin Chen","Hao Wu","Shu-Tao Xia","Changliang Zou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The proof replaces perturbation moments by text-law moments","plainExplanation":"The operational alignment statistic standardizes using perturbation law \\(P_\\theta\\), whereas Equation (23) uses AI-text law \\(q\\) without equating them. These standardizations need not agree.","primaryUrl":"https://arxiv.org/html/2604.16923","noteUrl":"/review?paper=review-neurips-2026-151776","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151777","title":"Deep Barycentric Regression for Optimal Transport Map Estimation and its Statistical Optimality","authors":["Kunwoong Kim","Insung Kong","Yongdai Kim"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Deep Barycentric Regression for Optimal Transport Map Estimation and its Statistical Optimality","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.06598","noteUrl":"/review?paper=review-neurips-2026-151777","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151780","title":"Spectral-Aligned Pruning for Universal Error-Correcting Code Transformers","authors":["Sanghyeon Cho","Taewoo Park","Seong-Joon Park","Dae-Young Yun","Hee-Youl Kwak","Sang-Hyo Kim","Yongjune Kim"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Independent frames do not make decoded bit errors independent","plainExplanation":"The BER uncertainty justification treats all decoded bits in independent AWGN frames as binomial trials for a Wilson interval. In a length-49 repetition code, maximum-likelihood decoding returns all zeros or all ones, so every bit error in a frame is the same Bernoulli event.","primaryUrl":"https://arxiv.org/html/2602.01602","noteUrl":"/review?paper=review-neurips-2026-151780","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151781","title":"Boule or Baguette? A Study on Task Topology, Length Generalization, and the Benefit of Reasoning Trace","authors":["William Tong","Ege Çakar","Cengiz Pehlevan"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Fixed coefficient magnitudes do not ensure Gaussian cancellation","plainExplanation":"Proposition B.2 controls Gaussian feature cross terms assuming equal coefficient magnitudes, then applies the result to learned coordinates. Let \\(x_1,\\ldots,x_V\\) be independent standard Gaussian vectors in \\(\\mathbb R^H\\). Data-dependent signs can remove cancellation.","primaryUrl":"https://arxiv.org/html/2602.14404","noteUrl":"/review?paper=review-neurips-2026-151781","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151782","title":"SymDrift: One-Shot Generative Modeling under Symmetries","authors":["Samir Darouich","Vinh Tong","Lluis Pastor Perez","Tanja Bien","Loay Mualem","Mathias Niepert"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Equivariant argmin sets do not ensure equivariant deterministic selections","plainExplanation":"Appendix B.6 claims universal strict equivariance for a drift using a selected optimal alignment, without a uniqueness or almost-sure qualifier in that lemma. Distance invariance gives \\(\\operatorname{Argmin}(hx,y)=h\\operatorname{Argmin}(x,y)\\) as sets.","primaryUrl":"https://arxiv.org/html/2605.06140","noteUrl":"/review?paper=review-neurips-2026-151782","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151783","title":"Reason to Play: Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners","authors":["Botos Csaba","Sreejan Kumar","Austin T D Andrews","Laurence T Hunt","Josh Tenenbaum","Christopher Summerfield","Rui Costa","Marcelo G Mattar","Momchil Tomov"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Isolated ridge prediction correlation does not measure unique information","plainExplanation":"The main-band ridge prediction is correlated with the target and interpreted as information unique beyond nuisance feature bands. Duplicate main and nuisance feature \\(x\\), with target \\(x\\), receive positive ridge coefficients in both bands.","primaryUrl":"https://arxiv.org/html/2605.08019","noteUrl":"/review?paper=review-neurips-2026-151783","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151784","title":"A Closer Look at Dynamic Scene Graph Generation in the Era of Multimodal Large Language Models","authors":["Xuanming Cui","Jaiminkumar Ashokbhai Bhoi","Chionh W Peng","Adriel Kuek","Ser Nam Lim"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"A Closer Look at Dynamic Scene Graph Generation in the Era of Multimodal Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2503.15846","noteUrl":"/review?paper=review-neurips-2026-151784","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151789","title":"Adaptive Multi-view Graph Contrastive Learning via Fractional Continuous Dynamics","authors":["Yanan Zhao","Feng Ji","Jingyang Dai","Jiaze Ma","Keyue Jiang","Kai Zhao","Wee Peng Tay"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Lipschitz continuity does not imply decaying perturbations with the horizon","plainExplanation":"A stationary graph mode preserves perturbations, contradicting decay with increasing evolution horizon. A valid fractional Grönwall estimate instead gives finite-horizon continuous dependence. In the spectral theorem, the second inverse-power coefficient can be negative; the positive leading correction and its slower decay for smaller fractional order remain valid under the stated restricted comparison. These issues do not establish a failure of the empirical learning method.","primaryUrl":"https://arxiv.org/html/2511.06216","noteUrl":"/review?paper=review-neurips-2026-151789","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151790","title":"Stabilizing Knowledge, Promoting Reasoning: Dual-Token Constraints for RLVR","authors":["Jiakang Wang","Runze Liu","Fuzheng Zhang","Xiu Li","Guorui Zhou","Kun Gai","Ling Pan"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Stabilizing Knowledge, Promoting Reasoning: Dual-Token Constraints for RLVR","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2507.15778","noteUrl":"/review?paper=review-neurips-2026-151790","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151793","title":"EventPrune: Cascaded Event-Assisted Token Pruning for Efficient First-Person Dynamic Spatial Reasoning","authors":["Pengtao Ma","Ziliang Zhou","Ciyu Ruan","Haoyang Wang","kaiyuan Li","Zihang GONG","Wenhua Ding","Chen Gao","Jingao Xu","Xinlei Chen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"EventPrune: Cascaded Event-Assisted Token Pruning for Efficient First-Person Dynamic Spatial Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.19506","noteUrl":"/review?paper=review-neurips-2026-151793","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151794","title":"Reinforcement Learning for Exponential Utility: Algorithms and Convergence in Discounted MDPs","authors":["Gugan Chandrashekhar Mallika Thoppe","Prashanth L.A.","Ankur Naskar","Sanjay P. Bhat"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Reinforcement Learning for Exponential Utility: Algorithms and Convergence in Discounted MDPs","plainExplanation":"Three mathematical issues concern the objective represented by the fixed point, coverage of all actions, and the signed-reward noise estimate.","primaryUrl":"https://arxiv.org/html/2605.08053","noteUrl":"/review?paper=review-neurips-2026-151794","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","incorrect_bound","missing_assumption"],"classifications":["critical"]},{"slug":"review-neurips-2026-151797","title":"ROCKET: Residual-Oriented Multi-Layer Alignment for Spatially-Aware Vision-Language-Action Models","authors":["Guoheng Sun","Tingting Du","Kaixi Feng","Chenxiang Luo","Xingguo Ding","Zheyu Shen","Ziyao Wang","Yexiao He","Ang Li"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"ROCKET: Residual-Oriented Multi-Layer Alignment for Spatially-Aware Vision-Language-Action Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.17951","noteUrl":"/review?paper=review-neurips-2026-151797","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151800","title":"Decoupling Conversational Dynamics in Full-Duplex Spoken Models through Reinforcement Learning","authors":["Yuxin Li","Donghang Wu","Guan-Ting Lin","Hung-yi Lee","Chengwei Qin","CHEN CHEN","Zhehuai Chen"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Decoupling Conversational Dynamics in Full-Duplex Spoken Models through Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.07148","noteUrl":"/review?paper=review-neurips-2026-151800","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151802","title":"Confidence-Based Decoding is Provably Efficient for Diffusion Language Models","authors":["Changxiao Cai","Gen Li"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Confidence-Based Decoding is Provably Efficient for Diffusion Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.22248","noteUrl":"/review?paper=review-neurips-2026-151802","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151803","title":"A Biconvex Formulation for Stable Transport of Mixture Models with a Unique Solution","authors":["Yeganeh Mohammadali Marghi","Kelly Jin","Uygar Sümbül"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A Biconvex Formulation for Stable Transport of Mixture Models with a Unique Solution","plainExplanation":"The single-pass optimization proof replaces aggregate marginals by stronger pairwise constraints. Pointwise Gaussian regularity also does not give its uniform covariance constant, and the perturbation proof needs a component-mismatch term when comparing different kernels.","primaryUrl":"https://arxiv.org/html/2606.02515","noteUrl":"/review?paper=review-neurips-2026-151803","collection":"neurips-2026","severity":"critical","issueTypes":["proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-151805","title":"Artificial Aphasias in Lesioned Language Models","authors":["Nathan Roll","Jill Kries","Laura Gwilliams","Cory Shain"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Artificial Aphasias in Lesioned Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16222","noteUrl":"/review?paper=review-neurips-2026-151805","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151807","title":"On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models","authors":["William Tong","Aryo Lotfi","Emmanuel Abbe","Konstantinos Vaggelakos","Vishnu Banna","Etai Littwin","Joshua Susskind","Eran Malach"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.16540","noteUrl":"/review?paper=review-neurips-2026-151807","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151809","title":"StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training","authors":["Bao Tang","Jiahao Guo","Haoxiang Cao","Wenyu Liu","Changqian Yu","Kun Gai","Xinggang Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Exact code matches make the proposed weight undefined","plainExplanation":"Equation (4) divides the minimum squared code distance by each assigned token's squared distance and claims weights in \\((0,1]\\). An exact match makes the minimum zero.","primaryUrl":"https://arxiv.org/html/2609.26774","noteUrl":"/review?paper=review-neurips-2026-151809","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151810","title":"Large Language Models Develop Belief State Geometry In-Context","authors":["Daniel Balcells","Andrew Jun Lee","Chirag Rastogi","Paul Riechers","Adam Shai","Xavier Poncini"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Large Language Models Develop Belief State Geometry In-Context","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2609.17376","noteUrl":"/review?paper=review-neurips-2026-151810","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151811","title":"Adaptive Inference for Functionals of M-Estimands","authors":["James Leiner","Aurelien Bibaut","Nathan Kallus","Aaditya Ramdas","Koulik Khamaru"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Growing inverse propensity weights invalidate the uniform moment bound","plainExplanation":"The functional expansion omits a nonlinear remainder independent of the Riesz representer, and the centering estimator omits its averaging factor. The weighted higher-moment lemma also loses the dependence on growing inverse propensities, which several downstream arguments already retain.","primaryUrl":"https://arxiv.org/html/2609.39274","noteUrl":"/review?paper=review-neurips-2026-151811","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151812","title":"Fast and Accurate Probing of In-Training LLMs' Downstream Performances","authors":["Zhichen Liu","Tianle Lun","Zhibin Wen","Hao An","Yulin Ou","Jianhui Xu","Hao Zhang","Wenyi Fang","YANG ZHENG","Yang Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Fast and Accurate Probing of In-Training LLMs' Downstream Performances","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.01025","noteUrl":"/review?paper=review-neurips-2026-151812","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151813","title":"Breaking the Grid: Distance-Guided Reinforcement Learning in Large Discrete Action Spaces","authors":["Heiko Hoppe","Fabian Akkerman","Wouter van Heeswijk","Maximilian Schiffer"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The finite rounded DBU target need not be an ascent direction","plainExplanation":"A finite rounded DBU target can point downhill even for a linear value function and arbitrarily small noise. The proof instead analyzes an unrounded population target, so it does not establish deterministic improvement for the stated update.","primaryUrl":"https://arxiv.org/html/2602.08616","noteUrl":"/review?paper=review-neurips-2026-151813","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151815","title":"Predicting and improving test-time scaling laws via reward tail-guided search","authors":["Muheng Li","Jian Qian","Wenlong Mou"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Intermediate-state generation consumes additional sampling budget","plainExplanation":"The problem charges a sample from any state, including the prompt, one unit. Algorithm 2 generates \\(K\\) states, takes \\(Km\\) pilots, then \\(N-Km\\) further responses while calling \\(N\\) the total budget. The total charged work is \\(K+Km+N-Km=N+K\\).","primaryUrl":"https://arxiv.org/html/2602.01485","noteUrl":"/review?paper=review-neurips-2026-151815","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151817","title":"Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion","authors":["ShiYing Huang","Liang Lin","Yuer Li","Kaiwen Luo","Zhenhong Zhou","An Zhang","Junhao Dong","Kun Wang","Zhigang Zeng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11679","noteUrl":"/review?paper=review-neurips-2026-151817","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151819","title":"A Differentiable Interior-Point Method in Single Precision","authors":["Jon Arrizabalaga","Kevin Tracy","Zac Manchester"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Complementarity averages over inequality coordinates","plainExplanation":"The setup has \\(m\\) equalities and \\(p\\) inequalities, with central-path products \\(s_i z_i=\\kappa\\). Algorithms 1 and 2 recover \\(\\kappa\\) by dividing \\(s^\\top z\\) by \\(m\\). The sum is \\(p\\kappa\\), so the divisor must be \\(p\\).","primaryUrl":"https://arxiv.org/html/2605.17913","noteUrl":"/review?paper=review-neurips-2026-151819","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151820","title":"Transferable SCF-Acceleration through Solver-Aligned Initialization Learning","authors":["Eike S. Eberhard","Viktor Kotsev","Timm Güthle","Stephan Günnemann"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Transferable SCF-Acceleration through Solver-Aligned Initialization Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.21657","noteUrl":"/review?paper=review-neurips-2026-151820","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151821","title":"DARTS: Targeting Prognostic Covariates in Budget-Constrained Sequential Experiments","authors":["Kateryna Husar","Alexander Volfovsky"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DARTS: Targeting Prognostic Covariates in Budget-Constrained Sequential Experiments","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06608","noteUrl":"/review?paper=review-neurips-2026-151821","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151824","title":"Certified but Private: Scalable Zero-Knowledge Proofs for the Formal Verification of Neural Networks","authors":["Youwei Zhong","Ben Merbaum","Timos Antonopoulos","Ning Luo","CHARALAMPOS PAPAMANTHOU","Katerina Sotiraki","Ruzica Piskac"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Local convexity does not guarantee a representable tangency witness","plainExplanation":"The local convexity assumption does not guarantee a representable finite-difference tangency witness on every permitted grid. An admissible smooth-sided activation and grid violate Theorem F.1’s universal existence bullet.","primaryUrl":"https://arxiv.org/html/2608.17070v2","noteUrl":"/review?paper=review-neurips-2026-151824","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151825","title":"Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction","authors":["Kaito Shiku","Kazuya Nishimura","Yasuhiro Kojima","Ryoma Bise"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33928","noteUrl":"/review?paper=review-neurips-2026-151825","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151826","title":"OmniJigsaw: Enhancing Omni-Modal Reasoning via Modality-Orchestrated Reordering","authors":["Yiduo Jia","Muzhi Zhu","Hao Zhong","Mingyu Liu","Yuling Xi","Hao Chen","Bin Qin","Yongjie Yang","Zhenbo Luo","Chunhua Shen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"OmniJigsaw: Enhancing Omni-Modal Reasoning via Modality-Orchestrated Reordering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.08209","noteUrl":"/review?paper=review-neurips-2026-151826","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151828","title":"HIMMEL: Hierarchical Interleaved Multi-stream Motion Encoding for Long Video Understanding","authors":["Haopeng Jin","Tiankun Yang","Zhenyu Guan","ShiQuan Dong","Wenlong Zhao","Hongzhu Yi","Chubin Chen","Tao Yu","Jinwen Luo","Yujia Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"HIMMEL: Hierarchical Interleaved Multi-stream Motion Encoding for Long Video Understanding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08158","noteUrl":"/review?paper=review-neurips-2026-151828","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151831","title":"CAIRN: Cross-Room 3D Scene Understanding with Topology-Aware Large Multimodal Models","authors":["He Liang","Chenyang Ma","Yiming Zhang","Sangyun Shin","Andrew Markham","Niki Trigoni","Yuhang He"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"CAIRN: Cross-Room 3D Scene Understanding with Topology-Aware Large Multimodal Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.06534","noteUrl":"/review?paper=review-neurips-2026-151831","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151832","title":"The Compliance Trap: Diagnosing How AI Agents Consume Conflicting Memory","authors":["Yixiong Chen","Xinyi Bai","Alan Yuille"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Compliance Trap: Diagnosing How AI Agents Consume Conflicting Memory","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.10608","noteUrl":"/review?paper=review-neurips-2026-151832","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151833","title":"Bridging Scene Generation and Planning: Driving with World Model via Unifying Vision and Motion Representation","authors":["Xingtai Gui","Meijie Zhang","Lei Zhou","Wencheng Han","jiahao gong","Feiyang Tan","Cheng-Zhong Xu","Jianbing Shen"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Bridging Scene Generation and Planning: Driving with World Model via Unifying Vision and Motion Representation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.14948","noteUrl":"/review?paper=review-neurips-2026-151833","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151835","title":"The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning","authors":["Jing Liang","Hongyao Tang","Yi Ma","Yancheng He","Weixun Wang","Xiaoyang Li","Wenbo Su","Jinyi Liu","YAN ZHENG","Jianye Hao","Bo Zheng"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A maximizer of the clipped Step-1 surrogate need not improve on the rollout policy","plainExplanation":"Clipping the update ratio does not impose a hard constraint on the new policy. The surrogate can have global maximizers with lower expected reward than the rollout policy, so the introductory nondegradation guarantee does not follow.","primaryUrl":"https://arxiv.org/html/2606.29526","noteUrl":"/review?paper=review-neurips-2026-151835","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151837","title":"Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot","authors":["Jorge Chang Ortega","Bastien Le Lan","Thomas Serre","Victor Boutin"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The scalar composition does not belong to the decoder domain","plainExplanation":"The appendix composes a scalar predictor \\(p_\\theta\\) into a decoder defined on \\(\\mathbb R^d\\), with \\(d>1\\), and derives saliency by its chain rule. The displayed composition is undefined because a scalar is not a decoder latent vector.","primaryUrl":"https://arxiv.org/html/2605.23819","noteUrl":"/review?paper=review-neurips-2026-151837","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151838","title":"PO-PDDL: Learning Symbolic POMDPs from Visual Demonstrations for Robot Planning Under Uncertainty","authors":["Wenjing Tang","Jin Xuanjin","Yuan Liu","RenMing Huang","Cewu Lu","Panpan Cai"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"PO-PDDL: Learning Symbolic POMDPs from Visual Demonstrations for Robot Planning Under Uncertainty","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.15654","noteUrl":"/review?paper=review-neurips-2026-151838","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151839","title":"CLeaR: A Unified Framework for Resolving the Leakage–Degradation Dilemma in Style Transfer","authors":["Teng Zhou","Yunhao Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"CLeaR: A Unified Framework for Resolving the Leakage–Degradation Dilemma in Style Transfer","plainExplanation":"The first-order covariance calculation does not transfer to the exact nonlinear minimizer under the stated assumptions, and the broad decrease-in-error claim needs control of geometry and noise as models are added.","primaryUrl":"https://arxiv.org/html/2609.38136","noteUrl":"/review?paper=review-neurips-2026-151839","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151840","title":"Training-Free Cultural Alignment of Large Language Models via Persona Disagreement","authors":["Dao S Minh","Trung-Kiet Huynh","Chi Nguyen Tran","Phu-Quy Nguyen-Lam","Phu-Hoa Pham","Tuan Nguyen","The Anh Han","Long Tran-Thanh"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A zero variance floor admits no finite multiplicative approximation","plainExplanation":"The consensus-variance model has zero asymptotic floor, making its finite multiplicative-floor claim impossible. The deployment cap also bounds only a residual correction and omits the algorithm’s separate persona-anchor displacement.","primaryUrl":"https://arxiv.org/html/2605.10843","noteUrl":"/review?paper=review-neurips-2026-151840","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151841","title":"FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution","authors":["Shuhong Liu","Xining Ge","quanfeng xu","Ziteng Cui","Liuzhuozheng Li","Gengjia Chang","Jun Liu","Ziying Gu","Dong Li","Xuangeng Chu","Lin Gu","Tatsuya Harada"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"More steps do not proportionally shrink this correction schedule","plainExplanation":"The explanation says increasing \\(T\\) weakens every correction \\(\\eta_i=\\eta_0(1-i/T)\\) proportionally. At fixed relative index \\(i/T\\), the coefficient is unchanged; at fixed positive \\(i\\), it increases with \\(T\\).","primaryUrl":"https://arxiv.org/html/2605.03749","noteUrl":"/review?paper=review-neurips-2026-151841","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151842","title":"CAFE: Causally-Guided Automated Feature Engineering with Multi-Agent Reinforcement Learning","authors":["Arun Vignesh Malarkkan","Wangyang Ying","Hongyu Cao","Dongjie Wang","Yanjie Fu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The reward multiplier strengthens negative penalties","plainExplanation":"For a negative performance change, stronger causal weighting makes the reward more negative. That is the opposite of attenuation.","primaryUrl":"https://arxiv.org/html/2602.16435","noteUrl":"/review?paper=review-neurips-2026-151842","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151843","title":"FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale","authors":["Runyuan He","Qiuyang Mang","Shang Zhou","Kaiyuan Liu","Hanchen Li","Huanzhi Mao","Qizheng Zhang","Zerui Li","Bo Peng","Lufeng Cheng","Tianfu Fu","Yichuan Wang","Wenhao Chai","Jingbo Shang","Alex Dimakis","Joseph Gonzalez","Alvin Cheung"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.14445","noteUrl":"/review?paper=review-neurips-2026-151843","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151845","title":"Transformers Converge to Invariant Algorithmic Cores","authors":["Joshua Schiffman"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Transformers Converge to Invariant Algorithmic Cores","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.22600","noteUrl":"/review?paper=review-neurips-2026-151845","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151847","title":"Prediction-Intervention Games and Invariant Sets","authors":["Linus Kühne","Felix Schur","Jonas Peters"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Prediction-Intervention Games and Invariant Sets","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16828","noteUrl":"/review?paper=review-neurips-2026-151847","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151849","title":"Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator","authors":["Zihan Wang","Seungjun Lee","Yinghao Xu","Gim Hee Lee"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.05765","noteUrl":"/review?paper=review-neurips-2026-151849","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151852","title":"Trust Guided Decision Transformer","authors":["Chainesh Gautam","Raghuram Bharadwaj Diddigi","Chandramouli Kamanchi","Pankaj Dayama","Sumanta Mukherjee","Kameshwaran Sampath"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Trust Guided Decision Transformer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31586","noteUrl":"/review?paper=review-neurips-2026-151852","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151853","title":"Ultra Fast PDE Solving via Physics Guided Few-step Diffusion","authors":["Xiangrui Cindy Kong","Yueqi Wang","Haoyang Zheng","Weijian Luo","Guang Lin"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Convergence of the physics-penalized objective does not imply teacher-score matching","plainExplanation":"The auxiliary objective targets the student’s score. Stationarity of the penalized generator objective only constrains the sum of its distillation and physics gradients; it does not make each term vanish or establish equality with the teacher score.","primaryUrl":"https://arxiv.org/html/2602.03627","noteUrl":"/review?paper=review-neurips-2026-151853","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151854","title":"Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning","authors":["Meng Lou","Hanzhong Guo","Linwei Chen","Yizhou Yu"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09640","noteUrl":"/review?paper=review-neurips-2026-151854","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151862","title":"FLAME: Flow Enhanced Legendre Memory Models for General Time Series Forecasting","authors":["Xingjian Wu","Hanyin Cheng","Xiangfei Qiu","Zhengyu Li","Jilin Hu","Chenjuan Guo","Bin Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FLAME: Flow Enhanced Legendre Memory Models for General Time Series Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.14253","noteUrl":"/review?paper=review-neurips-2026-151862","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151863","title":"Flash-SD-KDE: Accelerating SD-KDE with Tensor Cores","authors":["Elliot Epstein","Rajat Vadiraj Dwaraknath","John Winnicki"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Empirical score shifting does not equal Laplace correction","plainExplanation":"Empirical score shifting does not coincide with Laplace correction, and the transported empirical-density formula omits its smoothing operator. The roofline calculation also switches between quadratic tiled traffic and a linear input/output footprint.","primaryUrl":"https://arxiv.org/html/2602.10378","noteUrl":"/review?paper=review-neurips-2026-151863","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151867","title":"OmniSpace: Efficient Geometry Awareness for Autonomous Vehicles MLLMs","authors":["Anh Hao Vo","Phu Loc Nguyen","Khoa Vo","Sieu Tran","Duc Nguyen","Ngo X Cuong","Nghi Bui","Anh Nguyen","Duy M. H. Nguyen","Ngan Le"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The written geometry output needs its residual term","plainExplanation":"Without an additive skip term, zeroing the final projection removes the feature instead of preserving it.","primaryUrl":"https://arxiv.org/html/2606.22617","noteUrl":"/review?paper=review-neurips-2026-151867","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151868","title":"ReGDiff: Guided Diffusion in Regulated Latent Space for Exploring Metamaterial Voxel Geometry","authors":["Wangzhi Zhan","Jianpeng Chen","Dongqi Fu","Dawei Zhou"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"ReGDiff: Guided Diffusion in Regulated Latent Space for Exploring Metamaterial Voxel Geometry","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34231","noteUrl":"/review?paper=review-neurips-2026-151868","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151870","title":"VeriGraph: Towards Verifiable Data-Analytic Agents","authors":["Jiajie Jin","Zhao Yang","WenLe Liao","Yuyang Hu","Guanting Dong","Xiaoxi Li","Yutao Zhu","Zhicheng Dou"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"VeriGraph: Towards Verifiable Data-Analytic Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.16603","noteUrl":"/review?paper=review-neurips-2026-151870","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151871","title":"Drifting Field Policy: Wasserstein Gradient Flow on Policy Space for Offline-to-Online RL","authors":["Juil Koo","Mingue Park","Jiwon Choi","Yunhong Min","Minhyuk Sung"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A shared weak limit does not imply a vanishing total-variation gap","plainExplanation":"The empirical-measure argument uses impossible total-variation convergence to a nonatomic target, although weak kernel-integral convergence can repair it. Hard truncation and soft tilting can also share a sharp weak limit while their total-variation gap tends to one.","primaryUrl":"https://arxiv.org/html/2605.07727","noteUrl":"/review?paper=review-neurips-2026-151871","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151874","title":"Dynamic Video Generation: Shaping Video Generation Across Time and Space","authors":["Shikang Zheng","Jingkai Huang","Jiacheng Liu","Guantao Chen","Lixuan He","Yuqi Lin","Peiliang Cai","Linfeng Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The diffusion marginal needs cumulative retention","plainExplanation":"The signal retained over several Markov steps is the product of their per-step retentions.","primaryUrl":"https://arxiv.org/html/2605.21042","noteUrl":"/review?paper=review-neurips-2026-151874","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151878","title":"Denoising Distances in Metric Measure Spaces","authors":["Han Huang","Pakawut Jiradilok","Elchanan Mossel"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Denoising Distances in Metric Measure Spaces","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.18301","noteUrl":"/review?paper=review-neurips-2026-151878","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151879","title":"Bridging Structure and Language: Graph-Based Visual Reasoning for Autonomous Road Understanding","authors":["Lena Wild","Katie Luo","Marco Pavone"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Matching window endpoints does not mean the state stayed fixed","plainExplanation":"The answer generator asks whether any state change occurred in a fully known interval but compares only its first and last states. A red-green-green-red sequence has matching endpoints despite two changes.","primaryUrl":"https://arxiv.org/html/2605.20942","noteUrl":"/review?paper=review-neurips-2026-151879","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151883","title":"Towards Understanding Self-Pretraining for Sequence Classification","authors":["Omar Coser","Loredana Zollo","Paolo Soda","Antonio Orvieto"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Towards Understanding Self-Pretraining for Sequence Classification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21070","noteUrl":"/review?paper=review-neurips-2026-151883","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151884","title":"Token-to-Token Alignment of Text Embeddings for Semantic Blending","authors":["Saar Huberman","Ron Mokady","Or Patashnik","Daniel Cohen-or"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Token-to-Token Alignment of Text Embeddings for Semantic Blending","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.24021","noteUrl":"/review?paper=review-neurips-2026-151884","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151885","title":"OmniEgoCap: Camera-Agnostic Sequence-Level Egocentric Motion Reconstruction","authors":["Kyungwon Cho","Jeonghyeon Na","Hanbyul Joo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OmniEgoCap: Camera-Agnostic Sequence-Level Egocentric Motion Reconstruction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.19283","noteUrl":"/review?paper=review-neurips-2026-151885","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151886","title":"Dynamic Treatment on Networks","authors":["Bengusu Nar","Jiguang Li","Veronika Rockova","Panos Toulis"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Dynamic Treatment on Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06564","noteUrl":"/review?paper=review-neurips-2026-151886","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151887","title":"RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers","authors":["Mengyuan Fan","Bokai Huang","JiaMing Pan","Xiaokun Yuan","Peizhuang Cong","Zhewen Tan","Tong Yang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The main residual operator retains more directions than the advertised code alphabet","plainExplanation":"The default continuous residual-direction operator retains more per-entry angles than the finite code alphabet can store. The stated operator and effective-bit-width accounting therefore need a consistent quantization definition.","primaryUrl":"https://arxiv.org/html/2610.04457","noteUrl":"/review?paper=review-neurips-2026-151887","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151889","title":"What Does the Brain See? Multiview Neural Representation to Demystify the Brain-Visual Alignment","authors":["Salini Yadav","Taveena Lotey","Pravendra Singh","Partha P Roy"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The displayed contrastive loss contains only one alignment direction","plainExplanation":"The loss called symmetric contains only EEG-to-image alignment. Reverse alignment has different normalization denominators and is not supplied by freezing the image encoder. The mismatch concerns the mathematical symmetry claimed for the displayed objective. It does not establish an implementation or empirical failure.","primaryUrl":"https://arxiv.org/html/2606.25718v2","noteUrl":"/review?paper=review-neurips-2026-151889","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151892","title":"When the Merge Coefficient Stops Mattering: Proximity Regularized Merging for Continual LoRA Adaptation","authors":["Yixuan Liu","Yuhao Sun","Sen Song","Li Jin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The scalar merge ratio can leave the admissible interval","plainExplanation":"Equation (2) requires merge coefficient \\(\\alpha_t\\in[0,1]\\), while Algorithm 1 computes an unconstrained ratio. The ratio has no inherent interval guarantee.","primaryUrl":"https://arxiv.org/html/2609.32332","noteUrl":"/review?paper=review-neurips-2026-151892","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151894","title":"PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution","authors":["Zihang Xu","Xiaoyang Liu","郑 陈","Yulun Zhang","Xiaokang Yang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13027","noteUrl":"/review?paper=review-neurips-2026-151894","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151896","title":"ForceFlow: Learning to Feel and Act via Contact-Driven Flow Matching","authors":["Shuoheng Zhang","Yifu Yuan","Hongyao Tang","YAN ZHENG","Qiaojun Yu","Pengyi Li","Guowei Huang","Helong Huang","Xingyue Quan","Jianye Hao"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"ForceFlow: Learning to Feel and Act via Contact-Driven Flow Matching","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11048","noteUrl":"/review?paper=review-neurips-2026-151896","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151897","title":"Rosetta: Composable Native Multimodal Pretraining","authors":["Xiangyue Liu","Zijian Zhang","Miles Yang","Zhao Zhong","Liefeng Bo","Ping Tan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Momentum orthogonality does not guarantee preservation of an earlier task","plainExplanation":"The regularized projection leaves a small nonzero conflict residual despite its exact-orthogonality proof. Even exact orthogonality to historical momentum does not establish preservation of the current base-task loss under the actual optimizer.","primaryUrl":"https://arxiv.org/html/2607.00293","noteUrl":"/review?paper=review-neurips-2026-151897","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-151898","title":"DiagEval: Trajectory-Conditioned Diagnosis for Reliable Software Evaluation with GUI Agents","authors":["Sirui Hong","Liuzhijie","Tengfei Li","Wei Tao","Yifan Wu","Chenglin Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DiagEval: Trajectory-Conditioned Diagnosis for Reliable Software Evaluation with GUI Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17439","noteUrl":"/review?paper=review-neurips-2026-151898","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151899","title":"Learnable Spectral Activations","authors":["Tamir Shor","Or Litany","Alex M Bronstein"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learnable Spectral Activations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07419","noteUrl":"/review?paper=review-neurips-2026-151899","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151906","title":"Representation Learning Enables Scalable Multitask Deep Reinforcement Learning","authors":["Johan Obando Ceron","Lu Li","Scott Fujimoto","Pierre-Luc Bacon","Aaron Courville","Pablo Samuel Castro"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Representation Learning Enables Scalable Multitask Deep Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05555","noteUrl":"/review?paper=review-neurips-2026-151906","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151909","title":"The Sharp Directions Are Against You: Curvature Analysis of Activation Steering","authors":["Seyedarmin Azizi","Arya Fayyazi","Parsa Razmara","Erfan Baghaei Potraghloo","Souvik Kundu","Massoud Pedram"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A ratio of spectral bounds is not a guaranteed curvature reduction","plainExplanation":"The ratio of spectral upper bounds is not a guaranteed ratio of actual steering curvatures. The claimed invariance to steering strength also overlooks the Hessian’s changing evaluation point.","primaryUrl":"https://arxiv.org/html/2610.16084","noteUrl":"/review?paper=review-neurips-2026-151909","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151912","title":"The Dynamics of Policy Gradient in Social Dilemmas with Partner Selection","authors":["Benedict Russell","Chin-wing Leung","Paolo Turrini"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Reward covariance contains action-dependent partner-retention effects","plainExplanation":"The OFT variance calculation substitutes covariance of latent opponent cooperation types for covariance of sampled rewards, conditional on the focal first action being cooperation. Write \\(\\mu_j=\\mathbb E[Y^j]\\).","primaryUrl":"https://arxiv.org/html/2605.18185","noteUrl":"/review?paper=review-neurips-2026-151912","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151913","title":"GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models","authors":["Xiaohang Tang","Keyue Jiang","Che Liu","Zhao","Xiaoxiao Xu","Sangwoong Yoon","Ilija Bogunovic"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The stated optimizer uses different regularization coefficients","plainExplanation":"The objective rescales both the policy-mixture coefficients and advantage strength. Its stated optimizer does not use those rescaled coefficients. The objective/solution correspondence is affected; a differently parameterized regularized objective can yield the proposed policy.","primaryUrl":"https://arxiv.org/html/2605.29398","noteUrl":"/review?paper=review-neurips-2026-151913","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151914","title":"When Confidence Rises Too Early: Detecting Shortcut Reasoning via Premature Answer Commitment","authors":["Zhaohan Zhang","Junjie Liu","Chengzhengxu Li","Chen Shen","Xiaoming Liu","Chao Shen","Jieping Ye","Ziquan Liu","Ioannis Patras"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"When Confidence Rises Too Early: Detecting Shortcut Reasoning via Premature Answer Commitment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35074","noteUrl":"/review?paper=review-neurips-2026-151914","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151916","title":"Cost-Aware Best-LLM Identification using Dueling Feedback","authors":["Sarvesh Gharat","Nikhil Karamchandani","Jayakrishnan Nair"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Cost-Aware Best-LLM Identification using Dueling Feedback","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30360","noteUrl":"/review?paper=review-neurips-2026-151916","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151918","title":"NormLift: From Lifted Features To Semantic Reliability In 3D Gaussian Splatting","authors":["Yihan Zang","Da Li","Dominik Engel","Shinkyu Park","Ivan Viola"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Jaccard independence comparison uses the wrong denominator","plainExplanation":"For two independent \\(20\\%\\) selections, the population Jaccard baseline is \\(1/9\\), approximately \\(0.111\\), so \\(0.162\\) is above that benchmark. This corrects an auxiliary probability comparison; it does not alter the measured overlap or the separately reported rank correlations.","primaryUrl":"https://arxiv.org/html/2609.18898","noteUrl":"/review?paper=review-neurips-2026-151918","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151919","title":"VolFill: Single-View Amodal 3D Scene Reconstruction with Volumetric Flow Matching","authors":["Tuan D Ngo","Chuang Gan","Evangelos Kalogerakis"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"VolFill: Single-View Amodal 3D Scene Reconstruction with Volumetric Flow Matching","plainExplanation":"No issue was found in the checked claims. The loss-function mismatch is corrected.","primaryUrl":"https://arxiv.org/html/2605.31466","noteUrl":"/review?paper=review-neurips-2026-151919","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151920","title":"StateTree: Enhancing Long-term Dialogue Reasoning via Reinforcement Learning","authors":["Naen Xu","Wanqing Cui","Yibo Hu","Shixin Hong","Hengyu An","Meiguang Jin","Junfeng Ma","Tianyu Du"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"StateTree: Enhancing Long-term Dialogue Reasoning via Reinforcement Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.38809","noteUrl":"/review?paper=review-neurips-2026-151920","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151925","title":"Concepts Worth Having: Refining VLM-Guided Concept Bottleneck Models with Minimal Annotations","authors":["Nicola Debole","Stefano Teso","Andrea Passerini","Andrea Pugnana","Emanuele Marconato"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Concepts Worth Having: Refining VLM-Guided Concept Bottleneck Models with Minimal Annotations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16405","noteUrl":"/review?paper=review-neurips-2026-151925","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151926","title":"Generating Symmetric Materials using Latent Flow Matching","authors":["Anmar Karmush","Cedric M Brandenburg","Soheil Ershadrad","Johanna Rosen","Michael Felsberg","Filip Ekström Kelvinius"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The written integration grid includes an undefined endpoint","plainExplanation":"The printed grid includes \\(t=1\\), where the velocity expression is undefined, and its spacing is \\(1/(T-1)\\), inconsistent with the declared step.","primaryUrl":"https://arxiv.org/html/2605.10115","noteUrl":"/review?paper=review-neurips-2026-151926","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_ambiguity"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-151928","title":"Reward-Conditioned Reinforcement Learning","authors":["Michal Nauman","Marek Cygan","Pieter Abbeel"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Reward-Conditioned Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.05066","noteUrl":"/review?paper=review-neurips-2026-151928","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151929","title":"MESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Autoregressive Next-Frame Forecasting","authors":["Jiuming Liu","Jianing Li","Mengmeng Liu","Hongyang He","Hesheng Wang","Per O Kristensson"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Autoregressive Next-Frame Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10759","noteUrl":"/review?paper=review-neurips-2026-151929","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151933","title":"How Post-Training Shapes Biological Reasoning Models","authors":["Lukas Fesser","Hanlin Zhang","Michelle M Li","Eric Wang","Bryan Perozzi","Shek Azizi","Sham Kakade","Marinka Zitnik"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"How Post-Training Shapes Biological Reasoning Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.16517","noteUrl":"/review?paper=review-neurips-2026-151933","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151937","title":"Neural Garbage Collection: Learning to Forget while Learning to Reason","authors":["Michael Li","Jubayer Ibn Hamid","Emily Fox","Noah Goodman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The eviction score does not directly depend on every decoder parameter","plainExplanation":"Averaging only eviction scores changes the exact joint policy gradient under shared parameters. The optional forced-token likelihood is a surrogate rather than that action’s behavior score, and eviction scores cannot directly train parameters downstream of all scoring queries and keys.","primaryUrl":"https://arxiv.org/html/2604.18002","noteUrl":"/review?paper=review-neurips-2026-151937","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-151938","title":"Unlocking Fine-Grained Perception in CLIP via Structurally-Aware Latent Masked Modeling","authors":["Juntong Li","Lingwei Dang","Haomin Wu","Ziyan Qiu","Qingxin Xiao","Qingyao Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Unlocking Fine-Grained Perception in CLIP via Structurally-Aware Latent Masked Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07689","noteUrl":"/review?paper=review-neurips-2026-151938","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151940","title":"SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking","authors":["Zhiwei Li","Lei Zhu","Hao Gu","Xiang Hu","Yan Wang","Haitao Mi","Sirui Han","Leo Liang","Zhijiang Guo"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.13141","noteUrl":"/review?paper=review-neurips-2026-151940","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151942","title":"DrPO: Drifting Preference Optimization for One-Step Generative Models","authors":["Zhou Jiang","Yandong Wen","Zhen Liu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"DrPO: Drifting Preference Optimization for One-Step Generative Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02521","noteUrl":"/review?paper=review-neurips-2026-151942","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151943","title":"Explanation of Dynamic Physical Field Predictions using WassersteinGrad: Application to Autoregressive Weather Forecasting","authors":["Younes Essafouri","Laure Raynaud","Luciano DROZDA","Laurent Risser"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Compact spherical attention retains finite-time Lipschitz control","plainExplanation":"The argument transfers non-Lipschitz behavior of unbounded Euclidean attention to the stated attention ODE on a compact product of unit spheres, with fixed finite parameters. The vector field is smooth on that compact domain and therefore Lipschitz there.","primaryUrl":"https://arxiv.org/html/2604.22580","noteUrl":"/review?paper=review-neurips-2026-151943","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151946","title":"Computer Use at the Edge of the Statistical Precipice","authors":["Pierluca D Oro","Sneha Silwal","William R Wong","Yuxuan Sun","Fanyi Xiao","Manchen Wang","Eric Gan","Allen Bolourchi","Joseph Tighe"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Very small Wilson intervals need not have near-nominal pointwise coverage","plainExplanation":"Nominal Wilson coverage can fall well below its target with only three rollouts near a boundary probability. This limits the broad pointwise assurance while leaving the comparative recommendation and mixture simulation separate.","primaryUrl":"https://arxiv.org/html/2605.08261","noteUrl":"/review?paper=review-neurips-2026-151946","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151948","title":"Root-Selecting Fixed-Point Inversion for Rectified Flows via Trajectory Straightness","authors":["Semin Kim","Jihwan Yoon","Seunghoon Hong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Root-Selecting Fixed-Point Inversion for Rectified Flows via Trajectory Straightness","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.17584","noteUrl":"/review?paper=review-neurips-2026-151948","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151949","title":"AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems","authors":["Boxuan Zhang","Jianing Zhu","Zeru Shi","Dongfang Liu","Ruixiang Tang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08715","noteUrl":"/review?paper=review-neurips-2026-151949","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151951","title":"A Novel Schur-Decomposition-Based Weight Projection Method for Stable State-Space Neural-Network Architectures","authors":["Sergio Mauricio Vanegas Arias","Lasse Lensu","Fredy Ruiz"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The projection permits boundary dynamics that are not asymptotically stable","plainExplanation":"Projection onto the closed unit disc can preserve constant or growing zero-input trajectories, contradicting the strict asymptotic-stability claim. The spectral-matching definition also has a local constraint typo: binary row and column sums should be one.","primaryUrl":"https://arxiv.org/html/2605.14489","noteUrl":"/review?paper=review-neurips-2026-151951","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","definition_error","false_statement","formula_error"],"classifications":["major"]},{"slug":"review-neurips-2026-151952","title":"Relation-Aware Graph Foundation Model","authors":["Jianxiang Yu","Jiapeng Zhu","Yibo Zhao","HAO QIAN","Ziqi Liu","Zhiqiang Zhang","Xiang Li"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Relation-Aware Graph Foundation Model","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2505.12027","noteUrl":"/review?paper=review-neurips-2026-151952","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151954","title":"Post-hoc Selective Classification for Reliable Synthetic Image Detection","authors":["Kaixiang Zheng","Jacob Seidman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The AURC upper bound needs a harmonic factor","plainExplanation":"The proof imports $\\operatorname{AURC}<2\\Delta_{\\rm sele}$, which fails when errors have the highest confidence. Small score differences can make the pairwise logistic loss close to $\\log2$ without changing that harmful ordering. The imported inequality and the highest-confidence single-error construction were already refuted in [Zhou et al., Appendix A.2, Eqs. (34)–(35), arXiv v2](https://arxiv.org/pdf/2410.15361v2), posted on 6 February 2025. That earlier report concerns the same underlying inequality, before ReSIDe was public. The failure concerns Eqs. (7)–(8) for ordinary equally weighted empirical AURC. The harmonic-number correction bounds this empirical quantity; it does not by itself provide a population-risk bound or cover arbitrary mixture reweighting.","primaryUrl":"https://arxiv.org/html/2605.08574","noteUrl":"/review?paper=review-neurips-2026-151954","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151956","title":"Ensembits: an alphabet of protein conformational ensembles","authors":["Kaiwen Shi","Carlos Oliver"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Ensembits: an alphabet of protein conformational ensembles","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13789","noteUrl":"/review?paper=review-neurips-2026-151956","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151959","title":"PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing","authors":["Chengyu Fang","Chunming He","Yuelin Zhang","Chubin Chen","zhu","Hongqiu Wang","Longxiang Tang","Xiu Li","Sina Farsiu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.07048","noteUrl":"/review?paper=review-neurips-2026-151959","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151960","title":"TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention","authors":["David Huang","Guile Wu","Chengjie Huang","Bingbing Liu","DONGFENG BAI"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14315","noteUrl":"/review?paper=review-neurips-2026-151960","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151962","title":"AdapToPASS: Ambiguity-aware Adaptive Spherical Transformer for Panoramic Semantic Segmentation","authors":["Soumyaratna Debnath","weiming zhang","Shriram Damodaran","Dingwen Xiao","Lin Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"AdapToPASS: Ambiguity-aware Adaptive Spherical Transformer for Panoramic Semantic Segmentation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.29081","noteUrl":"/review?paper=review-neurips-2026-151962","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-151963","title":"Flash-KMeans: Fast and Memory-Efficient Exact K-Means","authors":["Shuo Yang","Haocheng Xi","Yilong Zhao","Muyang Li","Xiaoze Fan","Jintao Zhang","Han Cai","Yujun Lin","Xiuyu Li","Kurt Keutzer","Song Han","Chenfeng Xu","Ion Stoica"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Flash-KMeans: Fast and Memory-Efficient Exact K-Means","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.09229","noteUrl":"/review?paper=review-neurips-2026-151963","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151964","title":"Rethinking State Tracking in Recurrent Models Through Error Control Dynamics","authors":["Jiwan Chung","Heechan Choi","Seon Joo Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Rethinking State Tracking in Recurrent Models Through Error Control Dynamics","plainExplanation":"The projected recurrence needs a condition excluding cross-subspace leakage; common drift does not enlarge spread measured around current centroids; and mean radius alone cannot certify every point’s readability.","primaryUrl":"https://arxiv.org/html/2605.07755","noteUrl":"/review?paper=review-neurips-2026-151964","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151966","title":"Adaptive Calibration in Non-Stationary Environments","authors":["Junyan Liu","Haipeng Luo","Lillian Ratliff"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Sparse transition rows do not guarantee logarithmic power-iteration convergence","plainExplanation":"The runtime proof infers fast stationary-distribution convergence from a rounding matrix having two nonzero entries per row. For rows \\((1-e,e)\\) and \\((2e,1-2e)\\), the stationary law is \\((2/3,1/3)\\) and the other eigenvalue is \\(1-3e\\).","primaryUrl":"https://arxiv.org/html/2605.11490","noteUrl":"/review?paper=review-neurips-2026-151966","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151968","title":"PARE: Pruning and Adaptive Routing for Efficient Video Generation","authors":["Yutong Wang","Yunke Wang","Tianfan Xue","Yu Qiao","Yaohui WANG","Xinyuan Chen","Chang Xu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PARE: Pruning and Adaptive Routing for Efficient Video Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27336","noteUrl":"/review?paper=review-neurips-2026-151968","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151969","title":"Towards Direct Evaluation of Harness Optimizers via Priority Ranking","authors":["Kai Ong","Minseok Kang","Dongwook Choi","Junhee Cho","Seungwon Lim","Seungju Kim","Geunha Jang","Minwoo Oh","Bogyung Jeong","Sunghwan Kim","Taeyoon Kwon","Jihye Han","Juyoung wy","Jinyoung Yeo"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Towards Direct Evaluation of Harness Optimizers via Priority Ranking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22505","noteUrl":"/review?paper=review-neurips-2026-151969","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151970","title":"BACE: Behavior-Adaptive Connectivity Estimation from Multi-Region Neural Recordings","authors":["Mehrnaz Asadi","Sina Javadzadeh","Rahil Soroushmojdehi","Ali Mousavi","Terence Sanger"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The normalized off-diagonal sparsity penalty is constant","plainExplanation":"The pattern matrix has zero diagonal and unit absolute row sums, while its off-diagonal entrywise absolute sum is used as a sparsity penalty. The penalty is always \\(\\sum_{i\\ne j}|\\bar S_{ij}|=\\sum_i1=N\\).","primaryUrl":"https://arxiv.org/html/2510.20831v1","noteUrl":"/review?paper=review-neurips-2026-151970","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151971","title":"Annealing in variational inference mitigates mode collapse: a theoretical study on Gaussian mixtures","authors":["Luigi Fogliani","Bruno Loureiro","Marylou Gabrié"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The optimum requires the lower Lambert-W branch and a logarithmic correction","plainExplanation":"The optimum uses the lower Lambert-W branch, with a logarithmic correction absent from the printed closed form and asymptotic equivalence. Existence and uniqueness of the optimum and the objective being maximized remain valid.","primaryUrl":"https://arxiv.org/html/2602.12923","noteUrl":"/review?paper=review-neurips-2026-151971","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-151972","title":"SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion","authors":["Zizhao Hu","Ameya Godbole","Johnny Wei","Mohammad Rostami","Jesse Thomason","Robin Jia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The restricted loss removes the intended demotion gradient","plainExplanation":"The excluded logits do not enter this conditional loss. At initialization the two distributions on the retained set coincide, so the loss and its parameter gradient are zero.","primaryUrl":"https://arxiv.org/html/2605.07482","noteUrl":"/review?paper=review-neurips-2026-151972","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151973","title":"Long-Horizon Q-Learning: Accurate Value Learning via n-Step Inequalities","authors":["Armaan A Abraham","Lucy Xiaoyang Shi","Chelsea Finn"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Long-Horizon Q-Learning: Accurate Value Learning via n-Step Inequalities","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05812","noteUrl":"/review?paper=review-neurips-2026-151973","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151975","title":"Rectifying Categorical Flows on Statistical Manifolds for One-Step Generation","authors":["Xiaohuan Jia","Renzhe Xu","Xiao Wang","Jiayun Wu","Shaohua Fan"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The spherical-path prediction loss is not a likelihood upper bound","plainExplanation":"The spherical-path terminal prediction loss can fall below data entropy, so it cannot generally upper-bound any normalized model’s negative log-likelihood. Its normalized Text8 score is therefore a prediction-loss proxy rather than a likelihood-based BPC bound.","primaryUrl":"https://arxiv.org/html/2610.13225","noteUrl":"/review?paper=review-neurips-2026-151975","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-151978","title":"Localized Dynamics-Aware Domain Adaption for Off-Dynamics Offline Reinforcement Learning","authors":["Zhangjie Xia","Yu Yang","Pan Xu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The information identity, performance lemma and classifier ratio need separate corrections","plainExplanation":"A common representation marginal does not make these different divergences equal. The conditional distribution in the proof also changes when both next states are conditioned on.","primaryUrl":"https://arxiv.org/html/2602.21072","noteUrl":"/review?paper=review-neurips-2026-151978","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151980","title":"Reinforcement Learning with Multi-Step Lookahead Information Via Adaptive Batching","authors":["Nadav Merlis"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Reinforcement Learning with Multi-Step Lookahead Information Via Adaptive Batching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.10418","noteUrl":"/review?paper=review-neurips-2026-151980","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151981","title":"AttnDiff: Attention-based Differential Fingerprinting for Large Language Models","authors":["Haobo Zhang","Zhenhua Xu","Junxian Li","Shangfeng Sheng","Dezhang Kong","Meng Han"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"AttnDiff: Attention-based Differential Fingerprinting for Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.05502","noteUrl":"/review?paper=review-neurips-2026-151981","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151982","title":"On Generation in Metric Spaces","authors":["Jiaxun Li","Vinod Raman","Ambuj Tewari"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"On Generation in Metric Spaces","plainExplanation":"No issue was found in the checked claims. The obstruction in the 7 February 2026 manuscript does not apply to the revised closure dimension. Cover centers used to exhaust the closure must now be positive for every consistent hypothesis.","primaryUrl":"https://arxiv.org/html/2602.07710","noteUrl":"/review?paper=review-neurips-2026-151982","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151983","title":"DISTMOE: Rehearsal-free Routing in Mixture-of-Experts for Distributed Instruction Tuning","authors":["Mainak Singha","Niccolò Biondi","Elisa Ricci","Subhankar Roy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DISTMOE: Rehearsal-free Routing in Mixture-of-Experts for Distributed Instruction Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.09907","noteUrl":"/review?paper=review-neurips-2026-151983","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151986","title":"MobileWan: Closing the Quality Gap for Mobile Video Diffusion","authors":["Mohsen Ghafoorian","Denis Korzhenkov","Adil Karjauv","Ioannis Lelekas","Noor Fathima","Spyridon Stasis","Animesh Karnewar","Amirhossein Habibian"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A stabilizer applied only to one attention branch changes the function","plainExplanation":"The hybrid attention formula subtracts \\(c_t\\) only from local softmax logits while leaving its historical feature-map branch unscaled, calling this numerical stabilization. With one local key and one historical key of equal unshifted mass and values zero and one, the output is \\(1/(1+e^{-c_t})\\).","primaryUrl":"https://arxiv.org/html/2607.06173","noteUrl":"/review?paper=review-neurips-2026-151986","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151989","title":"FourierMoE: Fourier Mixture-of-Experts Adaptation of Large Language Models","authors":["Juyong Jiang","Fan Wang","Hong Qi","Sunghun Kim","Jing Tang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Batch reconstruction depends on the union of selected experts","plainExplanation":"The per-token active count does not bound the number of distinct reconstructions in a batch. Caching removes repeated reconstructions of the same expert, but different tokens can select disjoint expert sets.","primaryUrl":"https://arxiv.org/html/2604.01762","noteUrl":"/review?paper=review-neurips-2026-151989","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151992","title":"Shuffle and Joint Differential Privacy for Generalized Linear Contextual Bandits","authors":["Sahasrajit Sarmasarkar"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Shuffle and Joint Differential Privacy for Generalized Linear Contextual Bandits","plainExplanation":"Equal marginal counts do not make those counts independent of posterior error. The weighted regret inequality cannot be converted into an unweighted Hamming bound this way.","primaryUrl":"https://arxiv.org/html/2602.00417","noteUrl":"/review?paper=review-neurips-2026-151992","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","mathematical_error"],"classifications":["major"]},{"slug":"review-neurips-2026-151994","title":"BootstrapAgent: Turning Repository Setup into Reusable Agent Knowledge","authors":["Sihan Fu","Oucheng Liu","Shiyuan Wang","Jin Shi","Chengkun Wei"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BootstrapAgent: Turning Repository Setup into Reusable Agent Knowledge","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15815v1","noteUrl":"/review?paper=review-neurips-2026-151994","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-151995","title":"Beyond MMSE: Enhancing PnP Restoration with ProxiMAP","authors":["Kenta Vert","Giacomo Meanti","Scott Pesme","Michael Arbel","Julien Mairal"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Beating the MMSE density does not certify a global posterior mode","plainExplanation":"The residual-noise recursion is valid under the stated noiseless-output idealization, but an exact Gaussian MMSE example shows that it need not match the actual iterate distribution. A higher posterior density than the MMSE estimate also does not certify a global mode. The available one-sided curvature bound justifies the proposed step size after correcting the smoothness label. Separate local issues affect the printed DiffPIR noise scale and the Fourier forward model’s claimed nonlinearity and flip symmetry.","primaryUrl":"https://arxiv.org/html/2605.16396","noteUrl":"/review?paper=review-neurips-2026-151995","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement","formula_error","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-151998","title":"ProCLIP: Progressive Vision-Language Alignment via LLM-based Embedder","authors":["Xiaoxing Hu","Kaicheng Yang","Ziqi Ye","Ziyang Gong","Qi Ming","Zonghao Guo","Yu Tian","Xiang An","Ziyong Feng","Xue Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ProCLIP: Progressive Vision-Language Alignment via LLM-based Embedder","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.18795","noteUrl":"/review?paper=review-neurips-2026-151998","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152001","title":"CM2: Reinforcement Learning with Checklist Rewards for Multi-Turn and Multi-Step Agentic Tool Use","authors":["Zhen Zhang","Kaiqiang Song","Sean Wang","Yebowen Hu","Weixiang Yan","Chenyang Zhao","Henry P Zou","Haoyun Deng","Sathish Reddy Indurthi","Shujian Liu","Simin Ma","Xiaoyang Wang","Xin Wang","Song Wang"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"CM2: Reinforcement Learning with Checklist Rewards for Multi-Turn and Multi-Step Agentic Tool Use","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.12268","noteUrl":"/review?paper=review-neurips-2026-152001","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152002","title":"TracingFlow: A Simulation-Free Trajectory Inference Framework Based on Second-Order Dynamics","authors":["Yuhao Sun","Zekun Wu","Zixun Huang","Peijie Zhou"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"TracingFlow: A Simulation-Free Trajectory Inference Framework Based on Second-Order Dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.21070","noteUrl":"/review?paper=review-neurips-2026-152002","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152003","title":"Reinforcing Multimodal Reasoning Against Visual Degradation","authors":["Rui Liu","Dian Yu","Haolin Liu","Yucheng Shi","Tong Zheng","Runpeng(Leo) Dai","Haitao Mi","Pratap Tokekar"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The sampled token penalty squares the clean-law weights","plainExplanation":"The stopped clean probabilities are \\(p\\), augmented probabilities \\(q\\), and an on-policy token is sampled from \\(p\\). The penalty multiplies that token's log ratio by its clean probability again. The expected term is \\(\\sum_i p_i^2\\log(p_i/q_i)\\), not \\(\\mathrm{KL}(p\\|q)\\).","primaryUrl":"https://arxiv.org/html/2605.09262","noteUrl":"/review?paper=review-neurips-2026-152003","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152004","title":"Feature Information Dynamics in Diffusion","authors":["Jia-Shu Pan","Tao Zhang","Yufei Huang","Yanjun Sheng","Tailin Wu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Feature Information Dynamics in Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.08626","noteUrl":"/review?paper=review-neurips-2026-152004","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152006","title":"Conditional Counterfactual Mean Embeddings: Doubly Robust Estimation and Learning Rates","authors":["Thatchanon Anancharoenkij","Donlapark Ponnoprat"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The generic nuisance bound needs uniform control of the outcome estimator","plainExplanation":"The generic minimum-risk nuisance bound does not follow from an integrated outcome-estimator norm alone. The concrete Deep Feature and Neural-Kernel proofs already establish stronger pointwise bounds, which supply the missing condition for those uses. Separately, one Fourier indicator per annulus does not span the low-frequency subspace; the dependent approximation rate needs a complete-basis argument. CCME identification and the estimator definitions are not refuted by these findings.","primaryUrl":"https://arxiv.org/html/2602.04736","noteUrl":"/review?paper=review-neurips-2026-152006","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152012","title":"Flow-Transformed Implicit Processes for Function-Space Variational Inference","authors":["Luis A. Ortega","Andres Masegosa","Thomas Nielsen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Flow-Transformed Implicit Processes for Function-Space Variational Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.01954","noteUrl":"/review?paper=review-neurips-2026-152012","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152016","title":"Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models","authors":["Tobias Braun","Jonas Grebe","Hossein Shakibania","Anna Rohrbach","Marcus Rohrbach"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19227","noteUrl":"/review?paper=review-neurips-2026-152016","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152017","title":"Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents","authors":["Xiaofang Yang","Ziqi Miao","Dianbo Sui","Jing Shao","Lijun Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.01487","noteUrl":"/review?paper=review-neurips-2026-152017","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152018","title":"Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization","authors":["Yunzhen Yao","Hongye Wang","Yahong Wang","Michael Gastpar","Jiang Bo","Lie He"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06219","noteUrl":"/review?paper=review-neurips-2026-152018","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152019","title":"BiMoGen: Bidirectional Motion-Text Generation via Unified Masked Discrete Diffusion","authors":["Wanjiang Weng","Yongliang Wu","Xiaofeng Tan","Xingyu Zhu","Wenbo Zhu","Hongsong Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The unmasked inverse-time loss needs an integrability condition","plainExplanation":"Detaching the generated token controls differentiation through its selection; it does not make its reconstruction loss vanish with time.","primaryUrl":"https://arxiv.org/html/2609.35407","noteUrl":"/review?paper=review-neurips-2026-152019","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152021","title":"Flow Perturbation++: Multi-Step Unbiased Jacobian Estimation for High-Dimensional Boltzmann Sampling","authors":["xinpeng peng","Ang Gao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Splitting cancelling anisotropic steps can increase estimator variance","plainExplanation":"Cancelling anisotropic steps refute universal strict variance reduction. Separately, raw determinant unbiasedness does not extend to logarithms, and the VJP/JVP norm identity holds in distribution for isotropic probes rather than pointwise. The determinant identity and valid augmented-target SMC interpretation remain intact.","primaryUrl":"https://arxiv.org/html/2601.21177","noteUrl":"/review?paper=review-neurips-2026-152021","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","interpretation_error"],"classifications":["major"]},{"slug":"review-neurips-2026-152022","title":"Don’t Let Gains FADE: Breaking Down Policy Gradient Weights in RL","authors":["Juliette Decugis","Sean O'Brien","Francis Bach","Gabriel Synnaeve","Taco Cohen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Natural-gradient probability motion requires centering the action weight","plainExplanation":"The natural-gradient entropy argument omits centering, and the failure-mode residual factorization drops conditional covariance. The weight/mass table and difficulty objective have further definition mismatches. Its residual-norm comparison also confounds per-token variability with the number of tokens averaged.","primaryUrl":"https://arxiv.org/html/2607.01490","noteUrl":"/review?paper=review-neurips-2026-152022","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152023","title":"Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory","authors":["Jinzhuo Liu","Jiangning Zhang","Wencan Jiang","Yabiao Wang","Dingkang Liang","Xue zhucun","Ran Yi","Yong Liu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18733","noteUrl":"/review?paper=review-neurips-2026-152023","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152026","title":"Weisfeiler and Leman Follow the Arrow of Time: Expressive Power of Message Passing in Temporal Event Graphs","authors":["Franziska Heeg","Jonas Sauer","Petra Mutzel","Ingo Scholtes"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Weisfeiler and Leman Follow the Arrow of Time: Expressive Power of Message Passing in Temporal Event Graphs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.24438","noteUrl":"/review?paper=review-neurips-2026-152026","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152027","title":"LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation","authors":["Bo Jiang","Depu Meng","yihan hu","Yichen Xie","Tianshuo Xu","Wei Zhan"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23878","noteUrl":"/review?paper=review-neurips-2026-152027","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152028","title":"Control-Augmented Autoregressive Diffusion for Data Assimilation","authors":["Prakhar Srivastava","Farrin Marouf Sofian","Francesco Immorlano","Kushagra Pandey","Stephan Mandt"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The regularizer needs complete time support and an exact DDIM mean difference","plainExplanation":"A controlled transition can have positive KL even when no observation arrives. Restricting the sum to arrivals changes the objective rather than approximating each transition’s KL.","primaryUrl":"https://arxiv.org/html/2510.06637","noteUrl":"/review?paper=review-neurips-2026-152028","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152031","title":"Heterogeneous Judge-Aware Ranking with Sensitivity, Disagreement, and Confidence","authors":["Shibo Yu","Yingzhou Wang","Yan Chen","Guodong Li","Jin-Hong Du"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A lower bound on proximal coefficients does not ensure a geometric rate","plainExplanation":"The varying-parameter local convergence theorem assumes proximal coefficients bounded away from zero but supplies no upper bound. In the permitted centered rank-zero model with two items and four comparisons, three wins give curvature three at \\(m^*=\\log3/2\\).","primaryUrl":"https://arxiv.org/html/2605.05073","noteUrl":"/review?paper=review-neurips-2026-152031","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152032","title":"PEARL: Solver-in-the-Loop Interactive Optimization Modeling from Natural Language","authors":["Hongliang Lu","Zhong Li","Yuxuan Chen","Lan Yuan","Fan Zhang","Zaiwen Wen"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"PEARL: Solver-in-the-Loop Interactive Optimization Modeling from Natural Language","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.18256","noteUrl":"/review?paper=review-neurips-2026-152032","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152033","title":"Forecasting Downstream Performance of LLMs With Proxy Metrics","authors":["Arkil Patel","Siva Reddy","Marius Mosbach","Dzmitry Bahdanau"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Forecasting Downstream Performance of LLMs With Proxy Metrics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18607","noteUrl":"/review?paper=review-neurips-2026-152033","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152034","title":"Mean Testing under Truncation beyond Gaussian","authors":["Yuhao Wang","Roberto I Oliveira","Themis Gouleakis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mean Testing under Truncation beyond Gaussian","plainExplanation":"No issue was found in the checked claims. The source does not require that universal-scale interpretation. Theorem 11 does not label its \\(C\\) universal, and the following paragraph refers to suitable regularity parameters \\(c,r\\) for \\(P\\). Uniform error over the null and alternative within the stated class does not by itself require uniformity over all choices of class parameters. The Gaussian example changes those parameters as its variance grows, so it does not establish the alleged contradiction.","primaryUrl":"https://arxiv.org/html/2605.01335","noteUrl":"/review?paper=review-neurips-2026-152034","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152035","title":"Sort, Partition, Randomize: Optimal Binary Hypothesis Testing under Local Differential Privacy","authors":["Elena Ghazi","Jawad Nasser","Flavio Calmon","Ibrahim Issa"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Sort, Partition, Randomize: Optimal Binary Hypothesis Testing under Local Differential Privacy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.07443","noteUrl":"/review?paper=review-neurips-2026-152035","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152036","title":"MCGI: Manifold-Consistent Graph Indexing for Billion-Scale Disk-Resident Vector Search","authors":["Dongfang Zhao"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Reducing alpha can prune an edge that the larger-alpha rule keeps","plainExplanation":"The pruning proof omits the witness-distance restriction needed for RNG-lune containment. The high-LID edge-preservation claim also uses the wrong implication for smaller alpha, and a sorted two-candidate example contradicts it. A separate recall-gap lower bound can exceed the largest possible value of its left-hand side.","primaryUrl":"https://arxiv.org/html/2601.01930","noteUrl":"/review?paper=review-neurips-2026-152036","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152037","title":"LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution","authors":["Song Fei","Tian Ye","Sixiang Chen","Zhaohu Xing","Jianyu Lai","Lei Zhu"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.05947","noteUrl":"/review?paper=review-neurips-2026-152037","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152038","title":"Learning Survival Models with Right-Censored Reporting Delays","authors":["Yuta Shikuri","Hironori Fujisawa"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Survival Models with Right-Censored Reporting Delays","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.04421","noteUrl":"/review?paper=review-neurips-2026-152038","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152039","title":"GenEnv: Difficulty-Aligned Co-Evolution Between LLM Agents and Environment Simulators","authors":["Jiacheng Guo","Ling Yang","Peter Chen","Qixin Xiao","Yinjie Wang","Xinzhe Juan","Ke Shen","Mengdi Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The minimum-variance baseline is weighted by the policy score norm","plainExplanation":"The variance-minimizing REINFORCE baseline weights reward by squared score norm and need not equal mean reward under the stated boundedness assumption. The Bernoulli learning-signal lower bound also does not extend to arbitrary soft rewards by rescaling: constant rewards have zero centered learning signal at a nonextreme mean.","primaryUrl":"https://arxiv.org/html/2512.19682","noteUrl":"/review?paper=review-neurips-2026-152039","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152040","title":"Compute-Optimal Pretrain--Fine-tune in Ridge Gradient Flow","authors":["Alex Buna","Fanghui Liu","Patrick Rebeschini"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Compute-Optimal Pretrain--Fine-tune in Ridge Gradient Flow","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.16262v1","noteUrl":"/review?paper=review-neurips-2026-152040","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152041","title":"When is Warmstarting Effective for Scaling Language Models?","authors":["Neeratyoy Mallik","Maciej Janowski","Johannes Hog","Herilalaina Rakotoarison","Josif Grabocka","Frank Hutter","Aaron Klein"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"When is Warmstarting Effective for Scaling Language Models?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13405","noteUrl":"/review?paper=review-neurips-2026-152041","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152043","title":"SphereVAD: Training-Free Video Anomaly Detection via Geodesic Inference on the Unit Hypersphere","authors":["Chao Huang","Pengfei Wei","Wei Wang","Jie Wen","Zhihua Wang","Li Shen","Wenqi Ren","XIAOCHUN CAO"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Geodesic-distance logits are not the stated vMF posterior","plainExplanation":"Equation (4) interprets a sigmoid of \\(\\kappa(d_{\\rm norm}-d_{\\rm abn})\\) as a vMF likelihood-ratio anomaly probability. With one prototype per class, equal priors and equal concentration, exact log posterior odds are \\(\\kappa(\\mu_{\\rm abn}-\\mu_{\\rm norm})^\\top x=\\kappa(\\cos d_{\\rm abn}-\\cos d_{\\rm norm})\\).","primaryUrl":"https://arxiv.org/html/2605.08003","noteUrl":"/review?paper=review-neurips-2026-152043","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152044","title":"Beyond Adjacent Layers: Graph-Guided Layer Fusion for Compressing Large Language Models","authors":["Usama Muhammad","Summer Y Jung"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Beyond Adjacent Layers: Graph-Guided Layer Fusion for Compressing Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.15159","noteUrl":"/review?paper=review-neurips-2026-152044","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152049","title":"SCOPE-RL: Stable and Quantitative Control of Policy Entropy in RL Post-Training","authors":["Chen Wang","Zhaochun Li","Bai Jionghao","Hexuan Deng","Guanting Dong","Ge Lan","Yue Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SCOPE-RL: Stable and Quantitative Control of Policy Entropy in RL Post-Training","plainExplanation":"No issue was found in the checked claims. The arbitrary-temperature assertion criticized in the earlier manuscript is absent from v8. The corrected theorem evaluates the entropy drift and its derivative only at \\(T=1\\), and the appendix no longer transfers the argument to an arbitrary temperature by changing the base policy.","primaryUrl":"https://arxiv.org/html/2510.08141","noteUrl":"/review?paper=review-neurips-2026-152049","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152052","title":"Flow Map Language Models: One-step Language Modeling via Continuous Denoising","authors":["Chanhyuk Lee","Jaehoon Yoo","Manan Agarwal","Sheel Shah","Jerry Huang","Aditi Raghunathan","Seunghoon Hong","Nicholas Boffi","Jinwoo Kim"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The general denoiser-to-velocity formula differs from its own derivation","plainExplanation":"The general denoiser-to-velocity identity in Lemma C.8 has incorrect coefficients. Eliminating the conditional initial mean gives the formula already supported by the proof and by the correct linear-interpolation relation in Eq. (76).","primaryUrl":"https://arxiv.org/html/2602.16813","noteUrl":"/review?paper=review-neurips-2026-152052","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-152053","title":"On the Invariance and Generality of Neural Scaling Laws","authors":["Xing Han","Liu Ziyin","Suchi Saria","Paul Liang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"On the Invariance and Generality of Neural Scaling Laws","plainExplanation":"Losing input information does not imply losing information about the target.","primaryUrl":"https://arxiv.org/html/2605.07546","noteUrl":"/review?paper=review-neurips-2026-152053","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","invalid_inference","proof_gap","reporting_inconsistency"],"classifications":["major"]},{"slug":"review-neurips-2026-152054","title":"HYVE: Hybrid Views for LLM Context Engineering over Machine Data","authors":["Jian Tan","Fan Bu","Yuqing Gao","Devavrat Khanolkar","Jason Mackay","Boris Sobolev","Lei Jin","Li Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"HYVE: Hybrid Views for LLM Context Engineering over Machine Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.05400","noteUrl":"/review?paper=review-neurips-2026-152054","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152055","title":"Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation","authors":["Yuxuan Jiang","Runchao Li","Shubhashis Roy Dipta","Dawei Li","Zhao Yang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Conditioning a sampled log ratio differs from conditioning a context-wide KL","plainExplanation":"The token score is defined by conditioning the sampled log ratio on token identity, but the estimator instead conditions an already averaged context KL. At one fixed context, the context KL is identical for every token, whereas the token log ratio can differ and can be negative.","primaryUrl":"https://arxiv.org/html/2605.09253","noteUrl":"/review?paper=review-neurips-2026-152055","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152058","title":"Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures","authors":["José Lucas De Melo Costa","Seong Woo Ahn","Fabrice Popineau","Arpad Rimmel","Bich-Liên DOAN"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Exceeding a spectral threshold does not guarantee a nonzero stationary solution","plainExplanation":"The stationary equation requires exact spectral membership, so exceeding the critical scale does not by itself produce a nonzero solution. The origin also remains a global minimizer of the substituted nonnegative squared loss. A separate initialization calculation overstates the diagonal attention mass at the larger context size and uses an unjustified zero-mean-logit approximation.","primaryUrl":"https://arxiv.org/html/2610.02344","noteUrl":"/review?paper=review-neurips-2026-152058","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152060","title":"Learning to Drive in New Cities Without Human Demonstrations","authors":["Zilin Wang","Saeed Rahmani","Daphne Cornelisse","Bidipta Sarkar","Alexander D. Goldie","Jakob Foerster","Shimon Whiteson"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Learning to Drive in New Cities Without Human Demonstrations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.15891","noteUrl":"/review?paper=review-neurips-2026-152060","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152064","title":"Bayesian Additive Distribution Regression","authors":["Antonio R Linero","Jared Murray","Soumyabrata Bose"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Bayesian Additive Distribution Regression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.06462","noteUrl":"/review?paper=review-neurips-2026-152064","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152066","title":"Keep It CALM: Analyzing the Limits of Global Unsafety in Text-to-Image Generation","authors":["NaHyeon Park","Minhyun Lee","Hyunjung Shim"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Ridge-stabilized projection does not satisfy exact active constraints","plainExplanation":"The projection algorithm solves a stabilized active-set KKT system, then demands exact nonpositive feasibility residuals. Multiplying the stabilized solve back gives active residual \\(A_Sz+m=\\varepsilon\\mu\\), positive for positive active multipliers.","primaryUrl":"https://arxiv.org/html/2610.02300","noteUrl":"/review?paper=review-neurips-2026-152066","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152067","title":"Continuous Diffusion Scales Competitively with Discrete Diffusion for Language","authors":["Zhihan Yang","Wei Guo","Shuibai Zhang","Subham Sahoo","Yongxin Chen","Arash Vahdat","Morteza Mardani","John Thickstun"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Continuous Diffusion Scales Competitively with Discrete Diffusion for Language","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18530","noteUrl":"/review?paper=review-neurips-2026-152067","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152068","title":"EPIC: Efficient Predicate-Guided Inference-Time Control for Compositional Text-to-Image Generation","authors":["Sunung Mun","Sunghyun Cho","Jungseul Ok"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"EPIC: Efficient Predicate-Guided Inference-Time Control for Compositional Text-to-Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11722","noteUrl":"/review?paper=review-neurips-2026-152068","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152072","title":"GRASP: Guided Residual Adapters with Sample-wise Partitioning","authors":["Felix Nützel","Mischa Dombrowski","Bernhard Kainz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GRASP: Guided Residual Adapters with Sample-wise Partitioning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2512.01675","noteUrl":"/review?paper=review-neurips-2026-152072","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152075","title":"Learning CLI Agents with Structured Action Credit under Selective Observation","authors":["Haoyang Su","Ying Wen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning CLI Agents with Structured Action Credit under Selective Observation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08013","noteUrl":"/review?paper=review-neurips-2026-152075","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152076","title":"Collaborative Reasoning Distillation via Cross-Feedback and Coherent Curation","authors":["Taehoon Kim","Seunggeun Cho","Dongsu Han"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The dynamic program weights the first node differently","plainExplanation":"The path objective weights every quality by \\(\\lambda=0.7\\), but the initialization gives the first quality unit weight. This changes the selected path.","primaryUrl":"https://arxiv.org/html/2610.09587","noteUrl":"/review?paper=review-neurips-2026-152076","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152078","title":"BOOST: Power-Optimal Strong-FWER Testing for Block-Structured Multiplicity","authors":["Prasanjit Dubey","Xiaoming Huo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The ordinary Grenander estimator does not have the stated full-domain sup-norm rate","plainExplanation":"The analysis optimizes average rejection power but sometimes identifies it with the probability of rejecting the whole block. Several allocation and local-calibration formulas also need correction: concavity requires slope intervals, the ordinary Grenander estimator lacks the claimed full-domain rate, the printed constraint maps omit rejection gates, mixed-label sorting requires a permutation sum, and the Šidák perturbation proof reverses a budget implication. The two described block/tree baselines additionally lack the asserted individual strong-FWER guarantees; the conditional global theorem remains valid when its local-validity assumptions hold.","primaryUrl":"https://arxiv.org/html/2605.27664","noteUrl":"/review?paper=review-neurips-2026-152078","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","interpretation_error","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152079","title":"RotVLA: Rotational Latent Action for Vision-Language-Action Model","authors":["Qiwei Li","Xicheng Gong","Xinghang Li","Peiyan Li","Quanyun Zhou","Hangjun Ye","Jiahuan Zhou","Yadong Mu"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"RotVLA: Rotational Latent Action for Vision-Language-Action Model","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13403","noteUrl":"/review?paper=review-neurips-2026-152079","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152080","title":"When Riemann flows with Wasserstein: Generative Modeling of Probability Distributions on Manifolds","authors":["Doron Haviv","Edward De Brouwer","Rishabh Anand","Rex Ying","Gabriele Scalia","Hector Corrada Bravo"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"When Riemann flows with Wasserstein: Generative Modeling of Probability Distributions on Manifolds","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.25659","noteUrl":"/review?paper=review-neurips-2026-152080","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152081","title":"The Dynamic-Probabilistic Consistency Gap in Chaotic Surrogate Modeling","authors":["Andre Herz","Matthijs Pals","Daniel Durstewitz","Georgia Koppe"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Dynamic-Probabilistic Consistency Gap in Chaotic Surrogate Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31547","noteUrl":"/review?paper=review-neurips-2026-152081","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152083","title":"Temporal Backtracking Search for Test-time Generative Video Reasoning","authors":["SeJoon Jun","Zheng Ding","Huangyuan Su","Weirui Ye","Yilun Du"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Temporal Backtracking Search for Test-time Generative Video Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.13861","noteUrl":"/review?paper=review-neurips-2026-152083","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152084","title":"Ares: Adaptive Reasoning Effort Selection for Efficient LLM Agents","authors":["Jingbo Yang","Bairu Hou","Jiayun (Peter) Wang","Wei Wei","Yujia Bao","Shiyu Chang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Ares: Adaptive Reasoning Effort Selection for Efficient LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.07915","noteUrl":"/review?paper=review-neurips-2026-152084","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152087","title":"Clarification as Supervision: Reinforcement Learning for Vision-Language Interfaces","authors":["John Gkountouras","Ivan Titov"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The shared-parameter remark must distinguish detachment from a frozen reasoner","plainExplanation":"Detaching a computation changes the gradient estimator; it does not make the distribution of newly sampled reasoner outputs independent of updated shared parameters.","primaryUrl":"https://arxiv.org/html/2509.26594","noteUrl":"/review?paper=review-neurips-2026-152087","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["missing_assumption"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152089","title":"Expanding LLM Agent Boundaries with Strategy-Guided Exploration","authors":["Andrew Szot","Michael Kirchhof","Omar Attia","Alexander Toshev"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Expanding LLM Agent Boundaries with Strategy-Guided Exploration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.02045","noteUrl":"/review?paper=review-neurips-2026-152089","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152091","title":"Multi-view Relational Distillation for Spatial Reasoning with Vision-Language Models","authors":["Kiet Nguyen","Hanbo Shim","Jinwoo Kim","Seunghoon Hong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Multi-view Relational Distillation for Spatial Reasoning with Vision-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.10864","noteUrl":"/review?paper=review-neurips-2026-152091","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152097","title":"GMO-E²DIT: Grounded Multi-Operation Editing for E-Commerce Images","authors":["Zipeng Guo","Xiaoan Liu","Lichen Ma","Cheng Wang","Yu He","Xiaolong Fu","Jingling Fu","Xinyuan Shan","Shaojie Guo","Luohang Liu","Junshi Huang","Yan Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GMO-E²DIT: Grounded Multi-Operation Editing for E-Commerce Images","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.00920","noteUrl":"/review?paper=review-neurips-2026-152097","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152098","title":"AlignDrive: Aligned Lateral-Longitudinal Planning for End-to-End Autonomous Driving","authors":["Yanhao Wu","Haoyang Zhang","Fei He","Rui Wu","Yanhu Shan","Congpei Qiu","Liang Gao","Wei Ke","Tong Zhang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"AlignDrive: Aligned Lateral-Longitudinal Planning for End-to-End Autonomous Driving","plainExplanation":"A safe endpoint after an interior collision leaves the final original timestep eligible. The algorithm then chooses scale one and preserves the collision. More generally, pointwise safety before rescaling does not establish safety at the new spacetime positions.","primaryUrl":"https://arxiv.org/html/2601.01762","noteUrl":"/review?paper=review-neurips-2026-152098","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152099","title":"Language Model Goal Selection Differs from Humans' in a Self-Directed Learning Task","authors":["Gaia Molinaro","Dave August","Danielle Perszyk","Anne Collins"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Language Model Goal Selection Differs from Humans' in a Self-Directed Learning Task","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.03295","noteUrl":"/review?paper=review-neurips-2026-152099","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152102","title":"Training Optimal Large Diffusion Language Models","authors":["Jinjie Ni","Qian Liu","Chao Du","Longxu Dou","Hang Yan","Zili Wang","Tianyu Pang","Michael Shieh"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Training Optimal Large Diffusion Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.03280","noteUrl":"/review?paper=review-neurips-2026-152102","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152103","title":"Concise and Logically Consistent Conformal Sets for Neuro-Symbolic Concept-Based Models","authors":["Samuele Bortolotti","Emanuele Marconato","Andrea Pugnana","Andrea Passerini","Stefano Teso"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Concise and Logically Consistent Conformal Sets for Neuro-Symbolic Concept-Based Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18202","noteUrl":"/review?paper=review-neurips-2026-152103","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152104","title":"Reinforcement World Model Learning for LLM-based Agents","authors":["Xiao Yu","Baolin Peng","Ruize Xu","yelong shen","Pengcheng He","Suman Nath","Nikhil Singh","Jianfeng Gao","Zhou Yu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Importance weighting under old rollouts needs the old-policy denominator","plainExplanation":"The GRPO expectation samples the old policy but defines its importance ratio as new policy divided by a separate KL reference. To change an old-policy expectation into a new-policy one, the Radon–Nikodym ratio is \\(\\pi_{\\rm new}/\\pi_{\\rm old}\\).","primaryUrl":"https://arxiv.org/html/2602.05842","noteUrl":"/review?paper=review-neurips-2026-152104","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152105","title":"Shattered Compositionality: Counterintuitive Learning Dynamics of Transformers for Arithmetic","authors":["Xingyu Zhao","Darsh Sharma","Rheeya Uppaal","Yiqiao Zhong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The carry used in the proof differs from the local carry in the definition","plainExplanation":"Appendix B.2 defines the conditioning carry from current-column digits alone, whereas Appendix F includes the incoming carry as well. The incoming carry can change the outgoing integer carry with positive probability.","primaryUrl":"https://arxiv.org/html/2601.22510","noteUrl":"/review?paper=review-neurips-2026-152105","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152106","title":"UniPrefill: Universal Long-Context Prefill Acceleration via Block-wise Dynamic Sparsification","authors":["Qihang Fan","Huaibo Huang","zhiyingwu","Bingning Wang","Ran He"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The attention error bounds need per-query mass and operator controls","plainExplanation":"The selection rule bounds average dropped mass, not each query’s dropped mass. Recomputing softmax on retained keys also changes the retained weights.","primaryUrl":"https://arxiv.org/html/2605.06221","noteUrl":"/review?paper=review-neurips-2026-152106","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152108","title":"Principled Federated Random Forests for Heterogeneous Data","authors":["Rémi Khellaf","Erwan Scornet","Aurélien Bellet","Julie Josse"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Continuous quantile interpolation cannot uniformly approximate large empirical atoms","plainExplanation":"Theorem 3.1 claims uniform CDF error at most \\(1/B\\) for linear interpolation of reported empirical quantiles without an atom-free assumption or repeated-quantile rule. At an atom of mass \\(0.9\\), any continuous reconstructed CDF must differ by at least \\(0.45\\) from either the empirical left limit or its value at the atom.","primaryUrl":"https://arxiv.org/html/2602.03258","noteUrl":"/review?paper=review-neurips-2026-152108","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152109","title":"Not all uncertainty is alike: volatility, stochasticity, and exploration","authors":["Payam Piray"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The closed-form CAUSE bonus does not preserve both claimed monotonicities","plainExplanation":"The printed CAUSE approximation can reverse both claimed uncertainty responses: increasing volatility can lower its bonus, and increasing stochasticity can raise it. These counterexamples recompute all auxiliary quantities and concern the closed form, rather than the separate exact-index theorems.","primaryUrl":"https://arxiv.org/html/2605.19215","noteUrl":"/review?paper=review-neurips-2026-152109","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152111","title":"STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability","authors":["HAIPENG LUO","Qingfeng Sun","Song-Li Wu","Can Xu","Wenfeng Deng","Han Hu","Yansong Tang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Parameter updates change entropy through the logit Jacobian","plainExplanation":"A neural policy updates shared model parameters, not independent logits. With $J=\\partial z/\\partial\\theta$ and $u=e_a-p$, the parameter gradient induces logit velocity $\\widehat A JJ^\\top u$. Omitting $JJ^\\top$ can reverse the entropy derivative, rather than only rescale it. The entropy identity is valid for direct Euclidean logit updates, but not arbitrary neural-parameter GRPO updates. The corrected chain-rule expression covers the actual local gradient. This does not invalidate the reported empirical gains of STARE.","primaryUrl":"https://arxiv.org/html/2606.19236","noteUrl":"/review?paper=review-neurips-2026-152111","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152113","title":"Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume","authors":["Gregory Kang Ruey Lau","Hieu Dao","Nicole Hui Lin Kan","Bryan Kian Hsiang Low"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.24195","noteUrl":"/review?paper=review-neurips-2026-152113","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152114","title":"Generalized Wasserstein Flow Matching: Transport Plans, Everywhere, All at Once","authors":["Moritz Piening","Richard Duong","Gabriele Steidl"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Generalized Wasserstein Flow Matching: Transport Plans, Everywhere, All at Once","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08424","noteUrl":"/review?paper=review-neurips-2026-152114","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152117","title":"Generalized Priority-Aware Shapley Value","authors":["Kiljae Lee","Ziqi Liu","Weijing Tang","Yuan Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Generalized Priority-Aware Shapley Value","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15018","noteUrl":"/review?paper=review-neurips-2026-152117","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152121","title":"Why Learning Rediscovers the Closed-Form Diagonal Regularizer","authors":["Jeahn Han","Pyojin Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Why Learning Rediscovers the Closed-Form Diagonal Regularizer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.09656","noteUrl":"/review?paper=review-neurips-2026-152121","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152123","title":"From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs","authors":["Yuchuan Tian","Yuchen Liang","Shuo Zhang","Yingte Shu","Guangwen Yang","Wei He","Sibo Fang","Tianyu Guo","Kai Han","Chao Xu","Hanting Chen","Xinghao Chen","Yunhe Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.06776","noteUrl":"/review?paper=review-neurips-2026-152123","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152124","title":"Illusory Pattern Perception Drives Spurious Inference in Large Language Models","authors":["Peihua Mai","Zhuoyan Shao","Xinbao Qiao","Meng Zhang","Xinyue Zhou","Yan Pang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Illusory Pattern Perception Drives Spurious Inference in Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07791","noteUrl":"/review?paper=review-neurips-2026-152124","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152131","title":"Masked Visual Actions for Unified World Modeling","authors":["Hadi Alzayer","Wenlong Huang","Haonan Chen","Christopher Luey","Lvmin Zhang","Maneesh Agrawala","Gordon Wetzstein","Fei-Fei Li","Yilun Du","Jiajun Wu","Jia-Bin Huang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Masked Visual Actions for Unified World Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.19343","noteUrl":"/review?paper=review-neurips-2026-152131","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152133","title":"GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem","authors":["Rui-Xi Wang","Runzhong Wang","Connor Coley"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The separable Sinkhorn matrix removes every distance-dependent ranking signal","plainExplanation":"The contrastive construction defines \\(D_{ij}=d_i-d_j\\), exponentiates it as an entropic Sinkhorn kernel, and uses its normalized top-left entry as a ranking loss. The kernel factors as \\(e^{-D_{ij}/\\tau}=u_iv_j\\), a rank-one row/column potential.","primaryUrl":"https://arxiv.org/html/2606.29161","noteUrl":"/review?paper=review-neurips-2026-152133","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152137","title":"Channel-wise Vector Quantization","authors":["Wei Song","Tianhang Wang","Yitong Chen","Zuxuan Wu","Tong Zhang","Min Li","Jiaqi Wang","Kaicheng Yu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Channel-wise Vector Quantization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26089","noteUrl":"/review?paper=review-neurips-2026-152137","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152139","title":"Diffusion Path Samplers via Sequential Monte Carlo","authors":["James Matthew Young","Paula Cordero-Encinar","Sebastian Reich","Andrew Duncan","O. Deniz Akyildiz"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The inverse-polar target density omits the radial Jacobian","plainExplanation":"The rings target is described as an independent radius-and-angle pushforward, but its Cartesian density omits the radial Jacobian. Normalizing the printed expression gives second moment about $10.0675$, rather than the stated $7.5225$.","primaryUrl":"https://arxiv.org/html/2601.21951","noteUrl":"/review?paper=review-neurips-2026-152139","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152140","title":"TokenRouter: Efficient Serving System for Token-Level LLM Routing","authors":["Tianyu Fu","Tengxuan Liu","Ruoxi Wang","Yixin Dong","Yi Ge","Yichen You","Yu Wang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"TokenRouter: Efficient Serving System for Token-Level LLM Routing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.12242","noteUrl":"/review?paper=review-neurips-2026-152140","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152144","title":"Composing Diffusion Priors with Explicit Physical Context via Generative Gibbs Sampling","authors":["Weizhou Wang","Jonathan Weare","Aaron R Dinner"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Composing Diffusion Priors with Explicit Physical Context via Generative Gibbs Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10642","noteUrl":"/review?paper=review-neurips-2026-152144","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152145","title":"Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction","authors":["Ngoc Bui","Trung Hieu Nguyen","Arman Cohan","Rex Ying"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09649","noteUrl":"/review?paper=review-neurips-2026-152145","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152147","title":"Split the Differences, Pool the Rest: Provably Efficient Multi-Objective Imitation","authors":["Ziyad Sheebaelhamd","Luca Viano","Volkan Cevher","Claire Vernade"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Pareto-path proof needs a statewise-optimality condition, and the LP measures a domination gap","plainExplanation":"A policy can be optimal from the specified initial distribution while taking suboptimal actions at states it never reaches. Therefore arbitrary statewise combinations of two such policies need not remain optimal. This is a gap in the stated construction, not a counterexample to the existence of some Pareto path.","primaryUrl":"https://arxiv.org/html/2605.12000","noteUrl":"/review?paper=review-neurips-2026-152147","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","formula_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152148","title":"Generalization Error Bounds for Picard-Type Operator Learning in Nonlinear Parabolic PDEs","authors":["Koichi Taniguchi","Sho Sonoda"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Generalization Error Bounds for Picard-Type Operator Learning in Nonlinear Parabolic PDEs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10277","noteUrl":"/review?paper=review-neurips-2026-152148","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152149","title":"FedCF: Fair Federated Conformal Prediction","authors":["Anutam Srinivasan","Aditya T. Vadlamani","Amin Meghrazi","Srinivasan Parthasarathy"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FedCF: Fair Federated Conformal Prediction","plainExplanation":"The statistical certificate needs sampling uncertainty, and the private release needs a valid output distribution. The report separates those substantive failures from two explicit algebraic corrections.","primaryUrl":"https://arxiv.org/html/2509.22907","noteUrl":"/review?paper=review-neurips-2026-152149","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","incorrect_bound","invalid_inference","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-152150","title":"The SuperActivator Mechanism: Transformers Concentrate Reliable Concept Signals in the Tail","authors":["Cassandra Goldberg","Chaehyeon Kim","Adam Stein","Eric Wong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Bounded noise does not justify an exact successive contraction factor","plainExplanation":"An additive error that tends to zero need not be negligible relative to the shrinking deviation \\(u_\\ell\\). Thus this argument does not establish an exact successive-ratio limit. A geometric upper envelope is a weaker claim and is not contradicted here.","primaryUrl":"https://arxiv.org/abs/2512.05038v2","noteUrl":"/review?paper=review-neurips-2026-152150","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152151","title":"CoANeRV: Coordinate-Aware Token-Space Neural Video Representation","authors":["Jialong Guo","Ke Liu","MENGXUAN LI","Jiajun Bu","Haishuai Wang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"CoANeRV: Coordinate-Aware Token-Space Neural Video Representation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.13938","noteUrl":"/review?paper=review-neurips-2026-152151","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152152","title":"Generalization in Neural Networks Through the Lens of Magnitude Potential","authors":["Sahel Torkamani","Henry Gouk","Rik Sarkar"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The normalized potential limit counts tied nearest neighbors","plainExplanation":"The current v2 theorem still claims limit one at every exterior point. Equidistant nearest points contribute at the same leading scale, so the limit counts their multiplicity. The unique-nearest case and the separate empirical generalization proxy are not refuted.","primaryUrl":"https://arxiv.org/html/2610.01633","noteUrl":"/review?paper=review-neurips-2026-152152","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152153","title":"scTrilemma: Balancing Identity, Invariance, and Reconstruction in Single-Cell Representation Learning","authors":["Yunhak Oh","Yoonho Lee","Junseok Lee","Namkyeong Lee","Sang-Yeon Hwang","Yinhua Piao","Hyomin Kim","SEONGHWAN KIM","Jaechang Lim","Woo Youn Kim","Sungsoo Ahn","Chanyoung Park"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"scTrilemma: Balancing Identity, Invariance, and Reconstruction in Single-Cell Representation Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38840","noteUrl":"/review?paper=review-neurips-2026-152153","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152154","title":"Provable Speedups From Dynamic Population Sizes in Evolutionary Algorithms for Multiobjective Optimization","authors":["Andre Opris"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Uniform deletion is not the GSEMO replacement rule","plainExplanation":"Lemma 4 derives a population-distance recurrence by deleting uniformly from parent-plus-offspring, including the one-parent one-offspring GSEMO case. In the flat phase, the accepted equal-fitness child weakly dominates and replaces its parent with probability one, not one half.","primaryUrl":"https://arxiv.org/html/2607.23800","noteUrl":"/review?paper=review-neurips-2026-152154","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152156","title":"Information Templates: A New Paradigm for Intelligent Active Feature Acquisition","authors":["Hung-Tien Huang","Dzung Dinh","Junier Oliva"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Information Templates: A New Paradigm for Intelligent Active Feature Acquisition","plainExplanation":"The theorem names the wrong sign of the objective. Positive benefit does have diminishing returns, so that intuition should be retained.","primaryUrl":"https://arxiv.org/html/2508.18380","noteUrl":"/review?paper=review-neurips-2026-152156","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152157","title":"ITO: Multi-View Alignment and Training-Time Fusion for Image-Text Pretraining","authors":["Hanpeng Liu","Yaqian Li","Zidan Wang","Shuoxi Zhang","Zonglin Zhao","Zi-Hao Bo","rinyoichi takezoe","Kaiwen Long","Kun He"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ITO: Multi-View Alignment and Training-Time Fusion for Image-Text Pretraining","plainExplanation":"","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2603.02767","noteUrl":"/review?paper=review-neurips-2026-152157","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152161","title":"Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models","authors":["Ayush Rajesh Jhaveri","Anthony GX-Chen","Ilia Sucholutsky","Eunsol Choi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The stated rule admits 120 triples rather than 1,225","plainExplanation":"Table 9 overcounts the group-7 feasible set. Its ten allowed negative coordinate values yield $\\binom{10}{3}=120$ strict descending triples, rather than $1,225$.","primaryUrl":"https://arxiv.org/html/2604.02485","noteUrl":"/review?paper=review-neurips-2026-152161","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152162","title":"On the Depth of Monotone ReLU Neural Networks and ICNNs","authors":["Egor Bakaev","Florestan Brunck","Christoph Hertrich","Daniel Reichman","Amir Yehudayoff"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"One ReLU computes the allowed two-input maximum","plainExplanation":"The claimed ICNN depth lower bound fails for two inputs under the stated architecture: $\\max(x_1,x_2)=\\operatorname{ReLU}(x_1-x_2)+x_2$ uses one allowed ReLU, rather than the required two.","primaryUrl":"https://arxiv.org/html/2505.06169","noteUrl":"/review?paper=review-neurips-2026-152162","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152164","title":"Natural Synthesis: Outperforming Reactive Synthesis Tools with Large Reasoning Models","authors":["Frederik Schmitt","Matthias Cosler","Niklas Metzger","Julian Siber","Vladimir Krsmanovic","Oscar Ghanem","Bernd Finkbeiner"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Natural Synthesis: Outperforming Reactive Synthesis Tools with Large Reasoning Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15131","noteUrl":"/review?paper=review-neurips-2026-152164","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152165","title":"OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents","authors":["Shuang Chen","Kaituo Feng","Hangting Chen","Wenxuan Huang","Dasen Dai","Quanxin Shou","Yunlong Lin","Xiangyu Yue","Shenghua Gao","Tianyu Pang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05185","noteUrl":"/review?paper=review-neurips-2026-152165","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152166","title":"TriSpec: Ternary Speculative Decoding via Lightweight Proxy Verification","authors":["Haoyun Jiang","Junqi He","Feng Hong","Xinlong Yang","jianwei zhang","Zhengyang Zhuge","Zheng Li","Xiaofeng Cao","Zhiyong Chen","Bo Han","Junyang Lin","Jiangchao Yao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Rejection replacement needs the residual law","plainExplanation":"The stochastic decoder accepts draft law \\(q\\) toward proxy law \\(p\\) using the usual acceptance ratio, but samples rejected replacements from unchanged \\(p\\). Accepted mass is \\(\\min(q_i,p_i)\\).","primaryUrl":"https://arxiv.org/html/2601.23180","noteUrl":"/review?paper=review-neurips-2026-152166","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152168","title":"Tree Search With Predictions","authors":["Michael Dinitz","Bob Dong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Resetting inside a selected component can increase prediction error","plainExplanation":"The query proof assumes that projecting to a decomposition spine and resetting to its component start does not increase distance to the target. An allowed three-arm subdivided claw violates that assumption.","primaryUrl":"https://arxiv.org/html/2605.27490","noteUrl":"/review?paper=review-neurips-2026-152168","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152169","title":"Estimating the expected output of wide random MLPs more efficiently than sampling","authors":["Wilson Wu","Victor Lecomte","Michael Winer","George Robinson","Jacob Hilton","Paul Christiano"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Dropping vector-partition multiplicities changes the basic estimator and its error order","plainExplanation":"Equation (10) includes the number of labeled diagrams of each vector-partition type, but Equation (13) omits that multiplicity in the basic Hermite estimator. Two vertices with two legs each admit two cross-pairings of type \\(\\{(1,1),(1,1)\\}\\).","primaryUrl":"https://arxiv.org/html/2605.05179","noteUrl":"/review?paper=review-neurips-2026-152169","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152171","title":"Motion Cues from Image-based Point Tracking for LiDAR Scene Flow Estimation","authors":["Youngdong Jang","Gyeongrok Oh","Jong Wook Kim","Hyunju Ryu","Hyung-gun Chi","SeungHyeon Kim","Seungryong Kim","Jonghyun Choi","Sangpil Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The executable threshold has unintended operator precedence","plainExplanation":"Algorithm 1 uses a Python-style tensor threshold combining a numeric comparison with bitwise conjunction without the needed parentheses. Python parses the expression as a comparison against the bitwise conjunction.","primaryUrl":"https://arxiv.org/html/2605.16922","noteUrl":"/review?paper=review-neurips-2026-152171","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152172","title":"Overcoming Rank Collapse in Feedback Alignment","authors":["Gauthier Boeshertz","Razvan Pascanu","Claudia Clopath"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Overcoming Rank Collapse in Feedback Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11123","noteUrl":"/review?paper=review-neurips-2026-152172","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152173","title":"Agent-ToM: Learning to Monitor Autonomous LLM Agents via Theory-of-Mind Reasoning","authors":["Nesreen Ahmed","Nima Nafisi"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Agent-ToM: Learning to Monitor Autonomous LLM Agents via Theory-of-Mind Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24216","noteUrl":"/review?paper=review-neurips-2026-152173","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152175","title":"Entropy-Gated Latent Recursion","authors":["Soham Bhattacharjee","Dushyant Singh Chauhan","Salem Lahlou","Martin Takac","Nils Lukas"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Distinct replay spans can produce identical output distributions","plainExplanation":"The sufficient conditions for distinct EGLR distributions do not survive composition through norm matching and the replay blocks. An admissible open-gate transformer with nonzero intervening residuals produces identical distributions for two replay spans on every reachable continuation, refuting Proposition 1.","primaryUrl":"https://arxiv.org/html/2606.16620","noteUrl":"/review?paper=review-neurips-2026-152175","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-152180","title":"Continual Learning in Modern Hopfield Networks with an Application to Diffusion Models","authors":["Ken Takeda","Masafumi Oizumi","Ryo Karakida"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Continual Learning in Modern Hopfield Networks with an Application to Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27975","noteUrl":"/review?paper=review-neurips-2026-152180","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152181","title":"POETS: Uncertainty-Aware LLM Optimization via Compute-Efficient Policy Ensembles","authors":["Nicolas Menet","Andreas Krause","Abbas Rahimi"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"POETS: Uncertainty-Aware LLM Optimization via Compute-Efficient Policy Ensembles","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07775","noteUrl":"/review?paper=review-neurips-2026-152181","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152183","title":"Linguistic Trajectory Encoding for Efficient Long-Horizon Spatial Memory in Embodied Agents","authors":["Xie Tianyidan","Shenyi Wang","Qiang Tang","Mingjie Wang","Zhicheng Qiu","Xuanfu Li","Zhan Xu","Jian Yang","Lanjun Wang","Zili Yi"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Linguistic Trajectory Encoding for Efficient Long-Horizon Spatial Memory in Embodied Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.04802","noteUrl":"/review?paper=review-neurips-2026-152183","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152184","title":"Min-Max Optimization Requires Exponentially Many Queries","authors":["Martino Bernasconi","Matteo Castiglioni","Andrea Celli","Alexandros Hollender"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Min-Max Optimization Requires Exponentially Many Queries","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.13806v1","noteUrl":"/review?paper=review-neurips-2026-152184","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152185","title":"Hadamard Representation: Scaffolding Performance Across Model-free RL","authors":["Jacob Eeuwe Kooi","Zhao Yang","Mark Hoogendoorn","Vincent Francois-Lavet"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"A high activation-density peak does not identify saturation","plainExplanation":"The empirical dormancy statistic thresholds a Gaussian KDE peak, then is used to support a tanh saturation-probability mechanism involving neighborhoods of \\(\\pm1\\). A density-peak threshold measures scale and concentration, not location near saturation.","primaryUrl":"https://arxiv.org/html/2406.09079","noteUrl":"/review?paper=review-neurips-2026-152185","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152188","title":"Latent Riemannian Flow Matching for Geometry-Grounded 3D Foundation Models","authors":["Lisa Weijler","Irene Ballester","Guofeng Mei","Tolga Birdal","Pedro Hermosilla"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The sum product distance is not induced by the stated Riemannian norm","plainExplanation":"The method uses a sum of sphere-factor distances as its product Riemannian distance, while using the standard squared block-tangent norm. The infinitesimal norm of the sum distance is \\(F(v)=\\sum_i\\|v_i\\|\\).","primaryUrl":"https://arxiv.org/html/2607.19120","noteUrl":"/review?paper=review-neurips-2026-152188","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152191","title":"Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs","authors":["Quan Nguyen-Tri","Mukul Ranjan","Zhiqiang Shen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.26796","noteUrl":"/review?paper=review-neurips-2026-152191","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152192","title":"PrismFlow: Residual Dynamics for Flow Matching in Time-Series Generation","authors":["JUNRU ZHANG","Lang Feng","Jinbo Wang","Xu Guo","Yucheng Wang","Han Yu","Min Wu","Yabo Dong","Duanqing Xu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PrismFlow: Residual Dynamics for Flow Matching in Time-Series Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28867","noteUrl":"/review?paper=review-neurips-2026-152192","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152193","title":"Structured Transforms for Low-Overhead Quantization of Language Models","authors":["Daria Cherniuk","Alexander Rudikov","Boris Kashin","Ivan Oseledets"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Two signed components give at most four values, not four populated peaks","plainExplanation":"The construction allows at most four candidate values, rather than guaranteeing four populated peaks for every input. Reversing the projection schedule also need not preserve the realized residual norm on the same input. A separate background-algorithm branch projects the untransformed vector and can double its residual.","primaryUrl":"https://arxiv.org/html/2609.11687","noteUrl":"/review?paper=review-neurips-2026-152193","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152196","title":"Looped Diffusion Language Models","authors":["Sanghyun Lee","Chunsan Hong","Seungryong Kim","Jonghyun Lee","Jongho Park","Dongmin Park"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Looped Diffusion Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26106","noteUrl":"/review?paper=review-neurips-2026-152196","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152200","title":"When and Why SignSGD Outperforms SGD: A Theoretical Study Based on $\\ell_1$-norm Lower Bounds","authors":["Hongyi Tao","Dingzhi Yu","Lijun Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stochastic lower bound needs a finite-budget saturation regime","plainExplanation":"The lower bound is stated uniformly in noise and budget, but its proof drops the accuracy restriction used before inverting sample complexity. Smoothness and initial suboptimality \\(\\Delta\\) give \\(\\|\\nabla f(x_0)\\|\\le\\sqrt{2L\\Delta}\\), and the queried minimum includes this point.","primaryUrl":"https://arxiv.org/html/2605.06615","noteUrl":"/review?paper=review-neurips-2026-152200","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152201","title":"Multimodal LLMs Outperform Pathology Foundation Models in Cross-Domain Histological Similarity","authors":["Yishu Zhang","Yun Li","David Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Multimodal LLMs Outperform Pathology Foundation Models in Cross-Domain Histological Similarity","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.32876","noteUrl":"/review?paper=review-neurips-2026-152201","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152205","title":"Fix the Structural Bottleneck: Context Compression via Explicit Information Transmission","authors":["Jiangnan Ye","Hanqi Yan","Zhenyi Shen","Heng Chang","Ye Mao","Yulan He"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Fix the Structural Bottleneck: Context Compression via Explicit Information Transmission","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.03784","noteUrl":"/review?paper=review-neurips-2026-152205","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152206","title":"Co-Evolving Policy Distillation","authors":["Naibin Gu","Chenxu Yang","Qingyi Si","Chuanyu Qin","Dingyu Yao","Peng Fu","Zheng Lin","Weiping Wang","Nan Duan","Jiaqi Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Co-Evolving Policy Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.27083","noteUrl":"/review?paper=review-neurips-2026-152206","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152207","title":"When does LeJEPA learn a World Model?","authors":["David Klindt","Yann LeCun","Randall Balestriero"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Stationary Gaussian marginals do not identify an OU transition","plainExplanation":"The world-model assumptions permit arbitrary deterministic mean maps with independent additive noise, then infer a scalar Ornstein-Uhlenbeck conditional kernel. A radius-dependent sign permutation \\(s\\) preserves the standard Gaussian law.","primaryUrl":"https://arxiv.org/html/2605.26379","noteUrl":"/review?paper=review-neurips-2026-152207","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152208","title":"Overcoming State Inertia in Full-Duplex Spoken Language Models via Activation Steering","authors":["Cheng-Kuang Chang","Kai-Wei Chang","Alexander Liu","Jim Glass"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Overcoming State Inertia in Full-Duplex Spoken Language Models via Activation Steering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11386","noteUrl":"/review?paper=review-neurips-2026-152208","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152210","title":"Emergent Semantic Role Understanding in Language Models","authors":["Carla Griffiths","Mirco Musolesi"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Emergent Semantic Role Understanding in Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09187","noteUrl":"/review?paper=review-neurips-2026-152210","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152215","title":"HiFloat4 Format for Language Model Pre-training on Ascend NPUs","authors":["Mehran Taghian Jazi","Yunke Peng","Xing Huang","Yao Wang","Yaoyuan Wang","Wei Guo","Yuanyong Luo","Tianchi Hu","JUNSONG WANG","Xin Wang","Hu Liu","Yu Cheng","Yu Z Wei","Hongliang Li","Mehdi Rahimifar","Lei YAN","wxuefei","Zhuang Ma","Liulei","Hui Yu","Anandharaju D Raju","Hoang Le","Hei Yi Mak","Tanzila Rahman","Shadan Golestan"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Fixed-width dithering is not unbiased rounding on a nonuniform floating grid","plainExplanation":"The quantization appendix claims that adding uniform noise on \\([-1/2,1/2]\\) and rounding to the nearest FP4 value is unbiased. E2M1 has adjacent values four and six.","primaryUrl":"https://arxiv.org/html/2604.08826","noteUrl":"/review?paper=review-neurips-2026-152215","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152216","title":"Oversmoothing as Representation Degeneracy in Neural Sheaf Diffusion","authors":["Arif Dönmez","Ellen Fritsche","Axel Mosig","Katharina Koch"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Smaller edge stalks cannot produce the claimed nonempty stability chamber","plainExplanation":"The proposed smaller-edge-stalk architecture cannot exclude the trivial subrepresentation in a nonempty semistable locus under the paper’s quiver orientation. Edge-only subrepresentations force weight signs that make the claimed exclusion inequality impossible when $d_e<d_v$.","primaryUrl":"https://arxiv.org/html/2605.11178","noteUrl":"/review?paper=review-neurips-2026-152216","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-152218","title":"Improved State Mixing in Higher-order and Block Diagonal Linear Recurrent Networks","authors":["Igor Dubinin","Antonio Orvieto","Felix Effenberger"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The auxiliary expansions and eigenvalue estimate need correction","plainExplanation":"An eigenvalue is bounded by the maximum row sum, not by an arbitrary row’s budget. The companion shift rows have row sum one.","primaryUrl":"https://arxiv.org/html/2602.12021","noteUrl":"/review?paper=review-neurips-2026-152218","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152219","title":"VLM$^3$: Vision Language Models Are Native 3D Learners","authors":["zhipeng cai","Zhuang Liu","Yunyang Xiong","Zechun Liu","Vikas Chandra","Yangyang Shi"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"VLM$^3$: Vision Language Models Are Native 3D Learners","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30561","noteUrl":"/review?paper=review-neurips-2026-152219","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152222","title":"The Alien Space of Science: Sampling Coherent but Cognitively Unavailable Research Directions","authors":["Alejandro H. Artiles","Martin Weiss","Levin Brinkmann","Iyad Rahwan","Bernhard Schölkopf","Chris Pal","Hugo Larochelle","Anirudh Goyal","Nasim Rahaman"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Alien Space of Science: Sampling Coherent but Cognitively Unavailable Research Directions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.01092","noteUrl":"/review?paper=review-neurips-2026-152222","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152223","title":"Training on Documents About Monitoring Leads to CoT Obfuscation","authors":["Reilly Haskins","Bilal Chughtai","Joshua Engels"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Training on Documents About Monitoring Leads to CoT Obfuscation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15257","noteUrl":"/review?paper=review-neurips-2026-152223","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152224","title":"OneSearch-V2: The Latent Reasoning Enhanced Self-distillation Generative Search Framework","authors":["Ben Chen","Siyuan Wang","Yufei Ma","Zihan Liang","Han Li","Chenyi Lei","Wenwu Ou","Kun Gai"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The prefix reward normalization does not bound gates by one","plainExplanation":"The position reward is \\(R_l=l\\Delta R_l\\), its early positional weights are two, and gates divide prefix reward by prefix length while claiming values in \\([0,1]\\). For a perfectly matching target, the first two relevant prefix rewards produce gates of two at positions two and three.","primaryUrl":"https://arxiv.org/html/2603.24422","noteUrl":"/review?paper=review-neurips-2026-152224","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152225","title":"Localize Any Object in X-Ray Security Scans without Human Annotation","authors":["Yaqi Cai","Mingxuan Liu","Lorenzo Vaquero","Ning Wang","Nan Pu","Feng Xue","Elisa Ricci","Nicu Sebe"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Localize Any Object in X-Ray Security Scans without Human Annotation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.07326","noteUrl":"/review?paper=review-neurips-2026-152225","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152226","title":"Domination-Avoiding Learning Agents Cannot Collude","authors":["Noam Nisan","Emmanuel Zerah"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Domination-Avoiding Learning Agents Cannot Collude","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.01275","noteUrl":"/review?paper=review-neurips-2026-152226","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152228","title":"Spectral Reversal: Counteracting Singular Value Bias for Graph Prompting","authors":["Hanxu Yang","Yuhuan Zhao","Xiaodong He","Zhao Kang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Spectral Reversal: Counteracting Singular Value Bias for Graph Prompting","plainExplanation":"The calculation does not supply the asserted KL budget, and a Gibbs-risk bound does not directly bound the deterministic posterior mean. The norm estimate also needs bounded auxiliary coefficients.","primaryUrl":"https://arxiv.org/html/2609.32143","noteUrl":"/review?paper=review-neurips-2026-152228","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152229","title":"Differentiable Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation","authors":["Andrea Giuseppe Di Francesco","Andrea Rubbi","Rishabh Jain","Pietro Lió"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Differentiable Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation","plainExplanation":"","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2603.07233","noteUrl":"/review?paper=review-neurips-2026-152229","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152232","title":"How Hard is it to Rig a Benchmark? A Social Choice Analysis of Leaderboard Robustness","authors":["Polina Gordienko","Georg Schollmeyer","Frauke Kreuter","Christoph Jansen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"How Hard is it to Rig a Benchmark? A Social Choice Analysis of Leaderboard Robustness","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23628","noteUrl":"/review?paper=review-neurips-2026-152232","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152233","title":"To discretize continually: Mean shift interacting particle systems for Bayesian inference","authors":["Ayoub Belhadji","Daniel Sharp","Youssef Marzouk"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The regularizer factor changes the particle-weight system","plainExplanation":"The main objective adds \\(\\lambda\\|w\\|^2\\) to half squared MMD, while Equation (14) uses weight system \\((K+\\lambda I)w=v_0\\). Differentiation gives \\((K+2\\lambda I)w=v_0\\).","primaryUrl":"https://arxiv.org/html/2605.14142","noteUrl":"/review?paper=review-neurips-2026-152233","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152234","title":"BLADE: Scalable Bi-level Adaptive Data Selection for LLM Training","authors":["Jiaxing Wang","Deping Xiang","Jin Xu","Zirui Liu","Zicheng Zhang","Guoqiang Gong","Jun Fang","Chao Liu","Pengzhang Liu","Tongxuan Liu","Ke Zhang","Qixia Jiang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The convergence theorem must identify the relaxed update it analyzes","plainExplanation":"The proof’s persistent convex-combination update is absent from the printed practical algorithm. The source calls batch selection an approximation, but does not bound the oracle error needed to transfer this convergence statement.","primaryUrl":"https://arxiv.org/html/2606.18650","noteUrl":"/review?paper=review-neurips-2026-152234","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152236","title":"Playing Markov Games Without Observing Payoffs","authors":["Daniel Ablin","Alon Peled-Cohen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Global swap symmetry does not remove diagonal-history continuation values","plainExplanation":"The HSG symmetry condition allows nonzero skew continuation games after diagonal histories, invalidating the first-stage-only reduction. The lower-bound proof also selects payoff signs after random play, replacing the stated fixed-game objective with a stronger hindsight objective.","primaryUrl":"https://arxiv.org/html/2509.00179","noteUrl":"/review?paper=review-neurips-2026-152236","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152237","title":"Fast and slow gradient descent dynamics of logistic regression through weak alignment","authors":["Han Bao"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Fast and slow gradient descent dynamics of logistic regression through weak alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.04382v1","noteUrl":"/review?paper=review-neurips-2026-152237","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152238","title":"Noisy Data is Destructive to Reinforcement Learning with Verifiable Rewards","authors":["Yuxuan Zhu","Daniel Kang"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Noisy Data is Destructive to Reinforcement Learning with Verifiable Rewards","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.16140","noteUrl":"/review?paper=review-neurips-2026-152238","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152240","title":"Kernel Token Contradiction: a Fast and Principled Approach for LLM Claim Uncertainty Quantification","authors":["Jérémie Dentan","Alexi Canesse","Mahammed El Sharkawy","Sonia Vanier"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Kernel Token Contradiction: a Fast and Principled Approach for LLM Claim Uncertainty Quantification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.22506","noteUrl":"/review?paper=review-neurips-2026-152240","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152244","title":"Leveraging Latent Visual Reasoning in Silence","authors":["Dongyao Zhu","Zhen Wang","Xi Xiao","Han Jiang","Saeed Vahidian","Wei-Lun (Harry) Chao","Tanya Berger-Wolf","Yu Su","Ranga Raju Vatsavai","Jianyang Gu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Leveraging Latent Visual Reasoning in Silence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18641","noteUrl":"/review?paper=review-neurips-2026-152244","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152248","title":"Beyond DSA: Conjugacy-based Comparison of Dynamical Systems","authors":["Prakhar Godara","Pang S Tay","Marcelo G Mattar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Duplicated constants make the dictionary Gram inverse singular","plainExplanation":"The conjugacy dictionary concatenates endpoint blocks containing the same constant and uses an ordinary inverse Gram square root and inverse principal-angle sine. A nonzero coefficient vector subtracting the duplicate constants satisfies \\(Az=Gz=0\\).","primaryUrl":"https://arxiv.org/html/2607.04493","noteUrl":"/review?paper=review-neurips-2026-152248","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152249","title":"G-Zero: Self-Play for Open-Ended Generation from Zero Data","authors":["Chengsong Huang","Haolin Liu","Tong Zheng","Runpeng(Leo) Dai","Langlin Huang","JINYUAN LI","Zongxia Li","Zhepei Wei","Yu Meng","Jiaxin Huang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A large Hint-delta can reward a misleading hint on an easy question","plainExplanation":"A large Hint-delta does not require a difficult question or a useful hint. A short misleading hint can sharply reduce the probability of an already confident correct answer and produce a large score, so the statistic alone does not identify helpful missing reasoning.","primaryUrl":"https://arxiv.org/html/2605.09959","noteUrl":"/review?paper=review-neurips-2026-152249","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152252","title":"Live Music Diffusion Models: Efficient Fine-Tuning and Post-Training of Interactive Diffusion Music Generators","authors":["Zachary Novack","Stephen Brade","Haven Kim","Hugo Flores","Nithya Shikarpur","Chinmay Talegaonkar","Julian McAuley","Taylor Berg-Kirkpatrick","Anna Huang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Live Music Diffusion Models: Efficient Fine-Tuning and Post-Training of Interactive Diffusion Music Generators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22717","noteUrl":"/review?paper=review-neurips-2026-152252","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152260","title":"From Weak Data to Strong Policy: Q-Targets Enable Provable In-Context Reinforcement Learning","authors":["yichen lin","Xuyuan Xiong","xue wang","Xiangfu Meng","Mike Wei","Tao Yao"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Summed positive increments do not compute a running maximum","plainExplanation":"The construction expresses a maximum using the initial value plus positive adjacent increments, with arbitrary action ordering. Drops followed by recovery are counted again.","primaryUrl":"https://arxiv.org/html/2609.30391","noteUrl":"/review?paper=review-neurips-2026-152260","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152264","title":"Precise but Uncoupled: Reviewer Precision Does Not Guarantee Critique Uptake in Multi-Agent Math Reasoning","authors":["Chih-Hsuan Yang","Jingyan Jiang","Vikram Vasudevan","Cheng-Hau Yang","Huihuo Zheng","Benoit Cote","Le Chen","Eliu Huerta","Venkatram Vishwanath","Ian Foster","Rajeev Thakur"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Precise but Uncoupled: Reviewer Precision Does Not Guarantee Critique Uptake in Multi-Agent Math Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.15388","noteUrl":"/review?paper=review-neurips-2026-152264","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152266","title":"Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs","authors":["Wei Ding","Yudong Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.03879","noteUrl":"/review?paper=review-neurips-2026-152266","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152267","title":"Feedback World Model Enables Precise Guidance of Diffusion Policy","authors":["Tuo An","Jindou Jia","Gen Li","Jingliang Li","Chuhao Zhou","Pengfei Liu","Bofan Lyu","Jiaqi Bai","Xinying Guo","Geng Li","Jianfei Yang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Feedback World Model Enables Precise Guidance of Diffusion Policy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15705","noteUrl":"/review?paper=review-neurips-2026-152267","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152268","title":"The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network","authors":["Elias Sandmann","Sebastian Lapuschkin","Wojciech Samek"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2508.21380","noteUrl":"/review?paper=review-neurips-2026-152268","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152271","title":"Structural Rationale Distillation via Reasoning Space Compression","authors":["Jialin Yang","Jiankun Wang","Jiajun Wu","Henry Leung","Jiayu Zhou","Steve Drew"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A restricted global template bank need not beat the unrestricted conditional target","plainExplanation":"The generalization proof substitutes an entropy floor for the risk of a Gaussian posterior without requiring that posterior to attain the floor. Separately, minimizing loss over a restricted global template bank does not guarantee lower loss than an unrestricted input-dependent target; a two-input singleton-bank example reverses the comparison.","primaryUrl":"https://arxiv.org/html/2605.07139","noteUrl":"/review?paper=review-neurips-2026-152271","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152272","title":"On the Divergence of Differential Temporal Difference Learning without Local Clocks","authors":["David Antrobius","Shangtong Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Local clocks leave frequency-dependent ODE coefficients","plainExplanation":"The asynchronous ODE argument uses power stepsizes \\(\\alpha_t\\asymp t^{-\\beta}\\) and claims sampling-frequency-independent drift. A coordinate sampled with frequency \\(d_s\\) contributes multiplier \\(d_s\\alpha_{d_st}/\\alpha_t\\to d_s^{1-\\beta}\\).","primaryUrl":"https://arxiv.org/html/2605.06874","noteUrl":"/review?paper=review-neurips-2026-152272","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152273","title":"LiSA: Lifelong Safety Adaptation via Conservative Policy Induction","authors":["Minbeom Kim","Lesly Miculicich","Bhavana Dalvi Mishra","Mihir Parmar","Phillip Wallis","Bharath Chandrasekhar","Kyomin Jung","Tomas Pfister","Long T. Le"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LiSA: Lifelong Safety Adaptation via Conservative Policy Induction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14454","noteUrl":"/review?paper=review-neurips-2026-152273","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152275","title":"VC-OPD: Visual Counterfactual On-Policy Distillation for Grounded Vision-Language Reasoning","authors":["Jiahe Feng"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"VC-OPD: Visual Counterfactual On-Policy Distillation for Grounded Vision-Language Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.14939","noteUrl":"/review?paper=review-neurips-2026-152275","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152277","title":"VolCo: Volumetric Contact for High-Fidelity Human Grasp Generation","authors":["Zhuo Chen","Yihua Cheng","Ales Leonardis","Hyung Jin Chang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The contact conversion uses nonexistent ordinary inverses","plainExplanation":"The CSE vectors have four coordinates, while the conversion inverts a three-column embedding matrix and an origin-based three-point positional matrix. The embedding matrix is rectangular and has no ordinary inverse.","primaryUrl":"https://arxiv.org/html/2610.10197","noteUrl":"/review?paper=review-neurips-2026-152277","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152279","title":"OGPO: Offline Goal-conditioned Policy Optimization for Recoverable Vision-Language-Action Models","authors":["Xule Gao","Xi Wang","Zehua Zang","Rui Wang","Changwen Zheng","Chuxiong Sun"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"OGPO: Offline Goal-conditioned Policy Optimization for Recoverable Vision-Language-Action Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.12914","noteUrl":"/review?paper=review-neurips-2026-152279","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152284","title":"Understanding Layer Patching in Model Size Interpolation","authors":["Sara Kangaslahti","Jonathan Geuter","Nihal V Nayak","Marco Fumero","Francesco Locatello","David Alvarez-Melis"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Understanding Layer Patching in Model Size Interpolation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.08170","noteUrl":"/review?paper=review-neurips-2026-152284","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152286","title":"Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning","authors":["Baris Askin","Holger Roth","Zhenyu Sun","Carlee Joe-Wong","Gauri Joshi","Ziyue Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.00655","noteUrl":"/review?paper=review-neurips-2026-152286","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152288","title":"Soft Contamination Means Benchmarks Test Shallow Generalization","authors":["Ari Spiesberger","Juan J Vazquez","Nicky Pochinkov","Tomáš Gavenčiak","Peli Grietzer","Gavin Leech","Nandi Schoots"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Soft Contamination Means Benchmarks Test Shallow Generalization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.12413","noteUrl":"/review?paper=review-neurips-2026-152288","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152289","title":"Learning Reveals Invisible Structure in Low-Rank RNNs","authors":["Yoav Ger","Omri Barak"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The revised parameter gradients leave a scalar-flow normalization mismatch","plainExplanation":"The product rule for an overlap introduces another division by width; converting raw inner products to normalized norms cancels only one factor.","primaryUrl":"https://arxiv.org/html/2605.04115","noteUrl":"/review?paper=review-neurips-2026-152289","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152292","title":"Hitting Time Isomorphism for Multi-Stage Planning with Foundation Policies","authors":["Magnus Victor Boock","Abdullah Akgül","Mustafa Mert Çelikok","Melih Kandemir"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Hitting Time Isomorphism for Multi-Stage Planning with Foundation Policies","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06470","noteUrl":"/review?paper=review-neurips-2026-152292","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152295","title":"Neural Quantum Spectral Operator Learning for Solving Partial Differential Equations","authors":["Chanyoung Kim","Myeonghwan Seong","Kim Yujin","Daniel Kyungdeock Park","Youngjoon Hong"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The truncated-cost conclusion quotes the quantum term as classical cost","plainExplanation":"The proof’s final sentence assigns the quantum cost expression to the dominating classical work. The stated classical term has different dependence on problem size, matching the proposition at fixed accuracy. This is a mismatch in the proof’s final complexity sentence. The proposition matches the fixed-accuracy classical reading; uniform joint scaling requires both terms.","primaryUrl":"https://arxiv.org/html/2605.27408","noteUrl":"/review?paper=review-neurips-2026-152295","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152301","title":"FluxLite: Inference-Time Proposal Control for Discrete Diffusion Models","authors":["Yinuo Ren","Haoxuan Chen","Grant Rotskoff","Jiequn Han","Lexing Ying"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"FluxLite: Inference-Time Proposal Control for Discrete Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35947","noteUrl":"/review?paper=review-neurips-2026-152301","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152302","title":"AdERA: Adaptive Exponent Reuse for Lossless Allgather in Sharded MoE Training","authors":["Ali Zafar Sadiq","Haiying Shen","Masahiro Tanaka"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AdERA: Adaptive Exponent Reuse for Lossless Allgather in Sharded MoE Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.04609","noteUrl":"/review?paper=review-neurips-2026-152302","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152303","title":"Class Adaptive Conformal Training","authors":["Badr-Eddine Marani","Julio Silva-Rodríguez","Ismail Ayed","Maria Vakalopoulou","Stergios Christodoulidis","Jose Dolz"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The printed loss update ascends its gradient","plainExplanation":"The augmented objective is written as joint minimization over primal variables and a nonnegative multiplier, using the PHR active branch. At feasible residual \\(z=-1\\) and penalty parameter one, the active expression is \\(-\\lambda+1/2\\) for \\(\\lambda\\ge1\\), tending to minus infinity under minimization.","primaryUrl":"https://arxiv.org/html/2601.09522","noteUrl":"/review?paper=review-neurips-2026-152303","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","implementation_mismatch","interpretation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152304","title":"TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection","authors":["Tom Sander","Pierre Fernandez","Hongyan Chang","Tomáš Souček","Sylvestre-Alvise Rebuffi","Hady Elsahar","Valeriu Lacatusu","Alexandre Mourachko","Tuan Tran"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Moment-matched Gamma tails do not establish strict family-wise error control","plainExplanation":"The detector uses Gamma tails matched to two null moments and invokes a union bound requiring superuniform p-values for a universal finite-sample false-positive guarantee. Under ideal independent dual-key scores, each token contributes \\(0.9E_1+0.1E_2\\) for unit exponentials.","primaryUrl":"https://arxiv.org/html/2605.12456","noteUrl":"/review?paper=review-neurips-2026-152304","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152305","title":"SplineFlow: Flow Matching for Dynamical Systems with B-Spline Interpolants","authors":["Santanu Rathod","Pietro Lió","Xiao Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"SplineFlow: Flow Matching for Dynamical Systems with B-Spline Interpolants","plainExplanation":"An affine probability-flow field changes the variance at twice its radial coefficient. The printed field therefore doubles the prescribed variance derivative.","primaryUrl":"https://arxiv.org/html/2601.23072v2","noteUrl":"/review?paper=review-neurips-2026-152305","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152307","title":"The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment","authors":["Rishab Balasubramanian","Pin-Jie Lin","Rituraj Sharma","Mohit Bansal","Tu Vu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.06377","noteUrl":"/review?paper=review-neurips-2026-152307","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152309","title":"Scaling Limits of Long-Context Transformers","authors":["Giuseppe Bruno","Chen","Zhengjiang Lin","Yury Polyanskiy","Philippe Rigollet"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A strict smooth maximum need not have an inward gradient nearby","plainExplanation":"The global density comparison fails near the antipode, although its local form can remain valid. The Appendix B beta-integral estimate also diverges at the included dimension $d=2$. Separately, strict $C^2$ maximality does not ensure the gradient sign needed by the similarity-drift corollary.","primaryUrl":"https://arxiv.org/html/2605.08505","noteUrl":"/review?paper=review-neurips-2026-152309","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152311","title":"Verifying Neural Networks with Reinforcement Learning","authors":["Hai Duong","Thanh Le","ThanhVu Nguyen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The reward and verification pseudocode need consistent contracts","plainExplanation":"The reward counts resolved leaves, so a larger complete search tree can earn more reward. Discounting does not generally reverse that ordering.","primaryUrl":"https://arxiv.org/html/2609.34553","noteUrl":"/review?paper=review-neurips-2026-152311","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152314","title":"Block Sphere Vector Quantization","authors":["Heesang Ann","Joongkyu Lee","Min-hwan Oh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Total squared distortion needs the dimension factor","plainExplanation":"Definition C.1 uses normalized second moment \\(G\\), while the distortion is the full squared Euclidean norm rather than per-coordinate error. A cell of volume \\(v\\) has integrated error \\(mGv^{1+2/m}\\).","primaryUrl":"https://arxiv.org/html/2605.19972","noteUrl":"/review?paper=review-neurips-2026-152314","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152316","title":"CircuitSeer: Mining High-Quality Data by Probing Mathematical Reasoning Circuits in LLMs","authors":["Shaobo Wang","Yongliang Miao","YuanCheng Liu","Qianli Ma","Ning Liao","Dongrui Liu","Linfeng Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Row-normalized attention makes the selection variance zero","plainExplanation":"The selection score averages \\(\\sum_jA^h_{kj}\\) over reasoning heads, and the standing normalization sets every such row sum to one. Thus every token score \\(\\alpha_k=1\\), and its variance is zero for every sample.","primaryUrl":"https://arxiv.org/html/2510.18470","noteUrl":"/review?paper=review-neurips-2026-152316","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152317","title":"StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction","authors":["Xiangyuan Xue","Yifan Zhou","ZiDong Wang","Shengji Tang","Philip Torr","Wanli Ouyang","LEI BAI","Zhenfei Yin"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06642","noteUrl":"/review?paper=review-neurips-2026-152317","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152320","title":"Emergence of Distortions in High-Dimensional Guided Diffusion Models","authors":["Enrico Ventura","Beatrice Achilli","Luca Ambrogioni","Carlo Lucibello"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The zero-guidance variance baseline has the wrong time sign","plainExplanation":"The time-dependent CFG distortion compares a propagated variance with a baseline written \\(\\sigma^2-t\\). At zero guidance, the variance inside Equation (137) simplifies to \\(\\sigma^2+t\\).","primaryUrl":"https://arxiv.org/html/2602.00716","noteUrl":"/review?paper=review-neurips-2026-152320","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152321","title":"Invisible Ink, Visible Lies: How Production Watermarking Causes LLMs to Hallucinate","authors":["Haocheng Ye","Aoting Hu","Xinwei Zhang","Xunzhu Tang","Shuchao Pang","Minhui Xue"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Invisible Ink, Visible Lies: How Production Watermarking Causes LLMs to Hallucinate","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04860","noteUrl":"/review?paper=review-neurips-2026-152321","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152322","title":"Generating the Wild: Individual-Consistent Image-to-Video Generation for Wildlife","authors":["Yuzhuo Li","Di Zhao","Xinyu Zhang","Daniel Wilson","Yun Sing Koh"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Generating the Wild: Individual-Consistent Image-to-Video Generation for Wildlife","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.05587","noteUrl":"/review?paper=review-neurips-2026-152322","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152324","title":"Factual recall in linear associative memories: sharp asymptotics and mechanistic insights","authors":["Alessio Giorlandino","Sebastian Goldt","Antoine Maillard"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A Gaussian-maximum upper-tail rate needs the positive part of the threshold","plainExplanation":"The limiting spectral measure loses probability mass by weighting a truncated density twice. Two auxiliary replica formulas also contain a quadratic-coefficient half-factor and a reciprocal-derivative typo. The Gaussian-maximum lemma separately assigns an incorrect rate to sufficiently negative thresholds, although the following application treats that region correctly.","primaryUrl":"https://arxiv.org/html/2605.10795","noteUrl":"/review?paper=review-neurips-2026-152324","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152325","title":"Truncate Bad, Upweight Good: BoN-Style Distillation via Rank-Based Classification","authors":["Yarin Bar","Yaniv Romano"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Truncate Bad, Upweight Good: BoN-Style Distillation via Rank-Based Classification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.19748","noteUrl":"/review?paper=review-neurips-2026-152325","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152326","title":"DecQ: Detail-Condensing Queries for Enhanced Reconstruction and Generation in Representation Autoencoders","authors":["Tianhang Wang","Yitong Chen","Wei Song","Zuxuan Wu","Min Li","Jiaqi Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DecQ: Detail-Condensing Queries for Enhanced Reconstruction and Generation in Representation Autoencoders","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22777","noteUrl":"/review?paper=review-neurips-2026-152326","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152327","title":"Diagonalizing the Softmax: Hadamard Initialization for Tractable Cross-Entropy Dynamics","authors":["Connall Garrod","Jonathan Keating","Christos Thrampoulidis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Hadamard softmax identity omits scale factors","plainExplanation":"The normalized Hadamard basis is \\(U=\\Phi/\\sqrt K\\), and the displayed identity reconstructs column softmax from it. At zero input, column softmax is \\(\\mathbf1\\mathbf1^\\top/K\\), while the printed expression gives \\(\\mathbf1\\mathbf1^\\top/(K\\sqrt K)\\).","primaryUrl":"https://arxiv.org/html/2512.04006","noteUrl":"/review?paper=review-neurips-2026-152327","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152329","title":"Simulation-Ready Compositional 3D Scene Reconstruction from a Single Image","authors":["Inhee Lee","Sangwon Baik","Hyeonwoo Kim","Hyunsoo Cha","Sungjoo Kim","Hanbyul Joo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Simulation-Ready Compositional 3D Scene Reconstruction from a Single Image","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03994","noteUrl":"/review?paper=review-neurips-2026-152329","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152331","title":"Enhancing Novel View Synthesis via Geometry Grounded Set Diffusion","authors":["Farhad G. Zanjani","Herbert Cai","Amirhossein Habibian"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Enhancing Novel View Synthesis via Geometry Grounded Set Diffusion","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.07540","noteUrl":"/review?paper=review-neurips-2026-152331","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152333","title":"APIVOT: Adaptive Planning with Interleaved Vision-Language Thoughts","authors":["Emily Jin","Joy Hsu","Yiqing Xu","Weiyu Liu","Nick Haber","Jiajun Wu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"APIVOT: Adaptive Planning with Interleaved Vision-Language Thoughts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.08024","noteUrl":"/review?paper=review-neurips-2026-152333","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152334","title":"Fine-Tuning Improves Information Conveyance in Language Models","authors":["Yuwei Cheng","Weiyi Tian","Haifeng Xu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Prompt entropy variance is not controlled by prompt length variance","plainExplanation":"The refined error bound incorrectly controls prompt-to-prompt entropy variation by length variation; equal-length prompts can have different entropy. The BF estimator also converges to a pathwise local-entropy ratio that differs from the paper’s length-conditioned content-entropy definition.","primaryUrl":"https://arxiv.org/html/2605.30844","noteUrl":"/review?paper=review-neurips-2026-152334","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152337","title":"ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs","authors":["Xinlin Li","Timothy Chou","Joshua W Fromm","Zichang Liu","Yunjie Pan","Christina Fragouli"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A fixed-size upgrade can overshoot the bit budget","plainExplanation":"The feasible mean precision is at most \\(B\\), but the greedy rule adds \\(k\\) bits whenever the current mean is below the limit. The remaining slack can be smaller than the upgrade.","primaryUrl":"https://arxiv.org/html/2602.17698","noteUrl":"/review?paper=review-neurips-2026-152337","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152339","title":"Guided Data Generation for Understanding Model Behavior","authors":["Eren Mehmet KIRAL","Nursen Aydin","Ilker Birbil"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The Hastings proposal density must match the sampled Euler step","plainExplanation":"The sampler proposes an Euler Langevin step with drift scale \\(\\eta\\) and noise variance \\(2\\eta\\tau\\), but its displayed proposal density uses different scales. Up to a cancelling Gaussian constant, its log density is \\(-\\|y-x+\\eta\\nabla G(x)\\|^2/(4\\eta\\tau)\\).","primaryUrl":"https://arxiv.org/html/2502.06658","noteUrl":"/review?paper=review-neurips-2026-152339","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152341","title":"Learning to Inject: Automated Prompt Injection via Reinforcement Learning","authors":["Xin Chen, Cynthia","Jie Zhang","Florian Tramer"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Learning to Inject: Automated Prompt Injection via Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.05746","noteUrl":"/review?paper=review-neurips-2026-152341","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152343","title":"SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning","authors":["Ruijia Liu","Ancheng Hou","Xiang Yin"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.00775","noteUrl":"/review?paper=review-neurips-2026-152343","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152345","title":"LENS: Low-Frequency Eigen Noise Shaping for Efficient Diffusion Sampling","authors":["Haewon Jeon","Si-Hyeon Lee"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"LENS: Low-Frequency Eigen Noise Shaping for Efficient Diffusion Sampling","plainExplanation":"The appendix gives an explicit Jacobian approximation bound, but the proposition still omits that independent term from its claimed reward-tolerance guarantee.","primaryUrl":"https://arxiv.org/html/2605.07253","noteUrl":"/review?paper=review-neurips-2026-152345","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152346","title":"Even More Guarantees for Variational Inference in the Presence of Symmetries","authors":["Lena Zellinger","Antonio Vergari"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Even More Guarantees for Variational Inference in the Presence of Symmetries","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.21407","noteUrl":"/review?paper=review-neurips-2026-152346","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152348","title":"Remember with Confidence: Uncertainty Quantification for Spatio-temporal Memory with Probabilistic Guarantees","authors":["Harry Zhang","Nicolas Gorlo","Luca Carlone"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Remember with Confidence: Uncertainty Quantification for Spatio-temporal Memory with Probabilistic Guarantees","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.08277","noteUrl":"/review?paper=review-neurips-2026-152348","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152349","title":"Reverse to Advance: Teleoperation-Cost Effective Hard Policy Learning from Reversed Easy Tasks","authors":["Qiyuan Qiao","Ge Yuan","Can Wang","Dong Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Reverse to Advance: Teleoperation-Cost Effective Hard Policy Learning from Reversed Easy Tasks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.13455","noteUrl":"/review?paper=review-neurips-2026-152349","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152351","title":"Prospective Hindsight: Self-Calibrating Reinforcement Learning via Prediction–Reality Gaps","authors":["Jiaxin Zhang","XIANGYU PENG","Qinglin Chen","Yu Li","Hiroaki Hayashi","Chien-Sheng Wu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Action-derived beliefs need not be conditionally independent of gradients","plainExplanation":"The conditional-gradient identity conditions on belief \\(z\\) derived from state/action, then treats exclusion of an explicit belief input from the loss as sufficient independence. Action conditioning changes the gradient distribution.","primaryUrl":"https://arxiv.org/html/2610.02740","noteUrl":"/review?paper=review-neurips-2026-152351","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152354","title":"Generative Modeling by Value-Driven Transport","authors":["Pablo Moreno-Muñoz","Adrian Müller","Gergely Neu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The optimal boundary interpolation is shifted by one state","plainExplanation":"Theorem 3.1 assigns physical state \\(k\\) fraction \\(k/(H+1)\\), but Equation (9) assigns the shared boundary of stages \\(h\\) and \\(h+1\\) fraction indexed by \\(h\\). That boundary is state \\(h+1\\).","primaryUrl":"https://arxiv.org/html/2605.22507","noteUrl":"/review?paper=review-neurips-2026-152354","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152356","title":"ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving","authors":["Huimin Wang","Yue Wang","Bihao Cui","Pengxiang Li","Ben Lu","Mingqian Wang","Tong Wang","Chuan Tang","Teng Zhang","Kun Zhan"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.04647","noteUrl":"/review?paper=review-neurips-2026-152356","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152357","title":"Learning from Language Feedback via Variational Policy Distillation","authors":["Yang Li","Erik Nijkamp","Semih Yavuz","Shafiq Joty"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Exact dynamic-reference substeps can decrease the fixed-reference objective","plainExplanation":"The alternating dynamic-reference E/M subproblems can both be solved exactly while decreasing the fixed-reference RLVR objective. A Bernoulli bandit starts at that objective’s optimum and is moved away from it, contradicting Proposition 1’s monotonicity inference.","primaryUrl":"https://arxiv.org/html/2605.15113","noteUrl":"/review?paper=review-neurips-2026-152357","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-152360","title":"On the Faithfulness of Visual Thinking: Measurement and Enhancement","authors":["Zujing Liu","Junwen Pan","Qi She","Yuan Gao","Gui-Song Xia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Accuracy invariance does not identify answer invariance","plainExplanation":"The intervention replaces answer \\(A\\) by its correctness indicator and interprets zero average accuracy effect as reasoning based only on text. The indicator loses answer identity.","primaryUrl":"https://arxiv.org/html/2510.23482","noteUrl":"/review?paper=review-neurips-2026-152360","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152361","title":"Invariant Features in Language Models: Geometric Characterization and Model Attribution","authors":["Agnibh Dasgupta","Abdullah All Tanvir","Xin Zhong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The discovery covariance supports fewer directions than the fixed projection rank","plainExplanation":"The per-group discovery matrix uses ten centered semantic-change rows and a positive-definite denominator covariance, then retains 32 generalized eigenvectors and scores top-rank eigenvalue mass. Centering gives numerator rank at most nine, so at least 23 retained directions have zero generalized eigenvalue.","primaryUrl":"https://arxiv.org/html/2605.06458","noteUrl":"/review?paper=review-neurips-2026-152361","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152362","title":"Adaptive Generate-Rank-Verify: Inference-Time Search with Costly Verification","authors":["Shaddin Dughmi","Mahdi Haghifam","Yusuf H Kalayci"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Adaptive Generate-Rank-Verify: Inference-Time Search with Costly Verification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17609","noteUrl":"/review?paper=review-neurips-2026-152362","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152364","title":"FeatCal: Feature Calibration for Post-Merging Models","authors":["Yanggan Gu","Shuo CAI","Zihao Wang","Wenjun Wang","Yuanyi Wang","Pengkai Wang","Sirui Huang","Su Lu","Jianmin Wu","Hongxia Yang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"FeatCal: Feature Calibration for Post-Merging Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13030","noteUrl":"/review?paper=review-neurips-2026-152364","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152365","title":"ECHO-2: A Large-Scale Distributed Rollout Framework for Cost-Efficient Reinforcement Learning","authors":["Jingwei Song","Meng Chen","Jie Xiao","Qingnan Ren","Jiaqi Huang","Yangshen Deng","Senyu Tong","Wanyi Chen","Suli Wang","Zhisheng Chen","Ziqian Bi","Shuo Lu","Yiqun Duan","Xu Wang","Rymon Yu","Lynn Ai","Eric Yang","TIANYU SHI"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"ECHO-2: A Large-Scale Distributed Rollout Framework for Cost-Efficient Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.02192","noteUrl":"/review?paper=review-neurips-2026-152365","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152368","title":"Why DiT Models Underperform as Representation Learners without Long Skip Connections","authors":["Benyuan Meng","Yiliang Zhang","Jin-Wen Wu","Qianqian Xu","Qingming Huang","Longtao Huang"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Why DiT Models Underperform as Representation Learners without Long Skip Connections","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.13344","noteUrl":"/review?paper=review-neurips-2026-152368","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152371","title":"When Integral Meets Decomposition: A Signal-Level Self-Supervised Feature Decompose Paradigm for Multi-Modal Image Fusion","authors":["Zeyu Wang","Jiayu Wang","Haiyu Song","Haoran Duan"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"When Integral Meets Decomposition: A Signal-Level Self-Supervised Feature Decompose Paradigm for Multi-Modal Image Fusion","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.39004","noteUrl":"/review?paper=review-neurips-2026-152371","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152373","title":"Choosing Before Acting: Comparative Value Estimation for Long-Horizon Tool-Use Agents","authors":["Yu Li","Zheng Zhang","Xin Liu","shengtian yang","Guangfeng Cai","Lei Feng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Choosing Before Acting: Comparative Value Estimation for Long-Horizon Tool-Use Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02330","noteUrl":"/review?paper=review-neurips-2026-152373","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152374","title":"Controllable and Content Based Recommendations","authors":["Firat Oncel","Jihoon Jeong","Emiliano Penaloza","Mirco Ravanelli","Laurent Charlin","Cem Subakan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Controllable and Content Based Recommendations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.20938","noteUrl":"/review?paper=review-neurips-2026-152374","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152377","title":"Bayes-Optimal BER and AUC: Estimation and Evaluation of Estimators","authors":["Ryota Ushio","Takashi Ishida","Masashi Sugiyama"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Bayes-Optimal BER and AUC: Estimation and Evaluation of Estimators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2609.02304v1","noteUrl":"/review?paper=review-neurips-2026-152377","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152378","title":"Spectral Geometry of Attention: From Information Routing to Uncertainty","authors":["Giulio Viganò","Simone Melzi","Maks Ovsjanikov"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Spectral Geometry of Attention: From Information Routing to Uncertainty","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06012","noteUrl":"/review?paper=review-neurips-2026-152378","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152379","title":"Recursively Trained Diffusion Models: Limiting Collapse Distribution and Spectral Characterization","authors":["Nail B Khelifa","Richard Turner","Ramji Venkataramanan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The collapse-mixture covariance omits variance of the component means","plainExplanation":"The Hermite reconstruction uses unnormalized polynomials without dividing by their squared norms. The collapse covariance also omits variation among component means for noncentered data. A separate schedule formula assigns fresh mixture components the full smoothing history instead of their component-specific suffix times.","primaryUrl":"https://arxiv.org/html/2606.13796","noteUrl":"/review?paper=review-neurips-2026-152379","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152381","title":"Modelling Opinion Dynamics at Scale with Deep MARL","authors":["Lukas Seier","Brandon Kaplowitz","Sebastian Towers","Richard M Bailey","Jakob Foerster"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Modelling Opinion Dynamics at Scale with Deep MARL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.07487","noteUrl":"/review?paper=review-neurips-2026-152381","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152383","title":"Logistic Bandits with $\\tilde{O}(\\sqrt{dT})$ Regret without Context Diversity Assumptions","authors":["Seoungbin Bae","Dabeen Lee"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The terminal level floor can advance into uninitialized state","plainExplanation":"Algorithm 1 uses \\(S=\\lfloor\\tfrac12\\log_2T\\rfloor\\), exploration threshold \\(2^{-S}\\), and exploitation threshold \\(1/\\sqrt T\\). At \\(T=8\\), \\(S=1\\).","primaryUrl":"https://arxiv.org/html/2604.22161","noteUrl":"/review?paper=review-neurips-2026-152383","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152384","title":"MAGNET: Manifold-Aware Graph Diffusion Network for Connectome Generation","authors":["Protyay Dey","Ayush Roy","Hyuna Cho","Won Hwa Kim","Vishnu Lokhande"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The ambient log-Euclidean center need not be a correlation matrix","plainExplanation":"The current v2 proof correctly interprets the log-Euclidean center as an ambient SPD reference, but its proposition still claims correlation-constrained optimality. The corruption paragraph also continues to identify a linear variance reduction with an exact Gaussian product, whose precisions instead add. Both findings are restricted to these assertions.","primaryUrl":"https://arxiv.org/html/2610.17014","noteUrl":"/review?paper=review-neurips-2026-152384","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152385","title":"SP-CACW: Convergence-Aware Client Weighting for Selfish Personalized Learning","authors":["Yaron Kiselman","Kfir Y. Levy"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Simplex loss maxima do not bound the full Euclidean eigenvalue","plainExplanation":"The gradient-bound proof compares the largest eigenvalue of \\(H=b^\\top b+\\eta Lg^\\top g\\) to the maximum quadratic loss over simplex weights. The simplex does not contain every unit eigenvector.","primaryUrl":"https://arxiv.org/html/2606.29322","noteUrl":"/review?paper=review-neurips-2026-152385","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152386","title":"Seeing Is Not Screening: Multimodal Hidden Instruction Attacks on Agent Skill Scanners","authors":["Xiaojun Jia","Jie Liao","Simeng Qin","Ke Ma","Wenbo Guo","Yebo Feng","Aishan Liu","Yang Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Seeing Is Not Screening: Multimodal Hidden Instruction Attacks on Agent Skill Scanners","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.18198","noteUrl":"/review?paper=review-neurips-2026-152386","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152387","title":"Memoir: Let the Model Direct Its Own Story for Robust Cross-Domain Knowledge Editing","authors":["Jea Kwon","Jiwon Kim","Dong Kyum Kim","Meeyoung Cha"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Positive semidefiniteness alone does not give a unique minimizer","plainExplanation":"The main-text uniqueness claim is too strong for positive-semidefinite output weighting. Null objective directions can leave multiple feasible minimizers, as the appendix’s stronger uniqueness hypothesis recognizes. The selected minimizer remains valid. The main-text uniqueness hypothesis is too weak, and the appendix already provides the intended stronger version.","primaryUrl":"https://arxiv.org/html/2607.20433","noteUrl":"/review?paper=review-neurips-2026-152387","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152388","title":"Lattice Deduction Transformers","authors":["Liam Davis","Alberto Alfarano","Leopold Haller","Mark Santolucito"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Lattice Deduction Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08605","noteUrl":"/review?paper=review-neurips-2026-152388","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152392","title":"Rethinking On-Policy Self-Distillation for Thinking Models","authors":["Simran Kaur","Narutatsu Ri","Yinghui He","Liam Fowl","Sanjeev Arora"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rethinking On-Policy Self-Distillation for Thinking Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.05184","noteUrl":"/review?paper=review-neurips-2026-152392","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152393","title":"Pinpoint: Grounded Worldwide Image Geolocation via Cross-Source Retrieval and Reranking","authors":["Nika Chuzhoy","Brian Hu","Amit A Arora","Jae Ro","Sarthak Sahu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Pinpoint: Grounded Worldwide Image Geolocation via Cross-Source Retrieval and Reranking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04133","noteUrl":"/review?paper=review-neurips-2026-152393","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152394","title":"Probabilistic Circuits for Irregular Multivariate Time Series Forecasting","authors":["Christian Klötergens","Lars Schmidt-Thieme","Vijaya Krishna Yalavarthi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Probabilistic Circuits for Irregular Multivariate Time Series Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.27814","noteUrl":"/review?paper=review-neurips-2026-152394","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152401","title":"Layer Free-Riding in Forward-Forward Networks: Real, Repairable, but Not Accuracy-Dominant","authors":["Amirhossein Yousefiramandi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Layer Free-Riding in Forward-Forward Networks: Real, Repairable, but Not Accuracy-Dominant","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06240","noteUrl":"/review?paper=review-neurips-2026-152401","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152402","title":"CORA: Per-Slice Coherent Orthogonal Rotation for SVD-based Low-Rank Adaptation","authors":["Pengcheng Wang","Ziran Liu","Wei Wang","WEI JIANG"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The coordinate reduction omits admissible slice perturbations","plainExplanation":"The theorem's coordinate reduction discards perturbations outside the original right subspace. Further claims need narrower assumptions: slice conditioning is not inherited as stated, real spectrum scales can break SVD coherence, and the comparisons with additive and one-sided orthogonal adapters misidentify their matrix families.","primaryUrl":"https://arxiv.org/html/2607.02576","noteUrl":"/review?paper=review-neurips-2026-152402","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","interpretation_error","missing_assumption","proof_gap","unsupported_claim"],"classifications":["major"]},{"slug":"review-neurips-2026-152405","title":"TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment","authors":["Jiaming Li","Chenyu Zhu","Zhiyuan Ma","Nanxi Yi","Youjun Bao","Li Sun","Quanying Lv","Xiang Fang","Daizong Liu","Jianjun Li","Kun He","Bowen Zhou"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Small-concentration Beta noise converges to endpoint masses","plainExplanation":"Small-concentration Beta noise converges to endpoint masses. Averaging log densities changes normalized weights and importance ratios. Matching coefficient signs does not guarantee forward-KL descent. The marginal correction counts the step size twice.","primaryUrl":"https://arxiv.org/html/2605.10983","noteUrl":"/review?paper=review-neurips-2026-152405","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152406","title":"Multi-Variable Conformal Prediction: Optimizing Prediction Sets without Data Splitting","authors":["Laura Lützow","Simone Garatti","Marco Campi","Lars Lindemann","Matthias Althoff"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Multi-Variable Conformal Prediction: Optimizing Prediction Sets without Data Splitting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12341","noteUrl":"/review?paper=review-neurips-2026-152406","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152407","title":"Depth Exploration for LLM Decoding","authors":["Weisi Yang","Zipeng Sun","Stephen Xia"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Depth Exploration for LLM Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.29223","noteUrl":"/review?paper=review-neurips-2026-152407","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152409","title":"Boosting Text-to-Image Diffusion Models via Core Token Attention-Based Seed Selection","authors":["Yunzhe Zhang","Hongfu Liu","Pengyu Hong"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Boosting Text-to-Image Diffusion Models via Core Token Attention-Based Seed Selection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19532","noteUrl":"/review?paper=review-neurips-2026-152409","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152410","title":"Plausibility Is Not Prediction: Contrastive Evidence for LLM-Based Cellular Perturbation Reasoning","authors":["XINYU YUAN","Xixian Liu","Jianan Zhao","Ya Shi Zhang","Hongyu Guo","Jian Tang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Plausibility Is Not Prediction: Contrastive Evidence for LLM-Based Cellular Perturbation Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.01042","noteUrl":"/review?paper=review-neurips-2026-152410","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152416","title":"Align-RAG: Alignment Is All You Need for TSFM In-Context Learning","authors":["Mohammad Asadi","Soheil Hor","Bardiya Akhbari","Jack W O&amp;#x27;Sullivan","Tahoura Nedaee","Layne Price","Raviteja Anantha Ramesh","Euan Ashley","Ehsan Adeli"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Align-RAG: Alignment Is All You Need for TSFM In-Context Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.05571","noteUrl":"/review?paper=review-neurips-2026-152416","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152418","title":"DELTAVID: Enhancing Fine-Grained Spatiotemporal Perception with Cross-Video Differences","authors":["Yankai Yang","Yancheng Long","Bin Wen","Fan Yang","Tingting Gao","Han Li","Shuo Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Ground-truth averaging does not penalize duplicate predictions","plainExplanation":"The reward prose promises extra-prediction penalties and a highest matching score, while Appendix A.2 averages greedy matched scores over ground truths. Adding a duplicate perfect detection leaves the sole ground-truth match at one, giving no penalty.","primaryUrl":"https://arxiv.org/html/2607.02551","noteUrl":"/review?paper=review-neurips-2026-152418","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152419","title":"Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling","authors":["KunHo Heo","SuYeon Kim","Hayoung Lee","Chanse Oh","MyeongAh Cho"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35059","noteUrl":"/review?paper=review-neurips-2026-152419","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152420","title":"Total Variation Rates for Riemannian Flow Matching","authors":["Yunrui Guan","Krishnakumar Balasubramanian","Shiqian Ma"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Euler-interpolant divergence includes an inverse-map Jacobian term","plainExplanation":"Euler-interpolant divergence includes an inverse-map Jacobian term. A mean-square bound by epsilon does not imply a bound by epsilon squared.","primaryUrl":"https://arxiv.org/html/2602.05174","noteUrl":"/review?paper=review-neurips-2026-152420","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152421","title":"Patch Rebirth: Fast and Transferable Model Inversion of Vision Transformers","authors":["Seongsoo Heo","Dong-Wan Choi"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Patch Rebirth: Fast and Transferable Model Inversion of Vision Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.23235","noteUrl":"/review?paper=review-neurips-2026-152421","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152423","title":"Smooth Piecewise Cutting for Neural Operator to Handle Discontinuities and Sharp Transitions","authors":["Ha Dang","Sebastian Schmidt","Jürgen W Hesser"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Smooth Piecewise Cutting for Neural Operator to Handle Discontinuities and Sharp Transitions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19823","noteUrl":"/review?paper=review-neurips-2026-152423","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152426","title":"Addressable Memory for Video World Models","authors":["Xindi Wu","Sven Elflein","James Lucas","Olga Russakovsky","Laura Leal-Taixé","Despoina Paschalidou","Jonathan Lorraine","Aljosa Osep"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Addressable Memory for Video World Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.07408","noteUrl":"/review?paper=review-neurips-2026-152426","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152427","title":"Positional versus Symbolic Attention Heads: Learning Dynamics, RoPE Geometry, and Length Generalization","authors":["Felipe Urrutia","Juan J Alegría","Cinthia Sánchez","Jorge Salas","Cristian Buc Calderon","Cristobal Rojas"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Positional versus Symbolic Attention Heads: Learning Dynamics, RoPE Geometry, and Length Generalization","plainExplanation":"The nearby angular competitor found in the enlarged set need not be an available key.","primaryUrl":"https://arxiv.org/html/2605.31558","noteUrl":"/review?paper=review-neurips-2026-152427","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152431","title":"Self-Distilled RLVR","authors":["Chenxu Yang","Chuanyu Qin","Qingyi Si","Minghui Chen","Naibin Gu","Dingyu Yao","Zheng Lin","Weiping Wang","Nan Duan","Jiaqi Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Self-Distilled RLVR","plainExplanation":"The gradient-variance identity needs teacher-token cross-covariances, and positive detached token weights can change the summed parameter-update direction. The finite-batch update display also needs the policy factor already supplied by the GRPO-integration prose and score-function formula.","primaryUrl":"https://arxiv.org/html/2604.03128","noteUrl":"/review?paper=review-neurips-2026-152431","collection":"neurips-2026","severity":"critical","issueTypes":["lemma_error","notation"],"classifications":["critical"]},{"slug":"review-neurips-2026-152433","title":"Amortized Structured Stochastic Variational Inference for Gaussian Process Latent Variable Models","authors":["Maksym Tretiakov","Sarah Filippi","Vincent Fortuin","Ruth Misener","Ruby Sedgwick","James Odgers"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Amortized Structured Stochastic Variational Inference for Gaussian Process Latent Variable Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03647","noteUrl":"/review?paper=review-neurips-2026-152433","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152435","title":"Hydra-DP3: Frequency-Aware Right-Sizing of 3D Diffusion Policies for Visuomotor Control","authors":["Jinhao Zhang","Zhexuan Zhou","Huizhe Li","Yichen Lai","Wenlong Xia","Haoming Song","Youmin Gong","Jie Mei"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The numerical substitution omits a factor in the additive error","plainExplanation":"The numerical substitution omits a factor in the additive error. A forward-denoising theorem does not directly bound the generated DDIM midpoint.","primaryUrl":"https://arxiv.org/html/2605.01581","noteUrl":"/review?paper=review-neurips-2026-152435","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152437","title":"Importance-Aware OBS Pruning for Diffusion Models","authors":["Ba-Thinh Lam","Srijan Das","Hieu Minh Le"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Importance-Aware OBS Pruning for Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.20048","noteUrl":"/review?paper=review-neurips-2026-152437","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152440","title":"Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization","authors":["Xinwu Ye","He CAO","Li Hao","Bin Feng","Zijing Liu","Robert Tang","Yu Li","Shenghua Gao"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08659","noteUrl":"/review?paper=review-neurips-2026-152440","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152442","title":"VideoSTF: Stress-Testing Output Repetition in Video Large Language Models","authors":["Yuxin Cao","Yuxin Cao","Shangzhi Xu","Jingling Xue","Jin Song Dong"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"VideoSTF: Stress-Testing Output Repetition in Video Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.10639","noteUrl":"/review?paper=review-neurips-2026-152442","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152444","title":"Probing Persona-Dependent Preferences in Language Models","authors":["Oscar Gilg","Pierre Beckmann","Daniel Paleka","Patrick Butlin"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Probing Persona-Dependent Preferences in Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13339","noteUrl":"/review?paper=review-neurips-2026-152444","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152447","title":"Global Optimality for Constrained Exploration via Penalty Regularization","authors":["Florian Wolf","Ilyas Fatkhullin","Niao He"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Bounded softmax occupancies are not closed toward the original constrained optimum","plainExplanation":"Lemma 0.C.1 claims a bounded anchored-softmax class satisfies the local hidden-convexity assumption toward the original constrained optimizer. In a one-state two-action MDP, normalized occupancy is the action probability.","primaryUrl":"https://arxiv.org/html/2604.28144","noteUrl":"/review?paper=review-neurips-2026-152447","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152448","title":"Beyond Maximum Likelihood: Variational Inequality Estimation for Generalized Linear Models","authors":["Linglingzhi Zhu","Jonghyeok Lee","Yao Xie"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Monotonicity and Lipschitz continuity do not imply the cocoercivity used in the proof","plainExplanation":"Monotonicity and Lipschitz continuity do not imply the cocoercivity used in the proof. A residual bound at the true parameter is not a uniform bound at every iterate.","primaryUrl":"https://arxiv.org/html/2511.03087","noteUrl":"/review?paper=review-neurips-2026-152448","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152449","title":"Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents","authors":["Sikuan Yan","Ahmed Bahloul","Ercong Nie","Susanna Schwarzmann","Riccardo Trivisonno","Volker Tresp","Yunpu Ma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21768","noteUrl":"/review?paper=review-neurips-2026-152449","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152450","title":"ImmuVis: Hyperconvolutional Foundation Models for Imaging Mass Cytometry","authors":["Dawid Uchal","Marcin Możejko","Krzysztof Gogolewski","Piotr Kupidura","Fabian Ozga","Szymon Łukasik","Jakub Giezgała","Tomasz Nocoń","Kacper Pietrzyk","Robert Pieniuta","Mateusz Sulimowicz","Michał Zmysłowski","Michał Orzyłowski","Tomasz Siłkowski","Karol Zagródka","Eike Staub","Ewa Szczurek"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ImmuVis: Hyperconvolutional Foundation Models for Imaging Mass Cytometry","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04585","noteUrl":"/review?paper=review-neurips-2026-152450","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152453","title":"DiDE: Direct Injection with Color-Texture DEcoupling for 3D Stylization","authors":["Tao Wu","Alexandra Gomez-Villa","Senmao Li","Yaxing Wang","Joost van de Weijer","Kai Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DiDE: Direct Injection with Color-Texture DEcoupling for 3D Stylization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02044","noteUrl":"/review?paper=review-neurips-2026-152453","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152455","title":"MoBiQuant: Mixture-of-Bits Quantization for Token-Adaptive Any-Precision LLM","authors":["Dongwei Wang","Jinhee Kim","Seokho Han","Denis Gudovskiy","Yohei Nakata","Tomoyuki Okuno","KhayTze Peong","Kang E Jeon","Jong Hwan Ko","Yiran Chen","Huanrui Yang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"MoBiQuant: Mixture-of-Bits Quantization for Token-Adaptive Any-Precision LLM","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.20191","noteUrl":"/review?paper=review-neurips-2026-152455","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152456","title":"Training-Free Dynamic Upcycling of Expert Language Models","authors":["Eros Fanì","Oguzhan Ersoy"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Order-independent ridge statistics do not guarantee no forgetting","plainExplanation":"Order-independent ridge statistics do not guarantee no forgetting.","primaryUrl":"https://arxiv.org/html/2603.29765","noteUrl":"/review?paper=review-neurips-2026-152456","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152457","title":"Exact Unlearning via Quantized Sufficient Statistics","authors":["Ami Tavory","Shripad Gade","Tal Sarig","Noam Touitou","Ido Guy"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The simultaneous-batch outage threshold is forty-nine","plainExplanation":"The simultaneous-batch outage threshold is forty-nine. The relative outage correction does not have the claimed small parameter.","primaryUrl":"https://arxiv.org/html/2610.07197","noteUrl":"/review?paper=review-neurips-2026-152457","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152459","title":"Decoupled Descent: Exact Test Error Tracking Via Approximate Message Passing","authors":["Max Lovig"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The test expectation uses the parameter-overlap covariance","plainExplanation":"Definition 3.5 uses Gaussian variance \\(\\Omega\\), the parameter-overlap limit, whereas Equation (3.6) substitutes separate descent-statistic covariance \\(\\Sigma\\). The covariances can differ at initialization.","primaryUrl":"https://arxiv.org/html/2604.27883","noteUrl":"/review?paper=review-neurips-2026-152459","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152460","title":"STRAND: Sequence-Conditioned Transport for Single-Cell Perturbations","authors":["Boyang Fu","Sameer Gabbita","George Dasoulas","xiang lin","Shanghua Gao","Xiaorui Su","Soumya Ghosh","Marinka Zitnik"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Reversing positive variance increments does not make their sum decrease","plainExplanation":"The prose describes cumulative diffusion variance as increasing then decreasing in the bridge construction, although each increment is a nonnegative schedule value. Reordering nonnegative increments cannot decrease their cumulative sum.","primaryUrl":"https://arxiv.org/html/2602.10156","noteUrl":"/review?paper=review-neurips-2026-152460","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152461","title":"GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring","authors":["Zechen Li","Keerthana Natarajan","Weizhi Zhang","Simon Lee","Yuwei Zhang","Max Xu","Menglian Zhou","Zeinab Esmaeilpour","Flora Salim","Mark Malhotra","Lindsey Sunden","Shwetak Patel","Yuzhe Yang","Ahmed Metwally"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Beating a fixed mean-glucose rule does not prove extra information","plainExplanation":"The same-window comparison interprets better classification than a fixed GMI threshold as evidence of information beyond mean glucose; GMI is an invertible affine transform of that mean. A representation equal to mean glucose contains no additional information, yet a learned threshold at 120 can outperform a different fixed threshold applied to GMI.","primaryUrl":"https://arxiv.org/html/2605.30865","noteUrl":"/review?paper=review-neurips-2026-152461","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152462","title":"Learning to Decide with AI Assistance under Human-Alignment","authors":["Nina Corvelo Benz","Eleni Straitouri","Manuel Rodriguez"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stated regret rate drops the hypothesis-class factor","plainExplanation":"The stated regret rate drops the hypothesis-class factor. Strict cross-coordinate alignment does not imply within-coordinate monotonicity.","primaryUrl":"https://arxiv.org/html/2605.12646","noteUrl":"/review?paper=review-neurips-2026-152462","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-152464","title":"How You Move Tells What You'll Do: Trajectory-Conditioned Egocentric Prediction","authors":["SeJoon Jun","Hai Nguyen-Truong","Luigi Seminara","Lorenzo Torresani"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Endpoint scoring does not require monotone intermediate progress","plainExplanation":"Appendix B.2 explains CEM failure by saying its search requires intermediate latent distances to decrease, while the stated candidate score uses only final-goal distance. Two trajectories with the same final latent have the same score even if one initially moves away from the goal.","primaryUrl":"https://arxiv.org/html/2605.20388","noteUrl":"/review?paper=review-neurips-2026-152464","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152470","title":"Bias-Variance Optimized Preference Optimization for Large Reasoning Models","authors":["Mingkang Zhu","Xi Chen","Bei Yu","Hengshuang Zhao","Jiaya Jia"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Bias-Variance Optimized Preference Optimization for Large Reasoning Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.05095v1","noteUrl":"/review?paper=review-neurips-2026-152470","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152472","title":"Reconsidering Positional Supervision in Masked Diffusion Language Model Training","authors":["Mengyu Ye","Keito Kudo","Ryosuke Takahashi","Jun Suzuki"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The masking probability is the complement of the survival probability","plainExplanation":"The SFT sentence calls survival probability the masking probability. This reverses both endpoints of the forward process it says is unchanged. This is a local probability-convention slip. The equations specify the intended process; actual masking code is not established by the prose.","primaryUrl":"https://arxiv.org/html/2601.22947","noteUrl":"/review?paper=review-neurips-2026-152472","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152475","title":"V2M-Zero: Zero-Pair Time-Aligned Video-to-Music Generation","authors":["Yan-Bo Lin","Jonah Casebeer","Long Mai","Aniruddha Mahapatra","Gedas Bertasius","Nicholas J. Bryan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The sampling ODE sign opposes the trained interpolation derivative","plainExplanation":"The path is \\(x_t=(1-t)x_0+t\\epsilon\\) and the training target is \\(\\epsilon-x_0\\), but the printed sampler negates that velocity while integrating \\(t\\) from one to zero. The path derivative with respect to \\(t\\) is already \\(\\epsilon-x_0\\); decreasing time requires \\(dx=f\\,dt\\).","primaryUrl":"https://arxiv.org/html/2603.11042","noteUrl":"/review?paper=review-neurips-2026-152475","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152477","title":"AIRA-Compose: Agentic Discovery of Neural Architectures","authors":["Alberto Pepe","Chien-Yu Lin","Despoina Magka","Bilge Acun","Yannan N Wu","Anton Protopopov","Carole-Jean Wu","Yoram Bachrach"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AIRA-Compose: Agentic Discovery of Neural Architectures","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15871v2","noteUrl":"/review?paper=review-neurips-2026-152477","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152478","title":"An exact information theory of generalization phase transitions in Bayesian diffusion models","authors":["Henry Hunt","Mason Kamb","Surya Ganguli"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Real Gaussian mutual information retains half-log-determinants","plainExplanation":"The hierarchical Gaussian entropy calculation is for real covariance matrices, but Equation (162) drops one half from both log-determinant contributions. Real Gaussian entropy is \\(d\\log(2\\pi e)/2+\\log\\det C/2\\).","primaryUrl":"https://arxiv.org/html/2607.08041","noteUrl":"/review?paper=review-neurips-2026-152478","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152480","title":"Graph Energy Matching: Transport-Aligned Energy-Based Modeling for Graph Generation","authors":["Michal Balcerak","Suprosanna Shit","Chinmay Prabhakar","Sebastian Kaltenbach","Michael Albergo","Yilun Du","Bjoern Menze"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Graph Energy Matching: Transport-Aligned Energy-Based Modeling for Graph Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.23398","noteUrl":"/review?paper=review-neurips-2026-152480","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152482","title":"Beyond Suspicious Steps: Ontological Trust in Long-Horizon Agents","authors":["一个 他","Yao Wang","Haibin Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Beyond Suspicious Steps: Ontological Trust in Long-Horizon Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.17718","noteUrl":"/review?paper=review-neurips-2026-152482","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152485","title":"Accelerating Diffusion Language Models via Structured Suffix Modeling","authors":["Zifeng Cheng","Keda Li","Zhiwei Jiang","Cong Wang","Fei Shen","Qing Gu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Accelerating Diffusion Language Models via Structured Suffix Modeling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.23167","noteUrl":"/review?paper=review-neurips-2026-152485","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152487","title":"The Sparsity Whisperer","authors":["Linghao Kong","Inimai Subramanian","Micah Adler","Dan Alistarh","Dan Gutfreund","Nir Shavit"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Sparsity Whisperer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2608.06630v1","noteUrl":"/review?paper=review-neurips-2026-152487","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152489","title":"Done, But Not Sure: Disentangling World Completion from Self-Termination in Embodied Agents","authors":["Ying Chen","Lihuang Fang","Rui Jiang","Mingxu Wang","Zhifeng Gu","Lei Yi","Jie Chen"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Done, But Not Sure: Disentangling World Completion from Self-Termination in Embodied Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08747","noteUrl":"/review?paper=review-neurips-2026-152489","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152493","title":"Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?","authors":["Nils Grandien","David Steinmann","Felix Friedrich","Kristian Kersting"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.22994","noteUrl":"/review?paper=review-neurips-2026-152493","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152494","title":"Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing","authors":["Jiacheng Miao","Jin Mu","Guanhua Chen","James Zou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.07437","noteUrl":"/review?paper=review-neurips-2026-152494","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152496","title":"Early Semantic Grounding in Image Editing Models for Zero-Shot Referring Image Segmentation","authors":["Jingxuan He","Xiyu Wang","Yunke Wang","Mengyu Zheng","Chang Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Early Semantic Grounding in Image Editing Models for Zero-Shot Referring Image Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13122","noteUrl":"/review?paper=review-neurips-2026-152496","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152499","title":"Grounding Driving VLA via Inverse Kinematics","authors":["Junsung Park","Hyunjung Shim"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Future variance does not ensure useful visual prediction or a nonzero gradient","plainExplanation":"The certificate compares projected prediction error with raw future-feature variance conditional only on command, though the blind predictor may also use ego input. Conditional total variance makes command-only variance an upper bound on the ego-and-command blind Bayes floor, not a lower bound.","primaryUrl":"https://arxiv.org/html/2605.21061","noteUrl":"/review?paper=review-neurips-2026-152499","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152501","title":"Mitigating Label Bias with Interpretable Rubric Embeddings","authors":["Calvin Isley","Johann D. Gaebler","Sharad Goel"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"A Bayes bias identity need not hold for calibrated ridge predictions","plainExplanation":"Equation (4) concerns conditional expectations, but is applied as an exact identity to fitted ridge plus scalar recalibration. Unequal feature shrinkage can survive calibration.","primaryUrl":"https://arxiv.org/html/2605.21455","noteUrl":"/review?paper=review-neurips-2026-152501","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152502","title":"Solver-Aware Decompositions for Programming-by-Example: When Dividing Requires Knowing how to Conquer","authors":["Janis Zenkner","Tobias Sesterhenn","Tim Grams","Christian Bartelt"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Residual-prefix updates do not remove decomposition ambiguity","plainExplanation":"A ground-truth-promoting beam oracle is not an upper bound on useful reranking. Residual-prefix updates do not remove decomposition ambiguity. RobustFill completion must compare with the current residual.","primaryUrl":"https://arxiv.org/html/2608.03461","noteUrl":"/review?paper=review-neurips-2026-152502","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152503","title":"UNIVERSE: Unified Video Action Models for Autonomous Driving with Flexible Mask-Modulated Modality Generation","authors":["Mengmeng Liu","Diankun Zhang","Jiuming Liu","Jianfeng Cui","Hongwei Xie","Guang Chen","Hangjun Ye","Francesco Nex","Hao Cheng","Michael Yang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"UNIVERSE: Unified Video Action Models for Autonomous Driving with Flexible Mask-Modulated Modality Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.05133","noteUrl":"/review?paper=review-neurips-2026-152503","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152506","title":"Provable Test-Time Scaling for Beam Search in LLM Reasoning","authors":["Qijia He","Yu Huang","Yuan Cheng","Yuxin Chen","Yingbin Liang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Zero-sum Gaussian log perturbations can exceed unit probability","plainExplanation":"The simulator assigns token probabilities \\(\\exp(\\log p/L+\\xi_t)\\), with centered Gaussian perturbations satisfying \\(\\sum_t\\xi_t=0\\). The product equals \\(p\\), but individual factors need not lie in \\([0,1]\\).","primaryUrl":"https://arxiv.org/html/2609.38672","noteUrl":"/review?paper=review-neurips-2026-152506","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152511","title":"ALAM: Algebraically Consistent Latent Transitions for Vision-Language-Action Models","authors":["Zuojin Tang","Haoyun Liu","Xinyuan Chang","Changjie Wu","Dongjie Huo","Yandan Yang","Bin Liu","Zhejia Cai","Feng Xiong","Mu Xu","jiachen luo","De Ma","Zhiheng Ma","Gang Pan"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"ALAM: Algebraically Consistent Latent Transitions for Vision-Language-Action Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10819","noteUrl":"/review?paper=review-neurips-2026-152511","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152515","title":"Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE","authors":["Yu Xu","Yuxin Zhang","Xiao Yang","Haotian Yang","Yizhi Wang","Xinwei Huang","Minxuan Lin","Angtian Wang","Chongyang Ma","Fan Tang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38140v1","noteUrl":"/review?paper=review-neurips-2026-152515","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152518","title":"Disambiguating 2D-3D Correspondences in Gaussian Splatting-based Feature Fields for Visual Localization","authors":["Miso Lee","Sangeek Hyun","Yerim Jeon","Jae-Pil Heo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A view-wide nonempty test does not filter each Gaussian","plainExplanation":"The displayed filter tests whether any Gaussian qualifies in a view instead of whether the current Gaussian qualifies. It can retain a weak Gaussian solely because a different one passes the threshold. This is a missing membership condition in the printed filter. An implicit prior restriction could implement the intended algorithm, so no code behavior is inferred.","primaryUrl":"https://arxiv.org/html/2605.07351","noteUrl":"/review?paper=review-neurips-2026-152518","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152522","title":"Manifold Random Features","authors":["Ananya Parashar","Derek Long","Dwaipayan Saha","Krzysztof M Choromanski"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Discrete heat masses need volume scaling and a periodic limit","plainExplanation":"Equation (9) claims entries of the graph heat matrix converge directly to continuum heat-kernel densities. On a unit circle, a graph diagonal is \\(n^{-1}\\sum_ke^{-4tn^2\\sin^2(\\pi k/n)}\\), tending to zero at fixed positive time.","primaryUrl":"https://arxiv.org/html/2602.03797","noteUrl":"/review?paper=review-neurips-2026-152522","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152525","title":"Calibration without Ground Truth","authors":["Yuqing Kong","Mingyu Song","Yizhou Wang","Yifan Wu"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Pooling reports can make the calibration projection increase loss","plainExplanation":"The compatible transformations in Eq. (2) form a nonconvex set because a transformation can merge distinct reports. A finite example makes the selected projection increase squared loss. Defining compatibility conditional on the original reports gives a convex feasible set and restores the proper-loss improvement argument.","primaryUrl":"https://arxiv.org/html/2601.19862","noteUrl":"/review?paper=review-neurips-2026-152525","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152526","title":"KG-Guard: Graph-Based Hallucination Detection for Knowledge Base Question Answering","authors":["Albert Sawczyn","Piotr Bielak","Tomasz Kajdanowicz"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"KG-Guard: Graph-Based Hallucination Detection for Knowledge Base Question Answering","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.00328","noteUrl":"/review?paper=review-neurips-2026-152526","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152528","title":"How Useful Is Cross-Domain Generalization for Training LLM Monitors?","authors":["Sam Martin","Fabien Roger"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"How Useful Is Cross-Domain Generalization for Training LLM Monitors?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12265","noteUrl":"/review?paper=review-neurips-2026-152528","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152529","title":"LLM Flow Processes for Text-Conditioned Regression","authors":["Felix Biggs","Samuel Willis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The denoiser’s velocity coefficient needs a minus sign","plainExplanation":"The general denoiser identity gives its velocity coefficient the wrong sign. Solving the two linear mixing equations requires a minus sign absent from the displayed coefficient. This is a local coefficient typo. A change of time convention must alter all coefficients consistently; no implementation or sampling failure follows.","primaryUrl":"https://arxiv.org/html/2601.06147","noteUrl":"/review?paper=review-neurips-2026-152529","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152531","title":"Unified Noise Steering for Efficient Human-Guided VLA Adaptation","authors":["Junjie Lu","Xinyao Qin","Yuhua Jiang","Kaixin Wang","Chuheng Zhang","Bin Liang","Jun Yang","Min Xu","Li Zhao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Unified Noise Steering for Efficient Human-Guided VLA Adaptation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10821","noteUrl":"/review?paper=review-neurips-2026-152531","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152533","title":"Conformal Prediction with Paraphrase-Aware Scoring for LLM Uncertainty Quantification","authors":["Jiayi Xin","Evan Qiang","Zihan Zhu","Xiang Li","Weijie Su","Qi Long"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Conformal Prediction with Paraphrase-Aware Scoring for LLM Uncertainty Quantification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04239","noteUrl":"/review?paper=review-neurips-2026-152533","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152534","title":"Auto-Dreamer: Learning Offline Memory Consolidation for Language Agents","authors":["Chongrui Ye","Yuxiang Liu","Yu Wang","Haofei Yu","Yining Zhao","Ge Liu","Julian McAuley","Jiaxuan You"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Auto-Dreamer: Learning Offline Memory Consolidation for Language Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20616","noteUrl":"/review?paper=review-neurips-2026-152534","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152535","title":"Primal Generation, Dual Judgment: Self-Training from Test-Time Scaling","authors":["Yizhu Jiao","Ruixiang Zhang","Richard H Bai","Jiawei Han","Ronan Collobert","Yizhe Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Primal Generation, Dual Judgment: Self-Training from Test-Time Scaling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11299","noteUrl":"/review?paper=review-neurips-2026-152535","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152537","title":"Understanding Multi-View Transformers","authors":["Julien Gaubil","Michal Stary","Louis Martinez","Andreas Geiger","Maks Ovsjanikov","Ayush Tewari","Vincent Sitzmann"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Independent view normalization removes relative scale supervision","plainExplanation":"The training confidence loss normalizes each predicted view independently, while claiming shared coordinates constrain relative view scale. For every positive \\(c_v\\), view normalization cancels \\(c_v\\) exactly.","primaryUrl":"https://arxiv.org/html/2510.24907","noteUrl":"/review?paper=review-neurips-2026-152537","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152538","title":"$\\epsilon$-Good Action Identification in Fixed-Budget Monte Carlo Tree Search","authors":["Yinan Li","Ngo Tuan Nguyen","Kwang-Sung Jun"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"$\\epsilon$-Good Action Identification in Fixed-Budget Monte Carlo Tree Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.11324v1","noteUrl":"/review?paper=review-neurips-2026-152538","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152540","title":"PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs","authors":["Jiaqi Xue","Yifei Zhao","Mansour Al Ghanim","Shangqian Gao","Ruimin Sun","Qian Lou","Mengxin Zheng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The reverse-magnitude hinge is not convex","plainExplanation":"The regularizer includes \\(\\max(0,\\varepsilon-|u|)\\) for nonlinear MLP outputs and is said to be convex, even strongly convex, in network parameters. At \\(u=\\pm\\varepsilon\\) the scalar penalty is zero, but at their midpoint it equals \\(\\varepsilon\\), violating convexity.","primaryUrl":"https://arxiv.org/html/2510.23891","noteUrl":"/review?paper=review-neurips-2026-152540","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152543","title":"CLIOPATRA: Extracting Private Information from LLM Insights","authors":["Meenatchi Sundaram Muthu Selva Annamalai","Emiliano De Cristofaro","Peter Kairouz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CLIOPATRA: Extracting Private Information from LLM Insights","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.09781","noteUrl":"/review?paper=review-neurips-2026-152543","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152545","title":"On the Complexity of Preference-Based Bandits","authors":["Ahmed Ben Yahmed","Marc Abeille","Clément Calauzènes"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Feature differences contract RKHS norm but need not preserve it","plainExplanation":"The difference-kernel construction reuses coefficients and identifies the difference-function RKHS norm with the original function norm. Equivalent representatives can have smaller norm after the difference map removes a nullspace.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2609.39351","noteUrl":"/review?paper=review-neurips-2026-152545","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152546","title":"Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance","authors":["Gizem Yüce","Giorgos Nikolaou","Nicolas Flammarion"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06320","noteUrl":"/review?paper=review-neurips-2026-152546","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152547","title":"To Copy or Not to Copy: Controlling Speculative Decoding via Intrinsic Model Signals","authors":["Roy Eisenstadt","Ido Cohen","Edo Cohen-Karlik","Lior Wolf","Itamar Zimerman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Speculative acceptance is one minus total variation, not KL","plainExplanation":"The exact speculative acceptance identity uses the wrong divergence. Standard rejection acceptance averages to one minus total variation, whereas one minus KL can even be negative. This is a false background probability identity. It does not refute the measured decoding speedups or imply a failure of the acceptance implementation.","primaryUrl":"https://arxiv.org/html/2609.20186","noteUrl":"/review?paper=review-neurips-2026-152547","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152550","title":"AbFlowNet: Optimizing Antibody-Antigen Binding Energy via Diffusion-GFlowNet Fusion","authors":["Abrar Rahman Abir","Haz Sameen Shahgir","Rownok Z Ratul","Toki Tahmid","Greg Ver Steeg","Yue Dong"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The balance equation reverses diffusion kernels and includes an undefined endpoint","plainExplanation":"The indexing fixes \\(S^0\\) as data and \\(S^T\\) as noise, but Equation (17) reverses the generation/backward factors and extends to \\(S^{-1}\\). Generation must use \\(p(S^{t-1}\\mid S^t)\\), with reverse-path factors \\(q(S^t\\mid S^{t-1})\\), for \\(1\\le t\\le T\\).","primaryUrl":"https://arxiv.org/html/2505.12358","noteUrl":"/review?paper=review-neurips-2026-152550","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152556","title":"IDEAL: In-DEpth ALignment Makes A Discrete Representation AutoEncoder","authors":["Yitong Chen","Zijie Diao","Junke Wang","Lingyu Kong","Yixuan Ren","Bo He","Yu-Gang Jiang","Zuxuan Wu"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"IDEAL: In-DEpth ALignment Makes A Discrete Representation AutoEncoder","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.11096","noteUrl":"/review?paper=review-neurips-2026-152556","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152558","title":"Persuasive Prediction via Decision Calibration","authors":["Jingwu Tang","Jiahao Zhang","Fei Fang","Steven Wu"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The best-response sign and calibration tolerance need correction","plainExplanation":"The two printed inequalities force \\(r=\\gamma\\), not \\(-\\gamma\\leq r\\leq\\gamma\\). A perfectly calibrated zero residual is incorrectly penalized when \\(\\gamma>0\\).","primaryUrl":"https://arxiv.org/html/2505.16141","noteUrl":"/review?paper=review-neurips-2026-152558","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152559","title":"Generating Physically Consistent Molecules with Energy-Based Models","authors":["Christoph Griesbacher","Lea Bogensperger","Andreas Habring","Thomas Pock"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The replica-swap exponent has the reversed sign","plainExplanation":"The parallel-tempering equation uses the inverse of the required Metropolis–Hastings density ratio. In the analysis, the weighted norm comparison needs a distribution-dependent constant, and uniqueness of energy values at zero regularization holds only after fixing an additive constant. The finite-constant repair preserves the coercivity argument, while the paper’s mean-centered small-regularization proof already handles the additive ambiguity.","primaryUrl":"https://arxiv.org/html/2605.18381","noteUrl":"/review?paper=review-neurips-2026-152559","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error","incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152562","title":"GeoTransolver: Learning Physics on Irregular Domains using Multi-scale Geometry Aware Physics Attention Transformer","authors":["Corey Adams","Rishikesh Ranade","Sheel Nidhan","Ram Cherukuri","Sanjay Choudhry","Mohammad Amin Nabian","Deepak Akhare"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The geometric processor arguments reverse the query count","plainExplanation":"Two processor calls reverse the query and support arguments. Under the defined query-first operator they return the wrong number of rows for their declared geometry/input embeddings. This is a displayed argument-order error. The prose identifies the intended neighborhoods; no runtime fault is established.","primaryUrl":"https://arxiv.org/html/2512.20399","noteUrl":"/review?paper=review-neurips-2026-152562","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152563","title":"Language Model Memory and Memory Models for Language","authors":["Benjamin L Badger"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Uniform-baseline cross-entropy reduction is not a fraction of retained input information","plainExplanation":"Uniform-baseline cross-entropy reduction is not a fraction of retained input information. The denominator reverses prediction and target in cross-entropy.","primaryUrl":"https://arxiv.org/html/2602.13466","noteUrl":"/review?paper=review-neurips-2026-152563","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-152564","title":"Harnessing Streaming Video in the Wild","authors":["Dingyu Yao","Shuhuan Gu","Qingyi Si","Junhao Zhou","Chenxu Yang","Chuanyu Qin","Naibin Gu","Zheng Lin","Weiping Wang","Nan Duan","Jiaqi Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Harnessing Streaming Video in the Wild","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.08615","noteUrl":"/review?paper=review-neurips-2026-152564","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152570","title":"Derived Fields Preserve Fine-Scale Detail in Budgeted Neural Simulators","authors":["Wenshuo Wang","Fan Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Derived Fields Preserve Fine-Scale Detail in Budgeted Neural Simulators","plainExplanation":"No issue was found in the checked claims. The stronger pointwise interpretation challenged by the earlier report has been removed. An expected distortion scale is not a deterministic per-sample lower bound.","primaryUrl":"https://arxiv.org/html/2603.29224","noteUrl":"/review?paper=review-neurips-2026-152570","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152571","title":"Entropy Polarity in Reinforcement Fine-Tuning:  Direction, Asymmetry, and Control","authors":["Jiazheng Zhang","Ziche Fu","Junrui Shen","Yunbin Zhao","Yunke Zhang","Zhiheng Xi","LongMa","Chenxin An","Zhihao Zhang","Shichun Liu","Dingwei Zhu","Shihan Dou","Shaofan Liu","han li","Wiggin Zhou","Aiden Adams","Tao Gui","Fei Huang","Qi Zhang","Xuanjing Huang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The ablation bounds reverse expansion and contraction weighting","plainExplanation":"The appendix describes its exploit-biased interval as suppressing expansion, while the contracting branch receives \\(\\omega_{\\rm neg}\\) and the expanding branch its reciprocal. For \\(\\omega_{\\rm neg}\\in[0.96,1]\\), contraction is downweighted and expansion weakly amplified.","primaryUrl":"https://arxiv.org/html/2605.11775","noteUrl":"/review?paper=review-neurips-2026-152571","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152573","title":"Pretraining Curricula Enable Selective Fine-tuning","authors":["Kai J Sandbrink","Mia Whitefield","Sebastian Bruijns","Jirko Rubruck","Laurence T Hunt","Fazl Barez","Christopher Summerfield"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Pretraining Curricula Enable Selective Fine-tuning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.04846","noteUrl":"/review?paper=review-neurips-2026-152573","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152578","title":"Deep-Koopman-KANDy: Dictionary Discovery for Deep-Koopman Operators with Kolmogorov-Arnold Networks for Dynamics","authors":["Kevin Slote","Jeremie Fish"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Deep-Koopman-KANDy: Dictionary Discovery for Deep-Koopman Operators with Kolmogorov-Arnold Networks for Dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06000","noteUrl":"/review?paper=review-neurips-2026-152578","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152582","title":"Emergence of Physical Intelligence via Controllable Information Production","authors":["Tristan Shah","Stas Tiomkin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Emergence of Physical Intelligence via Controllable Information Production","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.22449","noteUrl":"/review?paper=review-neurips-2026-152582","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152589","title":"Locality Sensitive Hashing for p-Exponential Kernels with Applications to Density Estimation","authors":["Barak Gorodissky","Tal Wagner"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Locality Sensitive Hashing for p-Exponential Kernels with Applications to Density Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2610.05174v1","noteUrl":"/review?paper=review-neurips-2026-152589","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152591","title":"GenCOPE: Syn2Real Generalized Category-Level Object Pose Estimation for Robotic Picking","authors":["Jian Liu","Wei Sun","Zhenqi Dai","Hui Yang","Jian Xiao","Nicu Sebe","Na Zhao"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"GenCOPE: Syn2Real Generalized Category-Level Object Pose Estimation for Robotic Picking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2610.01758v1","noteUrl":"/review?paper=review-neurips-2026-152591","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152592","title":"Conservation Laws for Diffusion Models","authors":["Ziv Aharoni","Henry Pfister"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Conservation Laws for Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.10067","noteUrl":"/review?paper=review-neurips-2026-152592","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152593","title":"AmbientFM: A Foundation Model for Ambient Sensing","authors":["Guozhen Zhu","Yuqian Hu","Sakila S Jayaweera","Wei-Hsiang Wang","Weihang Gao","Jiaxuan Zhang","Beibei Wang","Chenshu Wu","K. Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AmbientFM: A Foundation Model for Ambient Sensing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.11200v1","noteUrl":"/review?paper=review-neurips-2026-152593","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152595","title":"DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence","authors":["Xu Huang","Ye Huang","Zijun Liao","Yuwei Niu","Xiaojie Li","Menghan Zhou","De Wen Soh","Xiaotong Li","Daquan Zhou"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39222","noteUrl":"/review?paper=review-neurips-2026-152595","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152597","title":"Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models","authors":["Bing Wang","Changchun Li","Xin-Qiang Cai","Lin Y Wu","Ximing Li","Gang Niu","Masashi Sugiyama"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Gumbel-softmax converges to a perturbed sample rather than the greedy token","plainExplanation":"Gumbel-softmax converges to a perturbed sample rather than the greedy token. Gradient orthogonality removes only first-order interference. A worst prompt at virtual weights need not bound losses at the updated weights.","primaryUrl":"https://arxiv.org/html/2609.30935","noteUrl":"/review?paper=review-neurips-2026-152597","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152599","title":"Rethinking Training Targets, Architectures and Data Quality for Universal Speech Enhancement","authors":["Szu-Wei Fu","Rong Chao","Xuesong Yang","Sung-Feng Huang","Ryandhimas E Zezario","Rauf Nasretdinov","Ante Jukić","Yu Tsao","Frank Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A small Lipschitz loss value does not imply a vanishing gradient","plainExplanation":"The feature-matching explanation uses a Lipschitz bound on an unsquared discrepancy to infer no gradient contribution when signal error is approximately zero. With identity feature map, target zero and prediction \\(\\epsilon>0\\), the unsquared absolute loss is \\(\\epsilon\\) but its derivative is one for every \\(\\epsilon\\).","primaryUrl":"https://arxiv.org/html/2603.02641","noteUrl":"/review?paper=review-neurips-2026-152599","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152600","title":"CoPE-VideoLM: Leveraging Codec Primitives For Efficient Video Language Modelling","authors":["Sayan Deb Sarkar","Rémi Pautrat","Ondrej Miksik","Marc Pollefeys","Iro Armeni","Mahdi Rad","Mihai Dusmanu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"CoPE-VideoLM: Leveraging Codec Primitives For Efficient Video Language Modelling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.13191","noteUrl":"/review?paper=review-neurips-2026-152600","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152604","title":"Fusion or Confusion? Multimodal Complexity Is Not All You Need","authors":["Tillmann Rheude","Roland Eils","Benjamin Wild"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Fusion or Confusion? Multimodal Complexity Is Not All You Need","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.22991","noteUrl":"/review?paper=review-neurips-2026-152604","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152605","title":"RVLoss: Runoff Vote Loss for Self-Supervised LiDAR Scene Flow Estimation","authors":["Shiming Wang","Liangliang Nan","Julian Kooij","Holger Caesar","Yancong Lin"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"RVLoss: Runoff Vote Loss for Self-Supervised LiDAR Scene Flow Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.18864","noteUrl":"/review?paper=review-neurips-2026-152605","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152609","title":"On the Convergence of Multicalibration Gradient Boosting","authors":["Daniel Haimovich","Fridolin Linder","Lorenzo Perini","Niek Tax","Milan Vojnovic"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Vanishing prediction steps need not imply multicalibration without bounded learners","plainExplanation":"Vanishing prediction steps need not imply multicalibration without bounded learners. The oracle residual bound applies at the full step rather than the shrunken iterate. The projector can leave an unlearnable residual that prevents quadratic convergence.","primaryUrl":"https://arxiv.org/html/2602.06773","noteUrl":"/review?paper=review-neurips-2026-152609","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152612","title":"AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs","authors":["Haizhong Zheng","Yizhuo Di","Jiahui Wang","Shuowei Jin","Xueshen Liu","Yongji Wu","Zhuoqing Morley Mao","Ion Stoica","Jiawei Zhao","Beidi Chen"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15565","noteUrl":"/review?paper=review-neurips-2026-152612","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152613","title":"The Rate-Distortion-Polysemanticity Tradeoff in SAEs","authors":["Tommaso Mencattini","Francesco Montagna","Francesco Locatello"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Rate-Distortion-Polysemanticity Tradeoff in SAEs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14694","noteUrl":"/review?paper=review-neurips-2026-152613","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152615","title":"Faster Rates For  Federated Variational Inequalities","authors":["Guanghui Wang","Satyen Kale"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The inner contraction bound has base three quarters, not one quarter","plainExplanation":"The proximal proof sets \\(\\gamma=[\\eta(L+1/\\eta)^2]^{-1}\\), assumes \\(\\eta L\\le1\\), and bounds repeated inner error using \\(4^{-H}\\). The displayed one-step factor is \\(1-2\\gamma/\\eta+\\gamma^2(L+1/\\eta)^2=1-(1+\\eta L)^{-2}\\).","primaryUrl":"https://arxiv.org/html/2602.09164","noteUrl":"/review?paper=review-neurips-2026-152615","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152622","title":"Sobolev Regularized MMD Gradient Flow","authors":["Chenyang Tian","Bharath Sriperumbudur","Arthur Gretton","Zonghao Chen"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The Stein-kernel Hessian need not be uniformly bounded","plainExplanation":"The first discrete-time bound needs the subsequent regularization coupling. The Stein-kernel Hessian need not be uniformly bounded.","primaryUrl":"https://arxiv.org/html/2605.11884","noteUrl":"/review?paper=review-neurips-2026-152622","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152623","title":"Displacement Geometry Captures Platonic Shared Reality Across Models and Modalities","authors":["Chenming Shang","Yujin Tang","Jun Jie Ou Yang","Ruize Xu","Adam Breuer","Nikhil Singh"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Displacement Geometry Captures Platonic Shared Reality Across Models and Modalities","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.24209","noteUrl":"/review?paper=review-neurips-2026-152623","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152624","title":"HyperTransport: Amortized Conditioning of T2I Generative Models","authors":["Valentino Maiorca","Eleonora Gualdoni","Xavier Suau","Marco Cuturi","Luca Zappella","Pau Rodriguez"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"HyperTransport: Amortized Conditioning of T2I Generative Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08254","noteUrl":"/review?paper=review-neurips-2026-152624","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152627","title":"Towards Understanding Momentum Acceleration in River-Valley Loss Landscape","authors":["Miao Lu","zeyu bian","Kaiyue Wen","Beining Wu","Siyu Chen","Tianhao Wang","Zhiyuan Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The neighborhood and gradient bounds are incompatible","plainExplanation":"The required neighborhood and gradient bounds make the nondegenerate river model empty, so the conditional tracking and acceleration results have no landscape to which they apply. Two auxiliary matrix statements also need a positive-product endpoint and a symmetric off-diagonal sign correction; their positive-product application and norm calculation avoid these local errors.","primaryUrl":"https://arxiv.org/html/2609.30957","noteUrl":"/review?paper=review-neurips-2026-152627","collection":"neurips-2026","severity":"major","issueTypes":["mathematical_error","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-152629","title":"Directional Confusions Reveal Divergent Inductive Biases Through Rate-Distortion Geometry in Human and Machine Vision","authors":["Leyla Roksan Caglar","Pedro A.M Mediano","Baihan Lin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Dissociated group contrasts do not establish statistical independence","plainExplanation":"The asymmetry analysis infers statistical independence from differing significance patterns of asymmetry and accuracy across groups. Groups with coordinates \\((0.6+z,1+2z)\\), \\((0.6+z,1.2+2z)\\) and \\((0.75+z,1+2z)\\) have the desired selective between-group changes while the two quantities are perfectly dependent within each group.","primaryUrl":"https://arxiv.org/html/2604.21909","noteUrl":"/review?paper=review-neurips-2026-152629","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152630","title":"Catch Your Breath: Adaptive Computation for Self-Paced Sequence Production","authors":["Alexandre Galashov","Matt Jones","Nan Rosemary Ke","Yuan Cao","Vaishnavh Nagarajan","Michael Mozer"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Marginalizing a random deadline inside the logarithm changes the objective","plainExplanation":"Marginalizing a random deadline inside the logarithm changes the objective. The reversed KL does not enforce the prior support restriction.","primaryUrl":"https://arxiv.org/html/2510.13879","noteUrl":"/review?paper=review-neurips-2026-152630","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-152631","title":"Localizing Input Uncertainty Quantification for Large Language Models via Shapley Values","authors":["Seongjun Lee","Suwan Yoon","Changhee Lee"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Localizing Input Uncertainty Quantification for Large Language Models via Shapley Values","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28170","noteUrl":"/review?paper=review-neurips-2026-152631","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"primal-dual-decoding-one-step-bound","title":"Primal-Dual Guided Decoding for Constrained Discrete Diffusion","authors":["Federico Tomasi","Dmitrii Moor","Alice Wang","Mounia Lalmas"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A post-sampling update cannot enforce the one-step constraint","plainExplanation":"A one-step decoder samples before its multiplier update, so its expected deficit can stay positive as the dual step size increases. The claimed bound nevertheless tends to zero. Strict feasibility does not resolve this mismatch; a valid guarantee must account for initialization and the order of sampling and optimization.","primaryUrl":"https://arxiv.org/html/2605.09749v1","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152633","title":"Robust Diffusion Models via Divergence-Induced Weighted Denoising","authors":["Lei Li","Yuexiao Dong"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The conditional-path equivalence is too strong","plainExplanation":"The reverse implication omits equality of the endpoint marginals.","primaryUrl":"https://arxiv.org/html/2606.22521","noteUrl":"/review?paper=review-neurips-2026-152633","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152635","title":"Targeted Synthetic Control Method","authors":["Yuxin Wang","Dennis Frauen","Emil Javurek","Konstantin Hess","Yuchen Ma","Stefan Feuerriegel"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Prediction-based tilting is not the derivative of the residual objective","plainExplanation":"The targeting equation uses weights proportional to \\(e^{\\epsilon S_j}\\), with centered predictions \\(S_j\\), while the claimed convex objective tilts by residuals \\(r_j=Y_j-m_j\\). Differentiating the log-sum-exp objective yields residual-weighted softmax \\(e^{\\epsilon r_j}\\), not the prediction-weighted residual equation.","primaryUrl":"https://arxiv.org/html/2602.04611","noteUrl":"/review?paper=review-neurips-2026-152635","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"visual-latents-best-state-selection","title":"Visual Latents Know More Than They Say: Unsilencing Latent Reasoning in MLLMs","authors":["xin zhang","Qiqi Tao","JIAWEI DU","Moyun Liu","Joey Tianyi Zhou"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The best-state test scores a different latent than it stores","plainExplanation":"The algorithm evaluates perturbed candidate \\(H+\\epsilon\\), then stores updated \\(H_{\\mathrm{next}}=H+\\alpha R(H+\\epsilon)\\epsilon/\\sigma^2\\) using the candidate's score. The update need not inherit that score.","primaryUrl":"https://arxiv.org/html/2605.02735v1","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152644","title":"Inline Critic Steers Image Editing","authors":["Weitai Kang","Xiaohang Zhan","Yizhou Wang","Mang Tik Chiu","Jason Kuen","Kangning Liu","Yan Yan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Inline Critic Steers Image Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12724","noteUrl":"/review?paper=review-neurips-2026-152644","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152645","title":"Stealthy World Model Manipulation via Data Poisoning","authors":["Yibin Hu","Zizhan Zheng"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The transition-gradient expansion needs stage-dependent discount powers","plainExplanation":"The deterministic finite-horizon expansion uses the same \\(\\gamma^t\\) for every future relative stage, despite its preceding discounted recursion. Repeated substitution multiplies the direct transition derivative at relative stage \\(i\\) by \\(\\gamma^i\\).","primaryUrl":"https://arxiv.org/html/2606.18697","noteUrl":"/review?paper=review-neurips-2026-152645","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152646","title":"DMax: Aggressive Parallel Decoding for dLLMs","authors":["Zigeng Chen","Gongfan Fang","Xinyin Ma","Ruonan Yu","Xinchao Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The fallback selects from an empty mask before checking termination","plainExplanation":"The decoder promotes positions using one threshold but stops only for unchanged top predictions or a separate confidence-acceptance threshold. Take decoding threshold zero, acceptance threshold \\(0.9\\), and 32 identical positions.","primaryUrl":"https://arxiv.org/html/2604.08302","noteUrl":"/review?paper=review-neurips-2026-152646","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152647","title":"Mixture of Probes: Learning with Privileged Modalities in Multimodal LLMs Through Probing","authors":["Dominick Reilly","Qiyu Wu","Hiromi Wakaki","Srijan Das","Yuki Mitsufuji"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mixture of Probes: Learning with Privileged Modalities in Multimodal LLMs Through Probing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.08839","noteUrl":"/review?paper=review-neurips-2026-152647","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152651","title":"CRAFT: Conflict-Resolved Aggregation for Federated Training","authors":["Ziqi Wang","Qiang Liu","Nils Thuerey"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CRAFT: Conflict-Resolved Aggregation for Federated Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21317","noteUrl":"/review?paper=review-neurips-2026-152651","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152652","title":"UniverSat: Resolution- and Modality-Agnostic Transformers for Earth Observation","authors":["Yohann PERRON","Guillaume Astruc","Nicolas Gonthier","Clement Mallet","Loic Landrieu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Pooling preserves channels rather than expanding them","plainExplanation":"The appendix defines max pooling of an \\(X\\times Y\\times D\\) tensor along \\(X\\), then uses the output as if its channel width were \\(\\alpha D\\). The output has shape \\(Y\\times D\\).","primaryUrl":"https://arxiv.org/html/2606.23503","noteUrl":"/review?paper=review-neurips-2026-152652","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152655","title":"Derivative-Informed Training of Neural Operators On-the-Fly via Sketched Tangent Consistency","authors":["Shancong (Sean) Mou","Xinhan Yang","Lu Lu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A variance ratio is not an additive noise attribution with covariance","plainExplanation":"The training-step diagnostic attributes total gradient variance to data and sketch terms using a component-variance ratio. For \\(D+W\\), variance is \\(\\operatorname{Var}D+\\operatorname{Var}W+2\\operatorname{Cov}(D,W)\\).","primaryUrl":"https://arxiv.org/html/2609.32797","noteUrl":"/review?paper=review-neurips-2026-152655","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152658","title":"Test-Time Defense Against Adversarial Attacks via Stochastic Resonance of Latent Ensembles","authors":["DONG LAO","Yuxiang Zhang","Haniyeh E Oskouie","Yangchao Wu","Alex Wong","Stefano Soatto"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Test-Time Defense Against Adversarial Attacks via Stochastic Resonance of Latent Ensembles","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.03224","noteUrl":"/review?paper=review-neurips-2026-152658","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152659","title":"Intrinsically Interpretable Attention via Sparse Post-Training","authors":["Florent Draye","Anson Lei","Hsiao-Ru Pan","Ingmar Posner","Bernhard Schölkopf"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Intrinsically Interpretable Attention via Sparse Post-Training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.05865","noteUrl":"/review?paper=review-neurips-2026-152659","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152660","title":"Q-ARVD: Quantizing Autoregressive Video Diffusion Models","authors":["Siao Tang","Xinyin Ma","Gongfan Fang","Xingyi Yang","Xinchao Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Q-ARVD: Quantizing Autoregressive Video Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21072","noteUrl":"/review?paper=review-neurips-2026-152660","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"space-group-torus-log-equivariance","title":"Space Group Conditional Flow Matching","authors":["Omri Puny","Yaron Lipman","Benjamin Kurt Miller"],"year":2025,"category":"Graphs & equivariance","status":"finding","issueTitle":"Componentwise torus displacements can break conditional-path symmetry","plainExplanation":"In nonorthogonal lattice coordinates, a valid crystal rotation need not preserve the componentwise principal displacement. The conditional logarithm and interpolation claims fail in that geometry, while the separate group-averaging inference construction and abstract equivariant-flow theorem remain intact.","primaryUrl":"https://arxiv.org/html/2509.23822v1","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152666","title":"Towards Fairness under Label Bias in Image Segmentation: Impact, Measurement and Mitigation","authors":["Aditya Parikh","Stella Christina Frank","Sneha Das","Aasa Feragen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Adaptive mean thresholds do not guarantee temperature invariance","plainExplanation":"Table 9 reports nearly unchanged metrics across the tested temperatures. Arithmetic class-conditional thresholds do not imply this invariance for arbitrary predictions. A nonlinear temperature transform can change whether an otherwise unchanged class assignment qualifies as a confident error. The main text already acknowledges calibration sensitivity; Appendix F.2 should retain that qualification when explaining its observed robustness.","primaryUrl":"https://arxiv.org/html/2605.06891","noteUrl":"/review?paper=review-neurips-2026-152666","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["false_statement"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152668","title":"Efficient Scaling of LLM Training with Flexible Context Parallelism","authors":["Yifan Niu","Han Xiao","Dongyi Liu","Wei zhou","Jia Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Efficient Scaling of LLM Training with Flexible Context Parallelism","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.21788","noteUrl":"/review?paper=review-neurips-2026-152668","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152669","title":"Optimal Representation Size: High-Dimensional Analysis of Pretraining and Linear Probing","authors":["Valentina Njaradi","Clémentine Dominé","Rachel A Swanson","Marco Mondelli","Andrew Saxe"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Optimal Representation Size: High-Dimensional Analysis of Pretraining and Linear Probing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20105","noteUrl":"/review?paper=review-neurips-2026-152669","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152671","title":"PGSB: Pretrained-Guided Shared Basis for LoRA Model Merging","authors":["Muqing Liu","Chongjie Si","Zhuoya Liu","Yuheng Jia"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PGSB: Pretrained-Guided Shared Basis for LoRA Model Merging","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.14558","noteUrl":"/review?paper=review-neurips-2026-152671","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152674","title":"The Sample Complexity of Multiple Change Point Identification under Bandit Feedback","authors":["Maximilian Graf","Victor Thuot"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The explicit sample-budget constant is too small","plainExplanation":"The current v2 description now matches the Algorithm 3 acceptance threshold. The proof still reverses its relation to the confidence radius, a minor wording error. Separately, Lemma 3 retains the insufficient explicit constant 223 and the same incorrect numerical step. The budget order and Theorem 1’s unspecified large constant are not refuted.","primaryUrl":"https://arxiv.org/html/2605.13252","noteUrl":"/review?paper=review-neurips-2026-152674","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-152675","title":"DeformMaster: An Interactive Physics-Neural World Model for Deformable Objects from Videos","authors":["Can Li","Zhoujian Li","Ren Li","Jie Gu","lei lei","Jingmin Chen","Lei Sun"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DeformMaster: An Interactive Physics-Neural World Model for Deformable Objects from Videos","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09586","noteUrl":"/review?paper=review-neurips-2026-152675","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152677","title":"Rethinking Personalized Generation: Test-time Alignment via Factorized Ranking Models","authors":["Qiyao Ma","Junshan Zhang","Zhe Zhao"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Rethinking Personalized Generation: Test-time Alignment via Factorized Ranking Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35695","noteUrl":"/review?paper=review-neurips-2026-152677","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152678","title":"GDMD: Guiding Distribution Matching Distillation with Gradient-Based Reinforcement Learning","authors":["Linwei Dong","Ruoyu Guo","Ge Bai","Quan Zheng","Yawei Luo","Changqing Zou"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Negative-aware weighting does not preserve every distillation descent direction","plainExplanation":"At the old policy, the preference gradient is a signed multiple of the base gradient, so rewards below one half can oppose it. The joint objective permits reversal for some positive loss weights; away from the old policy, an additional displacement term also affects the direction. This limits the unconditional alignment claim, without showing instability or reversal in the reported training runs.","primaryUrl":"https://arxiv.org/html/2604.19009","noteUrl":"/review?paper=review-neurips-2026-152678","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152680","title":"Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling","authors":["Alan Zhu","Mihran Miroyan","Carolyn Wang","Andrew Zhou","Lisa Dunlap","Narges Norouzi","Joseph Gonzalez"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26969","noteUrl":"/review?paper=review-neurips-2026-152680","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152681","title":"LensDesigner: A Self-Improving Agent for Optical Lens Design","authors":["Lei Sun","Haoran Liang","Dannong Xu","Yao Gao","Yuyu Geng","Jinjin Gu","Kaiwei Wang","Danda Pani Paudel","Luc V Gool"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LensDesigner: A Self-Improving Agent for Optical Lens Design","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.30450","noteUrl":"/review?paper=review-neurips-2026-152681","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152682","title":"SMILE: Bridging Continuous Optimization and Discrete Symbolic Recovery","authors":["Mansooreh Montazerin","Antonio Ortega","Ajitesh Srivastava"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The direct polynomial construction needs a positive-coordinate domain","plainExplanation":"The real logarithm construction is defined only for positive coordinates. Positive clamping does not preserve negative-input monomials.","primaryUrl":"https://arxiv.org/html/2609.04639","noteUrl":"/review?paper=review-neurips-2026-152682","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152683","title":"Your Embedding Model Is SMARTer Than You Think","authors":["Jianrui Zhang","Hyun Jung Lee","Sukanta Ganguly","Tae-Eui Kam","Donghyun Kim","Yong Jae Lee"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Your Embedding Model Is SMARTer Than You Think","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24938","noteUrl":"/review?paper=review-neurips-2026-152683","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152684","title":"TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs","authors":["Andong Hua","Colton Bishop","Igor Mordatch","Arian Hosseini","Jindong Gu","Aleksandra Faust","Rebecca Roelofs","Yao Qin"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.28640","noteUrl":"/review?paper=review-neurips-2026-152684","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152685","title":"Iterative Chow Filtering for Learning with Distribution Shift","authors":["Gautam Chandrasekaran","Georgios Gkrinias","Adam Klivans","Konstantinos Stavropoulos","Arsen Vasilyan"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Iterative Chow Filtering for Learning with Distribution Shift","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17251","noteUrl":"/review?paper=review-neurips-2026-152685","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152686","title":"FiRe: Fine-grained Multimodal Reasoning for Enhanced Image Generation","authors":["Yongjin Kim","Yoonjin Oh","Ye Rin Kim","Hyomin Kim","Jeeyoung Yun","Yu-Jung Heo","Min-Jun Kim","Sungwoong Kim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"FiRe: Fine-grained Multimodal Reasoning for Enhanced Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.13491","noteUrl":"/review?paper=review-neurips-2026-152686","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152687","title":"Mind the Heads: Topological Representation Alignment for Multimodal LLMs","authors":["Davide Caffagni","Alberto Compagnoni","Federico Melis","Sara Sarto","Pier Luigi Dovesi","Mark Granroth-Wilding","Marcella Cornia","Lorenzo Baraldi"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Mind the Heads: Topological Representation Alignment for Multimodal LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.23885","noteUrl":"/review?paper=review-neurips-2026-152687","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152690","title":"Multi-Agent Coordination via Support-Preserving Distillation","authors":["Sangmin Lee","Youngju Na","Chanmi Lee","Sung-eui Yoon"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Shared randomness does not represent every context-dependent product mixture","plainExplanation":"The fixed-context XOR example and the shared-latent mixture identity are correct. The broader representational claim allows mixture components to depend on observations unavailable to local actors. A four-context parity target cannot be represented by any context-independent shared random signal and local policies, even though its mixture weights are fixed and its individual marginals are consistent. The representation claim needs a jointly local decomposition or an explicitly richer coordination channel.","primaryUrl":"https://arxiv.org/html/2610.10087","noteUrl":"/review?paper=review-neurips-2026-152690","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-152696","title":"Securing AI Agents with Information-Flow Control","authors":["Manuel Costa","Boris Köpf","Aashish Kolluri","Andrew Paverd","Mark Russinovich","Ahmed Salem","Shruti Tople","Lukas Wutschitz","Santiago Zanella-Beguelin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Securing AI Agents with Information-Flow Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.23643","noteUrl":"/review?paper=review-neurips-2026-152696","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152698","title":"ASH: Agents that Self-Hone via Embodied Learning","authors":["Benjamin Schneider","Xavier Schneider","Victor Zhong","Sun Sun"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The written cloning loss has an extra index","plainExplanation":"The inclusive bounds request \\(w_s+1\\) actions, and the last hidden-state index exceeds the written sequence length.","primaryUrl":"https://arxiv.org/html/2605.14211","noteUrl":"/review?paper=review-neurips-2026-152698","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_ambiguity"],"classifications":["minor_cosmetic"]},{"slug":"persona-manifold-distance-ratio","title":"PersonaManifold: Revealing and Exploiting Curved Geometry in LLM Persona Representations","authors":["Rui Xu","Yinghui Xu","Libo Wu"],"year":2026,"category":"Representation learning","status":"addressed","issueTitle":"Metric scale changes the geodesic-to-Euclidean ratio on a flat line","plainExplanation":"A straight one-dimensional point cloud can have a ratio above or below one because only the numerator uses the inverse-covariance metric. Comparing path and chord in the same constant metric removes this scale effect; interpreting spatially varying metrics still requires a separate curvature argument.","primaryUrl":"https://arxiv.org/html/2609.34571v1","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152701","title":"SAFE-SVD: Sensitivity-Aware Fidelity-Enforcing SVD for Physics Foundation Models","authors":["Chengjie Hong","Feixiang He","Yiheng Zeng","Lulu Kang","He Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SAFE-SVD: Sensitivity-Aware Fidelity-Enforcing SVD for Physics Foundation Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17985","noteUrl":"/review?paper=review-neurips-2026-152701","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152702","title":"Reaching a Consensus in Predictive Loops","authors":["Jiduan Wu","Rediet Abebe","Celestine Mendler-Dünner"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Spillover cannot have a uniform positive lower bound when its source is unanchored","plainExplanation":"The homogeneous consensus theorem remains valid after replacing the appendix’s diagonal-conjugation display with the proper resolvent limit. The intervention closed form also has a local denominator error, while its amplification conclusion holds with the corrected expression when additional peers are present. Two substantive qualifications remain: a positive spillover bound needs an anchored source, and the finite-step spectral decomposition describes perfect prediction rather than the steering policy cited in its statement.","primaryUrl":"https://arxiv.org/html/2603.12137","noteUrl":"/review?paper=review-neurips-2026-152702","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152703","title":"Learning Theory of Transformers: Local-to-Global Approximation via Softmax Partition of Unity","authors":["Zhongjie Shi","Wenjing Liao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A nonempty approximation domain needs at least one center","plainExplanation":"A nonempty approximation domain needs at least one center. The supplied softmax scale can be negative. The parameter bound reciprocates a denominator that need not be positive.","primaryUrl":"https://arxiv.org/html/2605.08811","noteUrl":"/review?paper=review-neurips-2026-152703","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152704","title":"Rethinking Vector Field Learning for Generative Segmentation","authors":["Chaoyang Wang","Yaobo Liang","Boci Peng","Fan Duan","Yunhai Tong"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The target centroid need not be a stationary point of the potential","plainExplanation":"The target centroid need not be a stationary point of the potential.","primaryUrl":"https://arxiv.org/html/2603.19218v1","noteUrl":"/review?paper=review-neurips-2026-152704","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152705","title":"FlowSteer: Towards Agents Designing Agentic Workflows via Reinforced Progressive Canvas Editing","authors":["Mingda Zhang","Wenjin Liu","Tiesunlong Shen","Qika Lin","Rui Mao","Erik Cambria","Xiaoying Tang","Haoran Luo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Three strict policy-improvement claims need additional hypotheses","plainExplanation":"Zero–one Bayes risk is concave but not strictly concave. Informative signals can leave the same class optimal at every posterior.","primaryUrl":"https://arxiv.org/html/2602.01664","noteUrl":"/review?paper=review-neurips-2026-152705","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152707","title":"Value-Aware Stochastic KV Cache Eviction for Reasoning Models","authors":["Ting-Yun Chang","Harvey Yiyun Fu","Deqing Fu","Chenghao Yang","Jesse Thomason","Robin Jia"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Value-Aware Stochastic KV Cache Eviction for Reasoning Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.03928","noteUrl":"/review?paper=review-neurips-2026-152707","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152709","title":"Brenier Meets Adversarial Training: Optimal Transport Geometry for Robust Learning","authors":["alireza abdollahpour","Ehsan Sharifian","Buse Şen","Marco Cuturi","Daniel Kuhn"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Brenier Meets Adversarial Training: Optimal Transport Geometry for Robust Learning","plainExplanation":"An everywhere-positive Gaussian density is not uniformly bounded below on this expanding center set.","primaryUrl":"https://arxiv.org/html/2609.31363","noteUrl":"/review?paper=review-neurips-2026-152709","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152711","title":"Efficient One-Step Diffusion Restoration Model with Compact Token Compression and Linear Attention","authors":["Bingtian Qiao","Yue Shi","Yingjie Zhou","Yong Guo","Guangtao Zhai","Jiezhang Cao"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The KL value and calibration weight need separate corrections","plainExplanation":"The two definitions assign different timestep weights to the curvature proxy.","primaryUrl":"https://arxiv.org/html/2605.23451","noteUrl":"/review?paper=review-neurips-2026-152711","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152712","title":"Task Vector Geometry Underlies Dual Modes of Task Inference in Transformers","authors":["Hao Yan","Haolin Yang","Yiqiao Zhong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Hidden-state projection is not Euclidean coefficient projection","plainExplanation":"The method identifies projection onto a convex hull of hidden anchors with a two-step Euclidean projection of their coefficients. The state residual induces a Gram-weighted coefficient metric.","primaryUrl":"https://arxiv.org/html/2605.03780","noteUrl":"/review?paper=review-neurips-2026-152712","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152713","title":"Persona Vectors: Monitoring and Controlling Character Traits in Language Models","authors":["Runjin Chen","Andy Arditi","Henry Sleight","Owain Evans","Jack Lindsey"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Persona Vectors: Monitoring and Controlling Character Traits in Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2507.21509","noteUrl":"/review?paper=review-neurips-2026-152713","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152714","title":"Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping","authors":["Hsun-Yu Kuo","El Mahdi Chayti","Patrik Reizinger","Wieland Brendel","Martin Jaggi"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.29983","noteUrl":"/review?paper=review-neurips-2026-152714","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152715","title":"Bayesian Test-Time Inference of Task-Aligned Similarity from Weak Interactive Feedback","authors":["Weida Liang","Kenji Kawaguchi"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Bayesian Test-Time Inference of Task-Aligned Similarity from Weak Interactive Feedback","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.18656","noteUrl":"/review?paper=review-neurips-2026-152715","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152716","title":"TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond","authors":["Shang-Ling Hsu","Mark Tenzer","Cyrus Shahabi","Khurram Shafique"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Truth-centered window mass is not a proper calibration score","plainExplanation":"Truth-centered window mass is not a proper calibration score.","primaryUrl":"https://arxiv.org/html/2605.06906","noteUrl":"/review?paper=review-neurips-2026-152716","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152717","title":"Active Learning for Gaussian Process Regression Under Self-Induced Boltzmann Weights","authors":["Jixiang Qing","Henry Moss","Matthias Sachs"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A fixed-round confidence radius does not cover every Thompson draw","plainExplanation":"The fixed-round posterior bound does not automatically cover every Thompson draw, and four kernel derivatives do not guarantee polylogarithmic information gain. The average-risk proof must also average its dataset-dependent posterior-variance term. These findings concern the stated proof steps and rate simplifications, without establishing a loss lower bound against the SID algorithm. The extension to general metrics additionally requires the dominating function to vanish with uncertainty; the separately derived squared-error calculation is preserved.","primaryUrl":"https://arxiv.org/html/2605.10654","noteUrl":"/review?paper=review-neurips-2026-152717","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152718","title":"D-PACE: Dynamic Position-Aware Cross-Entropy for Parallel Speculative Drafting","authors":["Tianyu Wu","Yu Yao","Zhenting Qi","Han Zheng","Chengxi Zhang","Zhuohan Wang","Haoran Ma","Zichun Liao","Hima Lakkaraju","Ju Li","Yilun Du"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"D-PACE: Dynamic Position-Aware Cross-Entropy for Parallel Speculative Drafting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18810","noteUrl":"/review?paper=review-neurips-2026-152718","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152722","title":"Itô maps for any-step SDEs","authors":["Zhengkai Pan","Peter Potaptchik","Wenxi Yao","Michael Albergo","Jakiw Pidstrigach"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The default BEL weight needs a normalization factor","plainExplanation":"The displayed weight integrates to two thirds, so it does not satisfy the theorem’s stated normalization.","primaryUrl":"https://arxiv.org/html/2606.11156","noteUrl":"/review?paper=review-neurips-2026-152722","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152723","title":"UniGS: Unified Geometry-Aware Gaussian Splatting for Multimodal Rendering","authors":["Yusen XIE","Zhenmin Huang","Jianhao Jiao","Dimitrios Kanoulas","Jun Ma"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The geometry, backward rules and loss definition need explicit corrections","plainExplanation":"A world-to-camera transformation rotates the displacement from the camera center, not the world position itself.","primaryUrl":"https://arxiv.org/html/2510.12174","noteUrl":"/review?paper=review-neurips-2026-152723","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error","missing_assumption","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152726","title":"TrackFish3D: Self-Supervised 3D Tracking of Schooling Fish from Multi-view Videos","authors":["Patt Phurtivilai","Zhiyang Dou","Yifan Wu","Kinfung Chu","Yuan Liu","Lei Yang","Wenping Wang","Taku Komura"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"TrackFish3D: Self-Supervised 3D Tracking of Schooling Fish from Multi-view Videos","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38347","noteUrl":"/review?paper=review-neurips-2026-152726","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152727","title":"Empowering Time Series Analysis with Large-Scale Multimodal Pretraining","authors":["Peng Chen","Siyuan Wang","hu","Xingjian Wu","Yang Shu","Zhongwen Rao","Meng Wang","Yijie Li","Bin Yang","Chenjuan Guo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The frequency masks are swapped in the decomposition equation","plainExplanation":"The named frequency components disagree with their explicitly defined masks.","primaryUrl":"https://arxiv.org/html/2602.05646","noteUrl":"/review?paper=review-neurips-2026-152727","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152729","title":"Trustworthy Retrosynthesis: Eliminating Hallucinations with a Diverse Ensemble of Reaction Scorers","authors":["Michał Sadowski","Tadija Radusinović","Maria Wyrzykowska","Lukasz Sztukiewicz","Jan Rzymkowski","Piotr Kozakowski","Mikolaj Sacha","Ruard van Workum","Stanislaw Jastrzebski"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Fractional powers of log scores need a real-domain transformation","plainExplanation":"The reaction plausibility score multiplies log-probability components raised to default fractional weights, including exponent \\(2.5\\). Mean log token probability is negative whenever a component contains a nonunit probability.","primaryUrl":"https://arxiv.org/html/2510.10645","noteUrl":"/review?paper=review-neurips-2026-152729","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152730","title":"Learning Energy-Based Models from Stochastic Interpolants using Spatiotemporal Differences","authors":["Hanlin Yu","RuiKang OuYang","Partha Kaushik","Arto Klami","Michael Gutmann","Omar Chehab"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Gaussian transition display drops its target-time factors","plainExplanation":"For linear interpolation coefficients \\(\\alpha_t=t\\), \\(\\beta_t=1-t\\), Equation (20) gives a transition variance that omits dependence on \\(t^{\\prime}\\). The printed variance reduces to \\((2t-1)/t^2\\), negative for \\(0<t<1/2\\), and fails the coincident-time limit.","primaryUrl":"https://arxiv.org/html/2605.26850","noteUrl":"/review?paper=review-neurips-2026-152730","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152731","title":"Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm","authors":["Chuang Yu","Jinmiao Zhao","Yunpeng Liu","Yaokun Li","Xiujun Shu","Yuanhao Feng","Bo Wang","Xiangyu Yue"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.05511","noteUrl":"/review?paper=review-neurips-2026-152731","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152732","title":"ARGUS: Stacked Multi-View Identity Mosaic Injection for Subject-Preserving Video Generation","authors":["Zijie Meng","Jiwen Liu","Yufei Liu","Chengzhuo Tong","Xiaoqiang Liu","Yuanxing Zhang","Yulong Xu","Pengfei Wan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Interpolating identity features is not equivalent to gating attention logits","plainExplanation":"A logit gate changes the relative attention weights and their normalization. Interpolating the input features changes their projected values and keys; it does not generally produce that same normalized operator.","primaryUrl":"https://arxiv.org/html/2606.11670","noteUrl":"/review?paper=review-neurips-2026-152732","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152733","title":"Representation Forcing for Bottleneck-Free Unified Multimodal Models","authors":["Yuqing Wang","Zhijie Lin","Ceyuan Yang","Yang Zhao","Fei Xiao","Hao He","Qi Zhao","Zihan Ding","Fu-Yun Wang","Shuai Wang","Youliang Zhang","Haoqi Fan","Xihui Liu"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Softmax of probabilities is not Sinkhorn column scaling","plainExplanation":"Algorithm 1 applies a row softmax and then a column softmax, identifying the second operation with Sinkhorn–Knopp normalization. Column scaling divides all column entries by the same sum and preserves their ratio \\(a/b\\).","primaryUrl":"https://arxiv.org/html/2605.31604","noteUrl":"/review?paper=review-neurips-2026-152733","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152734","title":"AdaOcc: Adaptive 3D Occupancy Prediction for Embodied Tasks","authors":["Jinglong Wang","Yunjie Wang","Zhiyang Zhang","Jiawei He","Ye Yuan","Bo Qiu","Jing Zhang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"AdaOcc: Adaptive 3D Occupancy Prediction for Embodied Tasks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38864","noteUrl":"/review?paper=review-neurips-2026-152734","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152735","title":"Unlocking LLM Creativity in Science through Analogical Reasoning","authors":["Andrew Shen","Shaul Druckmann","James Zou"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Unlocking LLM Creativity in Science through Analogical Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11258","noteUrl":"/review?paper=review-neurips-2026-152735","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152736","title":"VTV-FM: Flow Matching through Variational Terminal-Velocity Closure","authors":["Haoyang Jiang","Yuheng Li","Di Yang","Yanhai Xiong","Haipeng Chen","Yi He"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"VTV-FM: Flow Matching through Variational Terminal-Velocity Closure","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.00785","noteUrl":"/review?paper=review-neurips-2026-152736","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152737","title":"MolHIT: Advancing Molecular-Graph Generation with Hierarchical Discrete Diffusion Models","authors":["Hojung Jung","Rodrigo Hormazabal","Jaehyeong Jo","Youngrok Park","Kyunggeun Roh","Se-Young Yun","Sehui Han","Dae-Woong Jeong"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"A stochastic-projection KL is not a log inner product plus a constant","plainExplanation":"Idempotent level maps need not be compatible across levels. Pointwise schedule ordering does not imply valid conditional survival ordering. The intermediate-state NELBO coefficient has the wrong first-order rate. A stochastic-projection KL is not a log inner product plus a constant. The printed uniform edge kernel is missing its dimension normalization.","primaryUrl":"https://arxiv.org/html/2602.17602","noteUrl":"/review?paper=review-neurips-2026-152737","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","lemma_error","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-152739","title":"Full-Duplex Speech-Motion Model for Dyadic Interaction","authors":["Koki Nagano","Hongyu Liu","Wookie Park","Tianye Li","Amrita Mazumdar","Christian Jacobsen","Shengze Wang","Michael Stengel","Ka Chun Cheung","Simon See","Shalini De Mello"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Full-Duplex Speech-Motion Model for Dyadic Interaction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03874","noteUrl":"/review?paper=review-neurips-2026-152739","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152743","title":"On the Memorization of Consistency Distillation for Diffusion Models","authors":["Bingqing Jiang","Difan Zou"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"On the Memorization of Consistency Distillation for Diffusion Models","plainExplanation":"The stated curvature formula can have negative normalized trace although the true curvature is positive semidefinite. Its scalar shift also uses the wrong flow increment, and its residual calculation needs centering.","primaryUrl":"https://arxiv.org/html/2604.23552","noteUrl":"/review?paper=review-neurips-2026-152743","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-152744","title":"Same Signal, Opposite Meaning: Direction-Informed Adaptive Learning for LLM Agents","authors":["Ziming Li","Jiatan Huang","Xiaoguang Guo","Guiling \"Grace\" Wang","Chuxu Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An always-off gate satisfies both weak success-rate requirements","plainExplanation":"A gate that never invokes the optimizer equals the base policy in every setting, so it satisfies both weak success-rate requirements in Appendix B.1. Identical uncertainty signals with opposite computation values instead show that a signal-only gate cannot choose the state-aware optimal action in both cases. That narrower decision statement needs explicit state and comparison assumptions. The paper’s approximate empirical scaling is not treated as an exact mathematical law.","primaryUrl":"https://arxiv.org/html/2605.06908","noteUrl":"/review?paper=review-neurips-2026-152744","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152745","title":"GauS: Differentiable Scheduling Optimization via Gaussian Reparameterization","authors":["Yaohui Cai","Vesal Bakhtazad","CUNXI YU","Zhiru Zhang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The expected violation sums omit equality and zero-index events","plainExplanation":"The relaxation declares unit latency, so a forward edge requires \\(s_j\\geq s_i+1\\), and a recurrence requires \\(s_i+kII\\geq s_j+1\\). Forward invalidity is \\(s_j\\leq s_i\\), not only strict earlier scheduling.","primaryUrl":"https://arxiv.org/html/2602.20427","noteUrl":"/review?paper=review-neurips-2026-152745","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152747","title":"Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs","authors":["Vincent-Daniel (Juyoung) Yun","Junhyuk Jo","Sai Praneeth Karimireddy","Sunwoo Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A positive ridge parameter changes the exact least-squares optimum","plainExplanation":"A positive ridge parameter changes the exact least-squares optimum. Minimum correction norm does not imply minimum final-weight norm.","primaryUrl":"https://arxiv.org/html/2605.15491","noteUrl":"/review?paper=review-neurips-2026-152747","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152749","title":"The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment","authors":["Jiachen Zhao","Zhengxuan Wu","Aryaman Arora","Yiyou Sun","David Bau","Weiyan Shi"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06667","noteUrl":"/review?paper=review-neurips-2026-152749","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152750","title":"Norm Anchors Make Model Edits Last","authors":["Mingda Liu","Zhenghan Zhu","Ze‘an Miao","Katsuki Fujisawa"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A fixed target norm does not by itself bound an unlimited edit sequence","plainExplanation":"Norm anchoring controls the new value vector, but it does not guarantee the uniform restoring drift needed for an infinite-horizon bound. A finite-range fit also gives no bound on the recurrence's error at later edits.","primaryUrl":"https://arxiv.org/html/2602.02543","noteUrl":"/review?paper=review-neurips-2026-152750","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152752","title":"Bellman Residual Minimization for Control: Geometry, Stationarity, and Convergence","authors":["Donghwan Lee","Hyukjun Yang"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Bellman Residual Minimization for Control: Geometry, Stationarity, and Convergence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.18840","noteUrl":"/review?paper=review-neurips-2026-152752","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152753","title":"Diff-CA: Separating Common and Salient Factors with Diffusion Models","authors":["Michaël Soumm","Alexandre Fournier Montgieux","Yunlong HE","Pietro Gori","Alasdair Newson"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Diff-CA: Separating Common and Salient Factors with Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06120","noteUrl":"/review?paper=review-neurips-2026-152753","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152756","title":"AMUSE: Anytime Muon with Stable Gradient Evaluation","authors":["Jueun Kim","Baekrok Shin","Jihun Yun","Beomhan Baek","Minhak Song","Chulhee Yun"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"AMUSE: Anytime Muon with Stable Gradient Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22432","noteUrl":"/review?paper=review-neurips-2026-152756","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152759","title":"RoSA: Rotational Sparse Adaptation for Memory-Efficient Fine-Tuning","authors":["Muhammad Azeem Lodhi","chao zhou","Rebekka Burkholz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RoSA: Rotational Sparse Adaptation for Memory-Efficient Fine-Tuning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06243","noteUrl":"/review?paper=review-neurips-2026-152759","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152760","title":"Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks","authors":["Kevin Lu","Jannik Brinkmann","Stefan T Huber","Aaron Mueller","Yonatan Belinkov","David Bau","Chris Wendler"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"A projection coefficient does not identify a patched representation endpoint","plainExplanation":"Equation (3) projects a patched representation onto donor-minus-target direction \\(d\\ne0\\) and interprets coefficients zero and one as full endpoint equality. Write \\(z_{\\rm patched}-z_{\\rm target}=\\alpha d+r\\), with \\(r\\perp d\\).","primaryUrl":"https://arxiv.org/html/2602.06020","noteUrl":"/review?paper=review-neurips-2026-152760","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152761","title":"3D Skew Normal Splatting","authors":["Xiangru Wu","Ke Fan","Yanwei Fu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"3D Skew Normal Splatting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15010","noteUrl":"/review?paper=review-neurips-2026-152761","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152762","title":"Fast Learning Rates for Physics-Informed Kernel Methods","authors":["Luc Brogat-Motte","Joachim Bona-Pellissier","Giacomo Meanti","Lorenzo Rosasco"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Independent covariates do not supply independent observation noise","plainExplanation":"Noise can have the required conditional mean at every covariate while being perfectly correlated across the sample. Averaging such observations does not reduce the noise, so the stated assumptions do not imply the shrinking estimation error.","primaryUrl":"https://arxiv.org/html/2609.18901","noteUrl":"/review?paper=review-neurips-2026-152762","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152764","title":"Invaria: Learning Scale and Density Invariance in Point Clouds via Next-Resolution Prediction","authors":["Chun-Peng Chang","Shaoxiang Wang","Alain Pagani","Dariu Gavrila","Holger Caesar"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Invaria: Learning Scale and Density Invariance in Point Clouds via Next-Resolution Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15923","noteUrl":"/review?paper=review-neurips-2026-152764","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152767","title":"EntropyCache: Decoded Token Entropy Guided KV Caching for Diffusion Language Models","authors":["Minsoo Cheong","Donghyun Son","Woosang Lim","Sungjoo Yoo"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Timestamp ties can exceed the selected-index budget","plainExplanation":"The method selects exactly \\(k\\) recent indices but then replaces their identities by an inclusive threshold at the selected oldest timestamp. Ties enlarge the final set.","primaryUrl":"https://arxiv.org/html/2603.18489","noteUrl":"/review?paper=review-neurips-2026-152767","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152768","title":"RigidFormer: Learning Rigid Dynamics using Transformers","authors":["Zhiyang Dou","Minghao Guo","Haixu Wu","Doug Roble","Tuur Stuyck","Wojciech Matusik"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Pooling linear anchor phases retains only the centroid","plainExplanation":"The anchor descriptor first averages a coordinate-linear multiscale phase map, then applies trigonometric rotations and attributes extent information to the result. Linearity gives \\(K^{-1}\\sum_k\\psi(x_k)=\\psi(K^{-1}\\sum_kx_k)\\).","primaryUrl":"https://arxiv.org/html/2605.09196","noteUrl":"/review?paper=review-neurips-2026-152768","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152769","title":"UGO: Unified Architecture for General Multi-Object Tracking by Segmentation","authors":["Jer Pelhan","Alan Lukezic","Matej Kristan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The fixed-auxiliary labeling objective remains pixel-coupled","plainExplanation":"The auxiliary energy defines counts and scale factors as functions of candidate labeling \\(Y\\), then treats the labeling step as a separable unary minimization at fixed auxiliaries. Fixing the auxiliaries does not freeze those labeling-dependent quantities.","primaryUrl":"https://arxiv.org/html/2609.37339","noteUrl":"/review?paper=review-neurips-2026-152769","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152773","title":"A Theory on Flow Matching with Neural Networks","authors":["Yihan He","Qishuo Yin","Yuan Cao","Jianqing Fan","Han Liu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The printed two-term balance has a positive support-radius power","plainExplanation":"The sampling proof needs a positive density lower bound on a truncation ball, while the stated target/model assumptions do not supply that source-law condition. For \\(N(0,I_d)\\) on radius \\(\\zeta\\), the minimum density is \\((2\\pi)^{-d/2}e^{-\\zeta^2/2}\\), not the larger printed value with exponent \\(-\\zeta^2/(2d)\\).","primaryUrl":"https://arxiv.org/html/2606.10089","noteUrl":"/review?paper=review-neurips-2026-152773","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152774","title":"Improving Function Space Flow Matching with Kernel Optimal Transport","authors":["Fred Xu","Thomas Markovich","Barbora Barancikova","Yizhou Sun"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Improving Function Space Flow Matching with Kernel Optimal Transport","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38049","noteUrl":"/review?paper=review-neurips-2026-152774","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152775","title":"Exposing and Mitigating Temporal Attack in Deepfake Video Detection","authors":["Zheyuan Gu","Minghao Shao","Zhen Wang","Keyu Mao","Ailiang Lin","Shijie Zhang","hao jiang","mingyou liang","Mingkun Xu","Yusong Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Exposing and Mitigating Temporal Attack in Deepfake Video Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07398","noteUrl":"/review?paper=review-neurips-2026-152775","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152776","title":"From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation","authors":["Bora Kargi","David Salinas"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation","plainExplanation":"No issue was found in the checked claims. The historical ordering-only explanation was false for Elo distances, but it is no longer the assertion made by the current revision. This finding is resolved in arXiv v3.","primaryUrl":"https://arxiv.org/html/2606.13221","noteUrl":"/review?paper=review-neurips-2026-152776","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152777","title":"RAM-Net: Linear-Time Sequence Modeling with Sparsely Addressable State","authors":["Kaicheng Xiao","Haotian Li","Liran Dong","Guoliang Xing"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RAM-Net: Linear-Time Sequence Modeling with Sparsely Addressable State","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.11958","noteUrl":"/review?paper=review-neurips-2026-152777","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152779","title":"Entropy Guided Dynamic Patch Segmentation for Time Series Transformers","authors":["Sachith Abeywickrama","Emadeldeen Eldele","Min Wu","Xiaoli Li","Chau Yuen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"An entropy spike does not force a strict decrease in patch mean","plainExplanation":"Splitting at an entropy spike can leave both subpatch means unchanged; strict improvement requires unequal subpatch means. The weak weighted-average inequality and the cross-entropy lower bound remain valid. The dense fusion decoder also contributes a sequence-length times patch-count attention term omitted from the general complexity expression, independently of whether entropy inference is cached.","primaryUrl":"https://arxiv.org/html/2509.26157","noteUrl":"/review?paper=review-neurips-2026-152779","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement","incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152780","title":"Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry","authors":["Ningkang Peng","Xuanming Chen","Yanhui Gu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17799","noteUrl":"/review?paper=review-neurips-2026-152780","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152783","title":"Is Backpropagation Optimal? When Synthetic Gradients Improve Sample Efficiency","authors":["Yibo Jacky Zhang","Zeyu Tang","Sanmi Koyejo"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Is Backpropagation Optimal? When Synthetic Gradients Improve Sample Efficiency","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27946","noteUrl":"/review?paper=review-neurips-2026-152783","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152787","title":"GIST: Gauge-Invariant Spectral Transformers for Scalable Graph Neural Operators","authors":["Mattia Rigotti","Nicholas Thumiger","Thomas Frick"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Trace-class tails do not preserve a fixed-mode convergence rate","plainExplanation":"Summability makes a tail small as the cutoff grows; it does not specify how quickly. The proof supplies neither a quantitative tail bound at the sample-dependent cutoff nor uniform control of the mode-dependent error constants.","primaryUrl":"https://arxiv.org/html/2603.16849","noteUrl":"/review?paper=review-neurips-2026-152787","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152788","title":"The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric","authors":["Sheng-Yu Wang","Yotam Nitzan","Aaron Hertzmann","Jun-Yan Zhu","Eli Shechtman","Alexei Efros","Richard Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.18237","noteUrl":"/review?paper=review-neurips-2026-152788","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152793","title":"Latent Generative Solvers for Generalizable Long-Term Physics Simulation","authors":["Zituo Chen","Sili Deng"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Latent Generative Solvers for Generalizable Long-Term Physics Simulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.11229","noteUrl":"/review?paper=review-neurips-2026-152793","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152794","title":"Empirical Bayes Rebiasing","authors":["Wanyi Ling","Sida Li","Junming Guan","Nikolaos Ignatiadis"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Empirical Bayes Rebiasing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08069","noteUrl":"/review?paper=review-neurips-2026-152794","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152795","title":"From Reasoning Chains to Verifiable Subproblems: Curriculum Reinforcement Learning Enables Credit Assignment for LLM Reasoning","authors":["Xitai Jiang","Wenze Lin","Zihan Tang","Yang Yue","Shenzhi Wang","Gao Huang"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"From Reasoning Chains to Verifiable Subproblems: Curriculum Reinforcement Learning Enables Credit Assignment for LLM Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22074","noteUrl":"/review?paper=review-neurips-2026-152795","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152798","title":"MACRO: Training-free Multi-plane Attention for Closeup Render Optimization","authors":["Nitzan Hodos","Roy Amoyal","Lior Fritz","Ianir Ideses","Sagie Benaim","Netalee Efrat"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"MACRO: Training-free Multi-plane Attention for Closeup Render Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.03875","noteUrl":"/review?paper=review-neurips-2026-152798","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152799","title":"Score-based Variational Inference via Quantum Maximally Mixed States","authors":["Yuchen CONG","Zerui Tao","Chao Li","Zhe Sun","Qibin Zhao"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Score-based Variational Inference via Quantum Maximally Mixed States","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39164","noteUrl":"/review?paper=review-neurips-2026-152799","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152800","title":"Mesh-Free Convolution: Learned Spectral Attenuation and Transport","authors":["Niloufar Zakariaei","Manav Manav","Eldad Haber"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mesh-Free Convolution: Learned Spectral Attenuation and Transport","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.15931","noteUrl":"/review?paper=review-neurips-2026-152800","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152801","title":"Loop-Free Inverse Reinforcement Learning via Sequential Value Recovery","authors":["Yang Chen","Yitan Zhang","Michael Witbrock","Shuyue Hu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Loop-Free Inverse Reinforcement Learning via Sequential Value Recovery","plainExplanation":"No issue was found in the checked claims. An up-to-offset recovery statement does not assert that the regularizer identifies the true offset. A universal guarantee of pointwise calibration is not the mathematical assertion made by Proposition 1 or Equation (12).","primaryUrl":"https://arxiv.org/html/2609.38955","noteUrl":"/review?paper=review-neurips-2026-152801","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152804","title":"ReFPO: Reflow Regularization for Flow Matching Policy Gradients","authors":["Ge Wang","Yibo Peng","Fan Feng","Shenhao Yan","Chengsi Yao","Jiahao Yang","Honghao Cai","Yiming Zhao","Xi Li","Jinke Ren","Shuguang Cui","Yatong Han","Zhen Li"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The Reflow interpretation needs a seed–endpoint convention","plainExplanation":"On-policy action sampling alone specifies the action marginal, not its coupling with the training noise. The exact recursive-Reflow interpretation therefore needs a joint sampling convention beyond the displayed marginal loss.","primaryUrl":"https://arxiv.org/html/2606.21086","noteUrl":"/review?paper=review-neurips-2026-152804","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["algorithm_specification"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152805","title":"Minimax Optimal Estimation of Transport-Growth Pairs in Unbalanced Optimal Transport","authors":["Donlapark Ponnoprat","Noboru Isobe","Masaaki Imaizumi"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Unit growth collapses the proposed balanced hard subclass","plainExplanation":"The experiment supplies exact distribution masses, but the packing argument controls only normalized-sample KL while perturbing the target mass. Distinct deterministic mass coordinates make full observation laws disjoint and KL infinite.","primaryUrl":"https://arxiv.org/html/2605.08705","noteUrl":"/review?paper=review-neurips-2026-152805","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152806","title":"DCVD: Dual-Channel Cross-Modal Fusion for Joint Vulnerability Detection and Localization","authors":["Wenxin Tang","Wenbin Li","Junliang Liu","Jingyu Xiao","Xi Xiao","Mingzhe liu","Jinlong Yang","Xuan Liu","Yuehe Ma","Wang Luo","Qing Li","Lei Wang","Peng Xiangli"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DCVD: Dual-Channel Cross-Modal Fusion for Joint Vulnerability Detection and Localization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11015","noteUrl":"/review?paper=review-neurips-2026-152806","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152811","title":"Optimal Rates for Differentially Private Hypothesis Testing with E-values","authors":["Ben Jacobsen","Tomás González Lara","Gavin R Brown","Kassem Fawaz","Aaditya Ramdas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Optimal Rates for Differentially Private Hypothesis Testing with E-values","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28952","noteUrl":"/review?paper=review-neurips-2026-152811","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152812","title":"LaCache: Robust Semantic Caching for LLM Serving","authors":["Jiacheng Liang","Yuhui Wang","Tanqiu Jiang","Ting Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LaCache: Robust Semantic Caching for LLM Serving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.01718","noteUrl":"/review?paper=review-neurips-2026-152812","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152816","title":"FOGO: Forgetting-aware Orthogonalization Optimizer","authors":["Toan Nguyen","Yang Liu","Trung Le","Celso de Melo","Flora Salim"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"FOGO: Forgetting-aware Orthogonalization Optimizer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.10406","noteUrl":"/review?paper=review-neurips-2026-152816","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152817","title":"Test time training enhances in-context learning of nonlinear functions","authors":["Kento Kuwataka","Taiji Suzuki"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The alignment proof uses the opposite projected-update sign","plainExplanation":"The displayed replacement is not the descent-and-normalization update. Its tangent component has the opposite sign, so the subsequent leading alignment term does not follow from the printed algorithm.","primaryUrl":"https://arxiv.org/html/2509.25741","noteUrl":"/review?paper=review-neurips-2026-152817","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152819","title":"Distributionally Robust Multi-Task Reinforcement Learning via Adaptive Task Sampling","authors":["Nicholas Corrado","Wenyuan Huang","Josiah Hanna"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The policy player should minimize positive reference gaps","plainExplanation":"The pseudocode maximizes \\(g_i=\\max(J_i^{\\mathrm{ref}}-J_i,0)\\), while the min-max theory assumes a gap-minimizing low-regret policy player. For an unsaturated gap, increasing return decreases it with derivative minus one.","primaryUrl":"https://arxiv.org/html/2605.14350","noteUrl":"/review?paper=review-neurips-2026-152819","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152825","title":"Bridging Textual Profiles and Latent User Embeddings for Personalization","authors":["Zhaoxuan Tan","Xiang Zhai","Yan Zhu","Meng Jiang","Mohamed Hammad"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Bridging Textual Profiles and Latent User Embeddings for Personalization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06981","noteUrl":"/review?paper=review-neurips-2026-152825","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152826","title":"DoAtlas-1: A Causal Compilation Paradigm for Clinical AI","authors":["Yulong Li","Jianxu Chen","Xiwei Liu","Chuanyue Suo","Rong Xia","Yichen Li","Xinlin Zhuang","Niranjana A Menon","Jionglong Su","Eran Segal","Imran Razzak"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Two marginal effect buckets do not identify a joint intervention","plainExplanation":"Pattern C.4 requires nonempty \\(X_1\\to Y\\) and \\(X_2\\to Y\\) evidence and returns a numerical joint effect without requiring no interaction. Different interactions can preserve both marginal intervention laws.","primaryUrl":"https://arxiv.org/html/2602.19158","noteUrl":"/review?paper=review-neurips-2026-152826","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152827","title":"Beyond Domains: Reusing Web Skills via Transferable Interaction Patterns","authors":["Shiqi He","Yue Cui","Feijie Wu","Xinyu Ma","Jiaheng Lu","Yaliang Li","Bolin Ding","Mosharaf Chowdhury"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Beyond Domains: Reusing Web Skills via Transferable Interaction Patterns","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.17645","noteUrl":"/review?paper=review-neurips-2026-152827","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152830","title":"Multilevel and Sequential Monte Carlo for Training-Free Diffusion Guidance","authors":["Aidan Gleich","Scott C Schmidler"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Multilevel and Sequential Monte Carlo for Training-Free Diffusion Guidance","plainExplanation":"Negative level corrections can overwhelm the base average, making the proposed ordinary resampling weights invalid.","primaryUrl":"https://arxiv.org/html/2601.21104","noteUrl":"/review?paper=review-neurips-2026-152830","collection":"neurips-2026","severity":"major","issueTypes":["mathematical_error","reporting_inconsistency"],"classifications":["major"]},{"slug":"review-neurips-2026-152831","title":"A Spectral Framework for Closed-Form Relative Density Estimation","authors":["Francis Bach"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Spectral Framework for Closed-Form Relative Density Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10668","noteUrl":"/review?paper=review-neurips-2026-152831","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152832","title":"ReST-RL: Reinforcing LLM Reasoning through Unified Self-Training and Value-Guided Search","authors":["Sining Zhoubian","Dan Zhang","Jie Tang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The prefix sampler needs an explicit terminal-endpoint convention","plainExplanation":"The probability law is normalized, but it assigns positive mass to the full, already terminated response. That endpoint is not a proper partial trajectory with a remaining suffix under the stated stopping convention.","primaryUrl":"https://arxiv.org/html/2508.19576","noteUrl":"/review?paper=review-neurips-2026-152832","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["algorithm_specification"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152834","title":"FlowMAS: Learning Multi-Agent Workflow Topology via Information-guided Generative Flow Network","authors":["Haitao Wang","Chenjing Liang","Haipeng Zhang","Jiawei Hu","SiCheng Wang","Songzhu Mei","Chenglu Wen","Siqi Shen","Cheng Wang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Intermediate rewards change the terminal sampling distribution","plainExplanation":"The flow objective adds nonnegative intermediate rewards to outgoing mass while claiming terminal probability proportional only to terminal reward \\(R\\). On a tree, the resulting terminal mass is proportional to \\(R(x)\\) plus rewards along its path, with normalization \\(\\sum_xR(x)+\\sum r\\).","primaryUrl":"https://arxiv.org/html/2609.37151","noteUrl":"/review?paper=review-neurips-2026-152834","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152835","title":"AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting","authors":["Yongjae Lee","Jingxing Li","Abhay Yadav","Rama Chellappa","Deliang Fan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The geometric enclosure and opacity-conservation claims need different update rules","plainExplanation":"The constructed ellipsoid can exclude points on an eligible child even though every coordinate lies within its chosen semiaxis.","primaryUrl":"https://arxiv.org/html/2605.06876","noteUrl":"/review?paper=review-neurips-2026-152835","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152836","title":"Language-Induced Priors for Domain Adaptation","authors":["Qiyuan Chen","Jiayu Zhou","Raed AL Kontar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A nonzero population log-ratio need not force data-dependent weights to zero or one","plainExplanation":"A Gaussian example follows the stated initialization and M-step yet leaves the relevance weight away from zero while its population log-ratio is negative at every sample size. The proof treats a vanishing, data-dependent log-ratio as a fixed nonzero margin. A relative concentration condition and diverging accumulated evidence repair the claim; sample splitting alone is insufficient. The separate target-sample consistency result and closed-form updates are not refuted.","primaryUrl":"https://arxiv.org/html/2605.14301v1","noteUrl":"/review?paper=review-neurips-2026-152836","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152840","title":"MoE-SpAc: Efficient MoE Inference Based on Speculative Activation Utility in Heterogeneous Edge Scenario","authors":["Shuhuai Li","Jianghao Lin","DongDong Ge","Yinyu Ye"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Two revised qualifications resolve earlier objections; five narrower issues remain","plainExplanation":"The balance objective alone does not control latency when the total device time depends on the threshold.","primaryUrl":"https://arxiv.org/html/2603.09983","noteUrl":"/review?paper=review-neurips-2026-152840","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","interpretation_error","invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152842","title":"Stein Kernelized Molecular Dynamics for Active Learning of Interatomic Potentials","authors":["Joanna Zou","Fraser Birks","Dallas Foster","Youssef Marzouk"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Stein Kernelized Molecular Dynamics for Active Learning of Interatomic Potentials","plainExplanation":"The SVGD self term is divided by \\(J\\), while the extra gradient in Equation (5) remains of order one. Its limiting field therefore need not have the Boltzmann density as a stationary solution.","primaryUrl":"https://arxiv.org/html/2606.04100","noteUrl":"/review?paper=review-neurips-2026-152842","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","invalid_reduction","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152847","title":"Graph-Regularized Sparse Autoencoders for LLM Safety Steering","authors":["Jehyeok Yeon","Federico Cinus","Yifan Wu","Luca Luceri"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Graph-Regularized Sparse Autoencoders for LLM Safety Steering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.06655","noteUrl":"/review?paper=review-neurips-2026-152847","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152848","title":"Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis","authors":["Keerthi Kaashyap","Dennis Anthony","Akshay Krishnan","Nhi Nguyen","Jeremy Collins","James Hays","Shreyas Kousik","Animesh Garg"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The decoder needs an explicit target-patch dependence","plainExplanation":"The displayed target transform depends on the view, not the target-patch index. Applying it to identical queries gives identical attention rows, so these operations alone cannot break the asserted spatial symmetry.","primaryUrl":"https://arxiv.org/html/2610.03717","noteUrl":"/review?paper=review-neurips-2026-152848","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152849","title":"Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards","authors":["Deokgyu Yoon","Hyungkyu Kang","Joongkyu Lee","Byeongchan Kim","Gyungin Shin","Sungrae Park","Min-hwan Oh"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The practical mask does not enforce uniform policy-ratio hypotheses","plainExplanation":"The theorem assumes supremum-state policy-distance and true likelihood-ratio bounds, while the implementation claim relies on an advantage-sign mask and clipped trace factors. The mask can admit high-distance tokens through its sign disjunction and never checks unsampled states.","primaryUrl":"https://arxiv.org/html/2605.20865","noteUrl":"/review?paper=review-neurips-2026-152849","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152850","title":"Not All Layers Need Tuning: Selective Layer Restoration Recovers Diversity","authors":["Bowen Zhang","Meiyi Wang","Harold Soh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Inclusive restoration intervals include single-layer choices","plainExplanation":"The restoration-interval count excludes singleton intervals even though the stated domain allows them. The exact count is \\(N(N+1)/2\\), not \\(N(N-1)/2\\). This is an exact combinatorial counting typo. Both counts are quadratic in depth, so the asymptotic search order is unchanged.","primaryUrl":"https://arxiv.org/html/2602.06665","noteUrl":"/review?paper=review-neurips-2026-152850","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152851","title":"iCATS: Fast Video Generation via Interaction-Aware Sparse Attention and Timestep-Adaptive Sparsity","authors":["Chengfeng Han","Baole Ai","Xianlu Bian","Jie Yao","Zilong Huang","Ang Wang","Dandan Ding"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"iCATS: Fast Video Generation via Interaction-Aware Sparse Attention and Timestep-Adaptive Sparsity","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.11302","noteUrl":"/review?paper=review-neurips-2026-152851","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152853","title":"Crowded in B-Space: Calibrating Shared Directions for LoRA Merging","authors":["Yixuan Tang","Yi Yang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Crowded in B-Space: Calibrating Shared Directions for LoRA Merging","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.16826","noteUrl":"/review?paper=review-neurips-2026-152853","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152854","title":"When and Why Adversarial Training Improves PINNs: A Neural Tangent Kernel Perspective","authors":["Yuandong Cao","Chi C So","Jun-Min Wang","He Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The weak-adversarial discriminator is undefined at zero residual","plainExplanation":"The logarithmic weak-adversarial critic remains undefined at zero despite the universal zero-anchor statement. The introductory objective still averages signed residuals, and the main Laplace boundary still conflicts with the correct appendix solution. The squared-energy analysis, the other critic families and the corrected appendix benchmark are credited separately.","primaryUrl":"https://arxiv.org/html/2605.15959","noteUrl":"/review?paper=review-neurips-2026-152854","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-152856","title":"Grounded-Exo2Ego: Structured Semantic Grounding for Robust Exocentric-to-Egocentric Video Generation","authors":["Shengze Wang","Michael Stengel","Tianye Li","Wookie Park","Amrita Mazumdar","Koki Nagano","Alex Trevithick","Shalini De Mello"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The rotary offset is assigned to different video streams","plainExplanation":"The main text aligns render and target positions while shifting exocentric positions; Appendix D.1 instead shifts render relative to target. Rotary dot products depend on relative phase.","primaryUrl":"https://arxiv.org/html/2608.20534","noteUrl":"/review?paper=review-neurips-2026-152856","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152859","title":"Adaptive LLM Routing for Multi-Turn Conversations with Continuously Evolving User Queries","authors":["Jiarui Zhang","Xiangyu Liu","Yong Hu","Chaoyue Niu","Hang Zeng","Shaojie Tang","Fan Wu","Guihai Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A negative cost denominator reverses expense preference","plainExplanation":"The ratio-reward ablation divides positive progress by a cost defined as the sum of negative expense rewards. For equal positive progress, the ratio is \\(-\\Delta/\\mathrm{expense}\\), increasing toward zero with larger expense.","primaryUrl":"https://arxiv.org/html/2604.12385","noteUrl":"/review?paper=review-neurips-2026-152859","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152862","title":"Statistical Convergence of Spherical First Hitting Diffusion Models","authors":["Simon Bienewald","Lukas Trottner"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Statistical Convergence of Spherical First Hitting Diffusion Models","plainExplanation":"The parameter printed in the theorem does not produce the advertised rate through the supplied error decomposition.","primaryUrl":"https://arxiv.org/html/2605.07625","noteUrl":"/review?paper=review-neurips-2026-152862","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152863","title":"VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling","authors":["Leigang Qu","Feng Cheng","Ziyan Yang","Bangbang Yang","Zhaoyang Huang","Wei Chow","Yicong Li","Wenjie Wang","Tat-Seng Chua","Yan Zeng"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Summed rotary phases retain inter-shot offsets","plainExplanation":"The explanation claims intra-position matching removes positional influence, but the final phase adds inter-shot and intra-video coordinates elementwise. Equal intra-video coordinates cancel only that component.","primaryUrl":"https://arxiv.org/html/2610.12104","noteUrl":"/review?paper=review-neurips-2026-152863","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152864","title":"Order Matters: Competition-Guided Query Ordering for RNN-Based Object Detection","authors":["Shengjian Wu","Li Sun","Yu Shangguan","Qingli Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Order Matters: Competition-Guided Query Ordering for RNN-Based Object Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.05191","noteUrl":"/review?paper=review-neurips-2026-152864","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152866","title":"Guiding Visual Autoregressive Models through Spectrum Weakening","authors":["Chaoyang Wang","Tianmeng Yang","Yunhai Tong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Deleting one conjugate Fourier partner need not lose real-signal information","plainExplanation":"The information-loss interpretation permits arbitrary Fourier masks on real inputs, without requiring conjugate symmetry. A one-partner mask can remain invertible after taking the real reconstruction.","primaryUrl":"https://arxiv.org/html/2511.22991","noteUrl":"/review?paper=review-neurips-2026-152866","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152867","title":"SpecLoR: Spectral Lookahead Rectification for Motion-Coherent Text-to-Video Generation","authors":["Xu Zhang","Yu Lu","Ruijie Quan","Zhaozheng CHEN","Bohan Wang","Yi Yang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The reset is not a no-op for general marginal flow trajectories","plainExplanation":"Even with no spectral amplitude change, the algorithm replaces \\(z\\) by \\((1-t)(z-tv)+t\\epsilon_{\\mathrm{init}}\\) and claims perfect baseline-dynamics preservation. The difference is \\(t[\\epsilon_{\\mathrm{init}}-z-(1-t)v]\\), not identically zero.","primaryUrl":"https://arxiv.org/html/2606.11969","noteUrl":"/review?paper=review-neurips-2026-152867","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152869","title":"P-EAGLE: Parallel-Drafting EAGLE with Scalable Training","authors":["Xin Huang","Mude Hui","Jaime Campos Salas","Yue Sun","Nathan Pemberton","Ashish Khetan","Xiang song","George Karypis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A joint generic-position argument does not cover every fixed key","plainExplanation":"The fixed-key recovery corollary extends an almost-everywhere query-key collision statement to all fixed keys. For allowed key zero, every rotated key is zero and all queries give identical positional scores.","primaryUrl":"https://arxiv.org/html/2602.01469","noteUrl":"/review?paper=review-neurips-2026-152869","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152871","title":"UniRank: Unified List-wise Reranking via Confidence-Ordered Denoising","authors":["Pengyue Jia","HailanYang","Shuchang Liu","Xiaobei Wang","Wanyu Wang","Xiang Li","Yongqi Liu","Kaiqiao Zhan","Kun Gai","Xiangyu Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Offline masked denoising is not generally an on-policy utility gradient","plainExplanation":"The target objective averages utility over generated slates, while Equation (5) samples logged ground-truth slates and differentiates masked token likelihoods. The sampling laws and score factors differ.","primaryUrl":"https://arxiv.org/html/2605.10527","noteUrl":"/review?paper=review-neurips-2026-152871","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152872","title":"Path Dependence under Adaptive AI Delegation","authors":["Lingxiao Huang","Nisheeth K. Vishnoi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Clamping a join does not preserve smooth pasting","plainExplanation":"The explicit separatrix approximation joins a power law \\(C\\theta^\\beta\\) to a saddle tangent line, with a clamp intended to handle all listed parameter settings. Value and derivative matching require \\(\\beta(p-ma)=m\\theta_l(1-\\beta)\\).","primaryUrl":"https://arxiv.org/html/2603.02950","noteUrl":"/review?paper=review-neurips-2026-152872","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152874","title":"TIDE: Every Layer Knows the Token Beneath the Context","authors":["AJAY JAISWAL","Lauren Hannah","Han-Byul Kim","Duc Hoang","Mehrdad Farajtabar","Minsik Cho"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A bias-free router cannot uniformly suppress arbitrary fixed memories","plainExplanation":"A bias-free router cannot uniformly suppress arbitrary fixed nonzero memories, although zeroing memory recovers the baseline exactly. The conditional sum of bankwise gradient lower bounds is valid, but it does not establish an exact factor \\(K\\) against an actual baseline or make gradient nondegeneracy follow from positive routing. The numerical rare-to-common gradient illustration also needs the exact batch-presence probability already supplied by the paper, giving about 0.00136 rather than 0.00001.","primaryUrl":"https://arxiv.org/html/2605.06216","noteUrl":"/review?paper=review-neurips-2026-152874","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152876","title":"SENSE: Semantic Neural Speech Synthesis from Brain Dynamics via Spatial Graph Encoding","authors":["Jisoo Park","Seonghak Lee","Hyojin Park","Junseok Kwon"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The CNN padding convention needs to match the branch dimensions","plainExplanation":"Under the usual symmetric padding convention, two such layers produce length \\(T+6\\), not \\(T\\). Rescaling and a scalar gate do not change that length.","primaryUrl":"https://arxiv.org/html/2609.37601","noteUrl":"/review?paper=review-neurips-2026-152876","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_ambiguity"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152877","title":"Re-examining Low Rank adaptation for private LLM fine-tuning","authors":["Ali Dadsetan","Frank Rudzicz"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Eigenvalue ordering does not bound a bilinear preconditioned gradient term","plainExplanation":"Eigenvalue ordering does not bound a bilinear preconditioned gradient term.","primaryUrl":"https://arxiv.org/html/2510.01137","noteUrl":"/review?paper=review-neurips-2026-152877","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-152878","title":"Spectral Feedback for Test-Time Alignment of Protein Diffusion Models","authors":["Shai Dickman","Mert Cemri","Landon Butler","Kannan Ramchandran"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The reverse jump generator needs the transposed forward rate","plainExplanation":"The convention uses column probabilities, with \\(Q(x,y)\\) the rate from \\(y\\) to \\(x\\), but the reverse formula keeps the wrong orientation and ratio. Short-time Bayes gives off-diagonal \\(\\bar Q(x,y)=Q(y,x)p(x)/p(y)\\), then diagonals must make outgoing column sums zero.","primaryUrl":"https://arxiv.org/html/2609.30456","noteUrl":"/review?paper=review-neurips-2026-152878","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152879","title":"OneVision-Encoder: Codec-Aligned Sparsity as a Foundational Principle for Multimodal Intelligence","authors":["feilong tang","Xiang An","Yunyao Yan","Yin Xie","Kaicheng Yang","Yifei Shen","Yuanhan Zhang","Shikun Feng","Chunyuan Li","Changrui Chen","Huajie Tan","Ming Hu","Manyuan Zhang","Bo Li","Ziyong Feng","Ziwei Liu","Zongyuan Ge","Jiankang Deng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"OneVision-Encoder: Codec-Aligned Sparsity as a Foundational Principle for Multimodal Intelligence","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.08683","noteUrl":"/review?paper=review-neurips-2026-152879","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152880","title":"ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection","authors":["Yijie Zhu","Rui Shao","Jie He","Wei Li","Bo Zhao","Yelin Wang","Xiaochen Yuan","Tao Tan","Miao Zhang","Xiaojiang Peng","Zitong YU"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"A shared codebook does not force paired modalities to share codes","plainExplanation":"The prediction objective uses a shared prototype codebook and decoder and is said to minimize paired cross-modal code distance. Paired codes can remain separated at zero prediction loss.","primaryUrl":"https://arxiv.org/html/2610.01741","noteUrl":"/review?paper=review-neurips-2026-152880","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152883","title":"Visual Enhanced Depth Scaling for Multimodal Latent Reasoning","authors":["Yudong Han","Yong Wang","Zaiquan Yang","Zhen Qu","Liyuan Pan","Xiangxiang Chu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The training-improvement proof drops a possible harm term","plainExplanation":"The training-gain proof deletes possible harm to selected tokens that are not improvable. The complexity argument treats row-stochastic attention as a Euclidean contraction and averages local sensitivities as though they bounded the full Jacobian. Those steps leave the claimed gain and complexity benefit unproved. The main and appendix theorem statements also use different width dependence, and the visual-replay equation selects from the visited set despite its coverage objective.","primaryUrl":"https://arxiv.org/html/2604.10500","noteUrl":"/review?paper=review-neurips-2026-152883","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","formula_error","incorrect_bound","invalid_inference","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152884","title":"RheoSampling: Resolving the One-Hot Dilemma in Stochastic Dynamic-Tree Speculative Decoding","authors":["Qiao Hu","Yepeng Weng","Bo Zhang","Takehisa Yairi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"RheoSampling: Resolving the One-Hot Dilemma in Stochastic Dynamic-Tree Speculative Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.21827","noteUrl":"/review?paper=review-neurips-2026-152884","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152888","title":"Variational Inference via Entropic Transport Descent","authors":["Vincent Pacelli","Akash Ratheesh Babu","Evangelos Theodorou"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The mutual-information optimal-value comparison has the wrong direction","plainExplanation":"Normalization prevents a uniform domination bound for the fixed-point image. The mutual-information optimal-value comparison has the wrong direction. Pooling proposals changes the population transition and its Hastings ratio.","primaryUrl":"https://arxiv.org/html/2606.25265","noteUrl":"/review?paper=review-neurips-2026-152888","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152890","title":"FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance","authors":["Jaihyun Lew","Mingi Jung","Minjun Park","Wooseok Song","Sungroh Yoon"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Fixed mixture time and fixed log-SNR cannot both hold under the stated rescaling","plainExplanation":"Equation (4) changes log-SNR by \\(2\\log(r_0/r)\\) at fixed schedule time when resolution changes, but the explanation holds both coordinates fixed. The two requirements are incompatible under that equation.","primaryUrl":"https://arxiv.org/html/2609.25716","noteUrl":"/review?paper=review-neurips-2026-152890","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152897","title":"Learning in Context, Guided by Choice: A Reward-Free Paradigm for Reinforcement Learning with Transformers","authors":["Juncheng Dong","Moyang Guo","Bowen He","Ethan Fang","Zhuoran Yang","Vahid Tarokh"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Learning in Context, Guided by Choice: A Reward-Free Paradigm for Reinforcement Learning with Transformers","plainExplanation":"The exponential tilt solves the objective with \\(D_{\\rm KL}(t\\|q)\\). With the printed KL direction, differentiating with respect to \\(t\\) gives a rational policy.","primaryUrl":"https://arxiv.org/html/2602.08244","noteUrl":"/review?paper=review-neurips-2026-152897","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152898","title":"A Theory of Training Profit-Optimal LLMs","authors":["Sophie Hao","Will Merrill"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A Theory of Training Profit-Optimal LLMs","plainExplanation":"The stationary production can be negative. Squaring its margin then gives fictitious positive revenue in a region where feasible production is zero.","primaryUrl":"https://arxiv.org/html/2605.16430","noteUrl":"/review?paper=review-neurips-2026-152898","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap","reporting_inconsistency"],"classifications":["major"]},{"slug":"review-neurips-2026-152900","title":"The Weight Gram Matrix Captures Sequential Feature Linearization in Deep Networks","authors":["Taehun Cha","Daniel Beaglehole","Adityanarayanan Radhakrishnan","Donghun Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strict activation monotonicity does not imply a strictly positive derivative","plainExplanation":"Strict activation monotonicity does not imply a strictly positive derivative.","primaryUrl":"https://arxiv.org/html/2605.06258","noteUrl":"/review?paper=review-neurips-2026-152900","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152901","title":"Drift Q-Learning","authors":["Achraf Anas El Houssaini","Mohamad Hosein Danesh","Amin Abyaneh","Scott Fujimoto","Hsiu-Chin Lin","David Meger"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Parallel candidate work is not sequential dependency depth","plainExplanation":"Table 2 permits independent candidate parallelism but labels \\(KN\\) as nonparallelizable depth for \\(N\\) independent chains of \\(K\\) steps. The chains have total work \\(KN\\) but sequential dependency depth \\(K\\).","primaryUrl":"https://arxiv.org/html/2606.00350","noteUrl":"/review?paper=review-neurips-2026-152901","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152902","title":"Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities","authors":["Hyunsik Kim","Youngmoon Jung"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01984","noteUrl":"/review?paper=review-neurips-2026-152902","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152903","title":"Do Deep Networks Forget Initialization? A Forgetting-Time View of Practical Inductive Bias","authors":["Mohua Das","Pierfrancesco Beneventano","Shibshankar Dey","Gareth H McKinley","Tomaso Poggio"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Do Deep Networks Forget Initialization? A Forgetting-Time View of Practical Inductive Bias","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.29152","noteUrl":"/review?paper=review-neurips-2026-152903","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152907","title":"Amortized Guidance for Image Inpainting with Pretrained Diffusion Models","authors":["Yilie Huang","XUNYU ZHOU"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Amortized Guidance for Image Inpainting with Pretrained Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13010","noteUrl":"/review?paper=review-neurips-2026-152907","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152908","title":"Verifiable Environments Are LEGO Bricks: Recursive Composition for Reasoning Generalization","authors":["Hao Xiang","Qiaoyu Tang","Le Yu","Yaojie Lu","Xianpei Han","Ben He","Le Sun","Bowen Yu","Peng Wang","Hongyu Lin","Dayiheng Liu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Verifiable Environments Are LEGO Bricks: Recursive Composition for Reasoning Generalization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.12373","noteUrl":"/review?paper=review-neurips-2026-152908","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152909","title":"Expressive Power of Deep Homomorphism Networks over Relational Databases","authors":["Balder ten Cate","Maurice Funk","Benny Kimelfeld","Carsten Lutz","Moritz Schönherr","Arie Soeteman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The extension sum needs labeled multiplicities","plainExplanation":"The border-propagation condition rules out a valid two-tile row. The displayed reduction is invalid as written; the undecidability theorem is not disproved after a consistent reduction repair.","primaryUrl":"https://arxiv.org/html/2605.22852","noteUrl":"/review?paper=review-neurips-2026-152909","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152910","title":"Driving Video Retrieval for Complex Queries with Structured Grounding","authors":["Manyi Yao","Sparsh Garg","Christian Shelton","Amit Roy-Chowdhury","Abhishek Aich"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Driving Video Retrieval for Complex Queries with Structured Grounding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.09109","noteUrl":"/review?paper=review-neurips-2026-152910","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152911","title":"VVTRec: Radio Interferometric Reconstruction through Visual and Textual Modality Enrichment","authors":["Kai Cheng","Ruoqi Wang","Qiong Luo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"VVTRec: Radio Interferometric Reconstruction through Visual and Textual Modality Enrichment","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.06475","noteUrl":"/review?paper=review-neurips-2026-152911","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152912","title":"OCTOPUS: Optimized KV Cache for Transformers via Octahedral Parametrization Under optimal Squared error quantization","authors":["Mark Boss","Vikram Voleti","Simon Donné","Shimon Vainer"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Zero-padded terminal blocks need a different norm prior","plainExplanation":"Zero-padded terminal blocks have a different norm law from full triplets, and orthogonal rotation does not make an arbitrary quantizer unbiased. The spherical octahedral map has a variable Jacobian, as the appendix already derives correctly. The diagonal bit-allocation sweep changes storage, so it cannot by itself establish a fixed-budget allocation gain. These findings do not refute the correct alternate density or the separately qualified ideal residual estimator.","primaryUrl":"https://arxiv.org/html/2605.21226","noteUrl":"/review?paper=review-neurips-2026-152912","collection":"neurips-2026","severity":"repairable","issueTypes":["documentation_error","formula_error","invalid_inference","missing_assumption","unsupported_claim"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152913","title":"Why Geometric Continuity Emerges in Deep Neural Networks: Residual Connections and Rotational Symmetry Breaking","authors":["Kyungwon Jeong","Won-Gi Paeng","Honggyo Suh"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The revised optimizer and activation arguments resolve two allegations","plainExplanation":"The literal product \\(W_OW_V\\) has incompatible inner dimensions. Grouped-query head expansion or a per-head aggregation is missing from the definition.","primaryUrl":"https://arxiv.org/html/2605.04971","noteUrl":"/review?paper=review-neurips-2026-152913","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_ambiguity"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-152916","title":"Distributional Alignment as a Criterion for Designing Task Vectors in In-Context Learning","authors":["Jihoon Kwon","Jiwon Choi","Jy-yong Sohn"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Distributional Alignment as a Criterion for Designing Task Vectors in In-Context Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20730","noteUrl":"/review?paper=review-neurips-2026-152916","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152917","title":"Exactness Matters for Physical Rule Enforcement","authors":["Bum Jun Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Exactness Matters for Physical Rule Enforcement","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08285","noteUrl":"/review?paper=review-neurips-2026-152917","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152922","title":"Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis","authors":["Shrestha Datta","Hongfu Liu","Anshuman Chhabra"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.20207","noteUrl":"/review?paper=review-neurips-2026-152922","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152924","title":"Distance Marching for Generative Modeling","authors":["Zimo Wang","Ishit Mehta","Haolin Lu","Xunpeng Huang","Chung-En Sun","Ge Yan","Lily Weng","Tzu-Mao Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Distance weighting saturates without becoming exactly length-neutral","plainExplanation":"The current formal distance-weight formula is correct, but it does not imply the repeated exact length-neutrality description. The renumbered scalar theorem still equates restricted scalar-field optimization with an unrestricted pointwise Bayes update without a realizability premise. The free weighted-regression identity and the decoupled vector architecture remain separate.","primaryUrl":"https://arxiv.org/html/2602.02928","noteUrl":"/review?paper=review-neurips-2026-152924","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152925","title":"ScreenSearch: Uncertainty-Aware OS Exploration","authors":["Michael Solodko","Justin Wagle"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"ScreenSearch: Uncertainty-Aware OS Exploration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16024","noteUrl":"/review?paper=review-neurips-2026-152925","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152927","title":"Deceptive Grounding: Entity Attribution Failure in Clinical Retrieval-Augmented Generation","authors":["Cedric Caruzzo","Donggeun Yoo","Tae Soo Kim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Deceptive Grounding: Entity Attribution Failure in Clinical Retrieval-Augmented Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.09349","noteUrl":"/review?paper=review-neurips-2026-152927","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152928","title":"Breaking the Quality–Privacy Tradeoff in Tabular Data Generation via In-Context Learning","authors":["Xinyan Han","yan Lu","Xiaoyu Lin","Yuanyuan Jiang","Yuanrui Wang","Xuanyue Li","Wenchao Zou","Xingxuan Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Breaking the Quality–Privacy Tradeoff in Tabular Data Generation via In-Context Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.04911","noteUrl":"/review?paper=review-neurips-2026-152928","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152930","title":"Coarsening Linear Non-Gaussian Causal Models with Cycles","authors":["Francisco Madaleno","Francisco Pereira","Alex Markham"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Acyclic-quotient partitions are not closed under ambient lattice joins","plainExplanation":"Stability does not uniquely identify an arbitrary cyclic structural model. The first-stable uniqueness claim is false; condensation identification is not refuted.","primaryUrl":"https://arxiv.org/html/2605.10163","noteUrl":"/review?paper=review-neurips-2026-152930","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","interpretation_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152931","title":"Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs","authors":["Jiahui Niu","Kefan Gu","Yucheng Zhao","shengwen Liang","Tiancai Wang","Xing Hu","ying wang","Huawei Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13778","noteUrl":"/review?paper=review-neurips-2026-152931","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152934","title":"AnyEdit: A Unified Framework for Speech and Singing Voice Editing with Real-World Environmental Consistency","authors":["Yunjia Zhang","Junan Zhang","Jing Yang","Xueyao Zhang","Fan Fan","Zhizheng Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The written flow-matching loss needs its squared norm","plainExplanation":"An unsquared norm generally learns a different field from the conditional mean required by the stated flow-matching construction.","primaryUrl":"https://arxiv.org/html/2610.15665","noteUrl":"/review?paper=review-neurips-2026-152934","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152937","title":"OASIS: Observation-Action Space Alignment via SE(3) Trajectory Prediction for Robotic Manipulation","authors":["Xinzhe Chen","Sihua Ren","Liqi Huang","Haowen Sun","Mingyang Li","Xingyu Chen","Zeyang Liu","Xuguang Lan"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"OASIS: Observation-Action Space Alignment via SE(3) Trajectory Prediction for Robotic Manipulation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.25829","noteUrl":"/review?paper=review-neurips-2026-152937","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152940","title":"To See Far, Look Close: Evolutionary Forecasting for Long-term Time Series","authors":["Jiaming Ma","Siyuan Mu","Ruilin Tang","Haofeng Ma","Qihe Huang","Zhengyang Zhou","Pengkun Wang","Binwu Wang","Yang Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"To See Far, Look Close: Evolutionary Forecasting for Long-term Time Series","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.23114","noteUrl":"/review?paper=review-neurips-2026-152940","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152942","title":"EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction","authors":["Shiyu Ji","Yixuan Wang","Yijun Liu","Qingfu Zhu","Wanxiang Che"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.22910","noteUrl":"/review?paper=review-neurips-2026-152942","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152945","title":"ANCRe: Adaptive Neural Connection Reassignment for Efficient Depth Scaling","authors":["Yilang Zhang","Bingcong Li","Niao He","Georgios Giannakis"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ANCRe: Adaptive Neural Connection Reassignment for Efficient Depth Scaling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09009","noteUrl":"/review?paper=review-neurips-2026-152945","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152947","title":"Intern-Atlas: A Methodological Evolution Graph as Research Infrastructure for AI Scientists","authors":["Yujun Wu","Dongxu Zhang","Xinchen Li","Jinhang Xu","Yiling Duan","Yumou Liu","Jiabao Pan","Qiyuan Zhu","Xuanhe Zhou","Jingxuan Wei","Siyuan Li","Jintao Chen","Conghui He","Cheng Tan"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Intern-Atlas: A Methodological Evolution Graph as Research Infrastructure for AI Scientists","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.28158","noteUrl":"/review?paper=review-neurips-2026-152947","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152949","title":"From Jumps to Signatures: a Generative Method for Temporal Point Processes","authors":["Niels Cariou-Kotlarek","Vasileios Lampos"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"From Jumps to Signatures: a Generative Method for Temporal Point Processes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.06652","noteUrl":"/review?paper=review-neurips-2026-152949","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152952","title":"Attention Drift: What Auto-Regressive Speculative Decoding Models Learn","authors":["Doğaç Eldenk","Payal Mohapatra","Yigitcan Comlek","Kaan Oktay","Hongyang Zhang","Stephen Xia"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Attention Drift: What Auto-Regressive Speculative Decoding Models Learn","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09992","noteUrl":"/review?paper=review-neurips-2026-152952","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152953","title":"MoBayes: A Modular Bayesian Framework for Separating Reasoning from Language in Conversational Clinical Decision Support","authors":["Yusuf Kesmen","Fay Elhassan","Jiayi Ma","Julien Stalhandske","Yena Chang","Alexandra V Kulinkina","David Sasu","Akhil Arora","Lars H Klein","Mary-Anne Hartley"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A binary indicator does not preserve all categorical-feature information","plainExplanation":"For one fixed reported value \\(v\\), the virtual node at full confidence is \\(V=\\mathbf1\\{X=v\\}\\), yet its endpoint is claimed to preserve all disease information in arbitrary finite-valued \\(X\\). Since \\(V\\) is deterministic from \\(X\\), \\(I(D;X)=I(D;V)+I(D;X\\mid V)\\); equality requires conditional sufficiency of that binary grouping.","primaryUrl":"https://arxiv.org/html/2604.20022","noteUrl":"/review?paper=review-neurips-2026-152953","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152954","title":"NPUsper: Eliminating Redundant Computation for Real-Time Whisper on Mobile NPUs","authors":["Hojeong Lee","Si H Lee","Sungwon Woo","Chengpo Yan","Suman Banerjee","Seyeon Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"NPUsper: Eliminating Redundant Computation for Real-Time Whisper on Mobile NPUs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.01108","noteUrl":"/review?paper=review-neurips-2026-152954","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152955","title":"TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization","authors":["Seokhyun Lee","Jaeho Kim","Changjun Oh","Mihaela van der Schaar","Changhee Lee"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Truncated sampled Gaussians do not obey the exact continuous semigroup identity","plainExplanation":"The consistency derivation composes Gaussian variance increments exactly, while the defined operators use normalized discrete kernels truncated at three standard deviations. Those finite sampled kernels do not form the continuous Gaussian convolution semigroup.","primaryUrl":"https://arxiv.org/html/2605.01418","noteUrl":"/review?paper=review-neurips-2026-152955","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152956","title":"Convex Compositional Reasoning Models","authors":["Meir Roketlishvili","Semen Semenov","Maksim Bobrin","Viktor Kovalchuk","Albert Baichorov","Abduragim Shtanchaev","Fakhri Karray","Dmitry V. Dylov","Martin Takac","Arip Asadulaev"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Satisfiability does not imply simultaneous falsifiability of all clauses","plainExplanation":"The Birkhoff domain is a relaxation rather than the N-Queens convex hull.","primaryUrl":"https://arxiv.org/html/2605.23395","noteUrl":"/review?paper=review-neurips-2026-152956","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152957","title":"A Theoretical Framework for Self-Play Theorem Proving Algorithms","authors":["Thomas Chen","Zhiyuan Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The eigengap disappears at the allowed zero augmentation time","plainExplanation":"The optimized spectral bound must retain both branches of the minimum. The eigengap disappears at the allowed zero augmentation time. The confidence-log comparison is not uniform near confidence parameter one.","primaryUrl":"https://arxiv.org/html/2606.01861","noteUrl":"/review?paper=review-neurips-2026-152957","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152958","title":"Ergodic Risk Measures: Towards a Risk-Aware Foundation for Continual Reinforcement Learning","authors":["Juan Rojas","Chi-Guhn Lee"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Nonlinear risk of the stationary reward mixture is not a fixed linear occupancy reward","plainExplanation":"Nonlinear risk of the stationary reward mixture is not a fixed linear occupancy reward. Time consistency does not make conditional risk insensitive to new observations.","primaryUrl":"https://arxiv.org/html/2510.02945","noteUrl":"/review?paper=review-neurips-2026-152958","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152960","title":"Align3D-AD: Cross-Modal Feature Alignment and Dual-Prompt Learning for Zero-shot 3D Anomaly Detection","authors":["Letian Bai","Xuanming Cao","Juan Du","Chengyu Tao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The discrepancy multiplier increases with agreement","plainExplanation":"The discrepancy multiplier increases with agreement. The general mechanism rationale is false; absolute residual size and measured anomaly accuracy are separate questions.","primaryUrl":"https://arxiv.org/html/2605.05850","noteUrl":"/review?paper=review-neurips-2026-152960","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152961","title":"Gluing Local Contexts into Global Meaning: A Sheaf-Theoretic Decomposition of Transformer Representations","authors":["Bryce Grant","Peng Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The assembled stalk Laplacian has a much larger algebraic nullspace","plainExplanation":"The coboundary adjoint annihilates harmonic edge features. The stated pullback cannot supply informative node features; this is not an accusation that the reported retrieval measurement was fabricated.","primaryUrl":"https://arxiv.org/html/2610.17117","noteUrl":"/review?paper=review-neurips-2026-152961","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152962","title":"Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning","authors":["Polina Dolgova","Sebastian Stich"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The bound decreases with absolute rather than signed cosine","plainExplanation":"The bound decreases with absolute rather than signed cosine. Only the signed-cosine inference fails; the preceding residual identity and numerical upper bound remain valid.","primaryUrl":"https://arxiv.org/html/2605.31317","noteUrl":"/review?paper=review-neurips-2026-152962","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152963","title":"Frontier Coding Agents Use Metaprogramming to Adapt to Unfamiliar Programming Languages","authors":["Aman Sharma","Sushrut Thorat","Paras Chopra"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Positive within-session dependence does not make binomial dispersion conservative","plainExplanation":"Positive within-session dependence does not make binomial dispersion conservative. The conservative-dispersion rationale has the wrong general direction. The explicit confidence disclaimer and independent-session ranking checks remain credited.","primaryUrl":"https://arxiv.org/html/2606.10933","noteUrl":"/review?paper=review-neurips-2026-152963","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152964","title":"Guidance For Prior Change via Density Ratio Estimation","authors":["Yichen Zang","Song Liu","Jiun-Yi Lin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The proposed sign flips do not preserve the printed covariance","plainExplanation":"Strict convexity alone does not ensure a finite ratio-estimation minimizer. The proposed sign flips do not preserve the printed covariance. The Two Moons likelihood has no claimed finite rejection envelope.","primaryUrl":"https://arxiv.org/html/2608.21729","noteUrl":"/review?paper=review-neurips-2026-152964","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152967","title":"Training Language Models to Explain Their Own Computations","authors":["Belinda Z Li","Zifan Carl Guo","Vincent Huang","Jacob Steinhardt","Jacob Andreas"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Training Language Models to Explain Their Own Computations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.08579","noteUrl":"/review?paper=review-neurips-2026-152967","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152968","title":"Asymmetric Invertible Threat: Learning Reversible Privacy Defense for Face Recognition","authors":["Jiabei Zhang","Ziyuan Yang","Andrew Beng Jin Teoh","Yi Zhang"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Asymmetric Invertible Threat: Learning Reversible Privacy Defense for Face Recognition","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.01217","noteUrl":"/review?paper=review-neurips-2026-152968","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152972","title":"BigCell: Generating Gigapixel Whole-Slide Images","authors":["Srikar Yellapragada","Alexandros Graikos","Zilinghan Li","Kostas Triaridis","Tarak N Nandi","Ji Dong K. Bai","Beatrice Knudsen","Tahsin Kurc","Prateek Prasanna","Rajarsi R Gupta","Ravi K Madduri","Joel Saltz","Dimitris Samaras"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BigCell: Generating Gigapixel Whole-Slide Images","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.14261","noteUrl":"/review?paper=review-neurips-2026-152972","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152974","title":"MID: Mask-Image Distributional Divergence for Evaluating Medical Image Segmentation","authors":["Vincenzo Marcianò","XIAOMING ZHANG","Gianluca Guglielmo","Sebastien Ourselin","Michela Antonelli","Maria A Zuluaga"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Discarded variance does not bound a discarded mean displacement","plainExplanation":"Discarded covariance variance does not control a discarded mean shift, and the nonlinear correspondence residual depends on marginal covariance blocks as well as cross blocks. The retained mathematical passages are unchanged; the metric and controlled dependence comparisons remain defined.","primaryUrl":"https://arxiv.org/html/2610.17354","noteUrl":"/review?paper=review-neurips-2026-152974","collection":"neurips-2026","severity":"major","issueTypes":["incorrect_bound","interpretation_error","invalid_inference","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-152976","title":"Symb-xMIL: Symbolic Explanations for Multiple Instance Learning in Digital Pathology","authors":["Yanqing Luo","Julius Hense","Niklas Prenißl","Andreas Mock","Klaus-Robert Müller","Thomas Schnake","Mina Jamshidi Idaji"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Exclusive instance labels do not make bag-level presence queries exclusive","plainExplanation":"Exclusive instance labels do not make bag-level presence queries exclusive.","primaryUrl":"https://arxiv.org/html/2606.06224","noteUrl":"/review?paper=review-neurips-2026-152976","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-152977","title":"Deep Minds and Shallow Probes","authors":["Su Hyeong Lee","Risi Kondor"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Dropping a nonzero SVD tail changes the exact mismatch identity","plainExplanation":"The exact thin-SVD theorem reconstructs every nonzero bank singular direction, but its practical truncated extension claims the same mismatch-only identity after dropping small nonzero directions. Take \\(W_1=[1]\\), \\(W_2=\\operatorname{diag}(1,10^{-4})\\), \\(M=[0,1]\\), target coordinates \\((a,b)\\) with independent signs, and \\(h_1=10^{-4}b\\).","primaryUrl":"https://arxiv.org/html/2605.11448","noteUrl":"/review?paper=review-neurips-2026-152977","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152978","title":"Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics","authors":["Jishnu S Nair","Patrice Bechard","Rishabh Maheshwary","SRAVAN RAMACHANDRAN","Surajit Dasgupta","Aakash Bhagat","Shruthan Radhakrishna","Pulkit Pattnaik","Johan Obando Ceron","Shiva Krishna Reddy Malay","Sagar Davasam","Seganrasan Subramanian","Vipul Mittal","Sridhar K Nemala","Chris Pal","Srinivas Sunkara","Sai Rajeswar Mudumba"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12178","noteUrl":"/review?paper=review-neurips-2026-152978","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152979","title":"OFBD: Object-Focused Background Debiasing for Long-Tailed Learning","authors":["Shenghan Chen","Yiming Liu","Zhipeng Deng","Haolin Wang","Jiale Zhou","Zhijian Wu","Lu Xiankai","yafei ou","Yefeng Zheng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A smaller distribution-distance bound need not mean smaller bias","plainExplanation":"A larger distribution discrepancy does not necessarily produce a larger feature bias, nor does reducing an upper bound necessarily reduce the actual bias. Feature reweighting alone also fails to control the full loss gradient, and separate early-training baselines can reverse the ordering of optimization shifts. Smaller corrections concern the support needed for the empirical-distribution rate and the separator sign in the appendix’s quadratic minimization; the explicitly acknowledged global-moment approximation is retained.","primaryUrl":"https://arxiv.org/html/2609.37331","noteUrl":"/review?paper=review-neurips-2026-152979","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152980","title":"AIRA 2: Overcoming Bottlenecks in AI Research Agents","authors":["Karen Hambardzumyan","Nicolas Baldwin","Edan Toledo","RISHI HAZRA","Michael Kuchnik","Bassel Al Omari","Thomas Foster","Anton Protopopov","Jean-Christophe Gagnon-Audet","Ishita Mediratta","Kelvin Niu","Michael Shvartsman","Alisia Lupidi","Alexis Audran-Reiss","Parth Pathak","Tatiana Shavrina","Despoina Magka","Hela Momand","Derek Dunfield","Nicola Cancedda","Pontus Lars Erik Saito Stenetorp","Carole-Jean Wu","Jakob Foerster","Yoram Bachrach","Martin Josifoski"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AIRA 2: Overcoming Bottlenecks in AI Research Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.26499","noteUrl":"/review?paper=review-neurips-2026-152980","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152984","title":"Knowledge-Level Consistency Reinforcement Learning: Dual-Fact Alignment for Long-Form Factuality","authors":["Junliang Li","Yucheng Wang","Yan Chen","Yu Ran","Ruiqing Zhang","Jing Liu","Hua Wu","Haifeng Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Knowledge-Level Consistency Reinforcement Learning: Dual-Fact Alignment for Long-Form Factuality","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2509.23765","noteUrl":"/review?paper=review-neurips-2026-152984","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-152988","title":"RAPTOR: Ridge-Adaptive Logistic Probes","authors":["Ziqi Gao","Yaotian Zhu","Qingcheng Zeng","Xu Zhao","Ziqing Wang","Feng Ruan","Kaize Ding"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The probe fold-back must invert the same standardization used in training","plainExplanation":"Algorithm 1 trains on \\((h-\\mu)/s\\) but folds weights back using an elementwise denominator \\(\\max(s,1)\\). A standardized affine logit has unique native parameters \\(\\omega=w/s\\) and intercept \\(b-\\omega^\\top\\mu\\).","primaryUrl":"https://arxiv.org/html/2602.00158","noteUrl":"/review?paper=review-neurips-2026-152988","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152990","title":"Rethinking Molecular Graph Backdoors under Chemistry-aware Admission","authors":["Thinh Nguyen","Sze Jue Yang","Khoa D Doan","Chee Seng Chan","Kok-Seng Wong"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Rethinking Molecular Graph Backdoors under Chemistry-aware Admission","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.23361","noteUrl":"/review?paper=review-neurips-2026-152990","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152992","title":"When Policies Cannot Be Retrained: A Unified Closed-Form View of Post-Training Steering in Offline Reinforcement Learning","authors":["Elias Hossain","Mohammad J Basher","Ivan Garibay","Ozlem Garibay","Niloofar Yousefi"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Old-occupancy Wasserstein action deviation does not control arbitrary occupancy changes","plainExplanation":"Old-occupancy Wasserstein action deviation does not control arbitrary occupancy changes. The variational Gaussian precision is missing a reciprocal beta factor. The actor endpoint loses actions excluded by the prior.","primaryUrl":"https://arxiv.org/html/2604.22873","noteUrl":"/review?paper=review-neurips-2026-152992","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-152993","title":"Transfer Learning Through Conditional Quantile Matching","authors":["Yikun Zhang","Steven Wilkins-Reeves","Wesley Lee","Aude Hofleitner"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Transfer Learning Through Conditional Quantile Matching","plainExplanation":"Equal response marginals can hide the wrong relationship with covariates. The report credits the later conditional code correction and gives explicit repairs for the remaining mathematical claims in arXiv v1.","primaryUrl":"https://arxiv.org/html/2602.02358","noteUrl":"/review?paper=review-neurips-2026-152993","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","formula_error","incorrect_bound","invalid_inference","lemma_error","missing_assumption","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-152994","title":"Budgeted Quotient-Residual Guidance for Frozen Pocket-Conditioned Molecular Diffusion","authors":["Xinyu Wang","Jinbo Bi","minghu song"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Budgeted Quotient-Residual Guidance for Frozen Pocket-Conditioned Molecular Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31222","noteUrl":"/review?paper=review-neurips-2026-152994","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-152996","title":"Tango3D: Towards Alignment for Global and Local 2D-3D Correspondence","authors":["Zebin He","Mingxin Yang","Shuhui Yang","Hanxiao Sun","Xintong Han","Chunchao Guo","Wenhan Luo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Post-pooling residual fusion does not exclude a semantic bypass","plainExplanation":"Post-pooling residual fusion does not exclude a semantic bypass. The universal architectural assurance is unsupported; actual learned bypass behavior is not established.","primaryUrl":"https://arxiv.org/html/2605.19727","noteUrl":"/review?paper=review-neurips-2026-152996","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-152999","title":"From Isolated feature to Orbits:\\\\Discovering Music Concepts via Multi-SAE Alignment","authors":["Liwei Lin","Gus Xia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Per-quality ring normalization needs a zero-mass convention","plainExplanation":"Per-quality ring normalization needs a zero-mass convention. The downstream formula lacks an admitted boundary case; its scale invariance on positive rings remains valid.","primaryUrl":"https://arxiv.org/html/2610.01864","noteUrl":"/review?paper=review-neurips-2026-152999","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153002","title":"Teaching Large Language Models When Not to Know: Learning Temporal Critique for Ex-Ante Reasoning","authors":["Chenlu Ding","Jiancan Wu","Yanchen Luo","Zheyuan (Frank) Liu","Yancheng Yuan","Xiang Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Conditional answer imitation can represent cutoff-dependent validity","plainExplanation":"Conditional answer imitation can represent cutoff-dependent validity.","primaryUrl":"https://arxiv.org/html/2605.14636","noteUrl":"/review?paper=review-neurips-2026-153002","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153003","title":"Differentiable Nonlinear Model Predictive Control","authors":["Jonathan Frey","Katrin Baumgärtner","Gianluca Frison","Dirk Reinhardt","Jasper Hoffmann","Leonard Fichtner","Joschka Boedecker","Sebastien Gros","Moritz Diehl"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"NLP second-order sufficiency does not make its full-space Hessian QP convex","plainExplanation":"The barrier complementarity product has the wrong sign. NLP second-order sufficiency does not make its full-space Hessian QP convex. A fixed-base quadratic model agrees in sensitivity only at its base point.","primaryUrl":"https://arxiv.org/html/2505.01353","noteUrl":"/review?paper=review-neurips-2026-153003","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153004","title":"M2A: Synergizing Mathematical and Agentic Reasoning in Large Language Models","authors":["JunJian Wang","Xin Zhou","Qiran Xu","Kun Zhan"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Positive ridge leaves a protected-subspace residual","plainExplanation":"Positive ridge leaves a protected-subspace residual. The ridge update is not exactly null-space preserving; the ideal projection theorem is not refuted.","primaryUrl":"https://arxiv.org/html/2605.09879","noteUrl":"/review?paper=review-neurips-2026-153004","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153005","title":"Towards Direct Latent-Space Synthesis for Parallel Branches in LLM-Agent Workflows","authors":["Shikun Liu","Mufei Li","Dongqi Fu","Haoyu Wang","Yinglong Xia","Hong Li","Hong Yan","Pan Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Towards Direct Latent-Space Synthesis for Parallel Branches in LLM-Agent Workflows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.14672","noteUrl":"/review?paper=review-neurips-2026-153005","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153010","title":"Clustered Randomized Smoothing for Stochastic Prediction Functions","authors":["Eduardo Figueiredo","Frederik B Mathiesen","Julian F Schumann","Jens Kober","Arkady Zgonnikov","Luca Laurenti"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Coordinatewise trimming does not preserve general convex coverage sets","plainExplanation":"Convexity preserves averages of whole vectors. Sorting and trimming each coordinate separately can instead combine coordinates into a vector outside the convex hull of all the samples. The containment step in Appendix A.1 therefore fails for general convex coverage sets, even with independent deterministic-predictor outputs. The contradiction concerns the general-convex extension in Theorem 3.1. The correction covers axis-aligned rectangles with independent draws; it does not cover shared stochastic weights.","primaryUrl":"https://arxiv.org/html/2608.12037","noteUrl":"/review?paper=review-neurips-2026-153010","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153011","title":"A Statistical Theory of Gated Attention through the Lens of Hierarchical Mixture of Experts","authors":["Viet Nguyen","Thinh Cao","Tuan M Pham","Tan Dinh","Huy Nguyen","Nhat Ho","Alessandro Rinaldo"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A Statistical Theory of Gated Attention through the Lens of Hierarchical Mixture of Experts","plainExplanation":"A lower bound does not supply a finite upper recovery rate, and the individual coefficients can differ while generating exactly the same regression law.","primaryUrl":"https://arxiv.org/html/2602.01468","noteUrl":"/review?paper=review-neurips-2026-153011","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153014","title":"Learning Where It Matters: Geometric Anchoring for Robust Preference Alignment","authors":["Youngjae Cho","Jongsuk Kim","Ji-Hoon Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Where It Matters: Geometric Anchoring for Robust Preference Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04909","noteUrl":"/review?paper=review-neurips-2026-153014","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153015","title":"Counterfactual Maps: What They Are and How to Find Them","authors":["Awa Khouna","Julien Ferry","Thibaut Vidal"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Counterfactual Maps: What They Are and How to Find Them","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09128","noteUrl":"/review?paper=review-neurips-2026-153015","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153017","title":"SGD Provably Prioritizes a Shortcut Spurious Feature in the XOR Model","authors":["Tyler LaBonte","Vidya Muthukumar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The universal Gaussian-moment error scale vanishes on a nonzero sparse vector","plainExplanation":"The Phase I recurrence replaces a quadrature norm by a sum, while the Phase II telescope uses the wrong exponential and a local-neuron bound drops network-wide contributions. The learned-weight conditioning argument also lacks its required independence. Further local errors affect norm centering, sparse-vector Gaussian moments, Gaussian-cubic tails, binomial confidence tolerances, shifted means, Taylor signs and the zero-start scalar recurrence. These are scoped proof defects; the final random-SGD theorem is not disproved by the counterexamples.","primaryUrl":"https://arxiv.org/html/2606.30444","noteUrl":"/review?paper=review-neurips-2026-153017","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153018","title":"Exploring Multi-Order Self-Similarity for Motion Understanding","authors":["Manjin Kim","Heeseung Kwon","Karteek Alahari","Minsu Cho"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Exploring Multi-Order Self-Similarity for Motion Understanding","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.20760v1","noteUrl":"/review?paper=review-neurips-2026-153018","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153019","title":"The Geometry of Alignment Collapse: When Fine-Tuning Breaks Safety","authors":["Max Springer","Chung Peng Lee","Bohdan Turbal","Blossom Metevier","Hayoung Jung","Jane Castleman","Zeyu Shen","Aleksandra Korolova"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Geometry of Alignment Collapse: When Fine-Tuning Breaks Safety","plainExplanation":"The claimed dimension suppression needs quantitative spectral assumptions, and the stated instability conditions do not force alignment loss.","primaryUrl":"https://arxiv.org/html/2602.15799","noteUrl":"/review?paper=review-neurips-2026-153019","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-153021","title":"DTA-GT: Direction- and Topology-Aware Graph Transformer for Neural Network Representation Learning","authors":["Yuxiang Zeng","Kun Xie","Yang Wang","Jigang Wen","Xiaocan Li","Guangxing Zhang","Gaogang Xie","Jiannong Cao"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"DTA-GT: Direction- and Topology-Aware Graph Transformer for Neural Network Representation Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16822","noteUrl":"/review?paper=review-neurips-2026-153021","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153024","title":"PixelPonder: Dynamic Patch Adaptation for Enhanced Multi-Conditional Text-to-Image Generation","authors":["Yanjie Pan","Qingdong He","Zhengkai Jiang","Pengcheng Xu","Chaoyi Wang","Jinlong Peng","Haoxuan Wang","Yun Cao","Zhenye Gan","Mingmin Chi","Bo Peng","Yabiao Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"An integrated straight-path velocity loss does not identify the learned transport","plainExplanation":"Equation (17) compares the integral of velocity along an explicitly prescribed straight interpolation with its endpoint displacement. The field \\(v(z,t)=(12t-6)z\\) integrates along every straight path to its endpoint difference, giving zero displayed training loss.","primaryUrl":"https://arxiv.org/html/2503.06684","noteUrl":"/review?paper=review-neurips-2026-153024","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153028","title":"Can Language Models Actually Retrieve In-Context? Drowning in Documents at Million Token Scale","authors":["Siddharth Gollapudi","Prasann Singhal","Nilesh Gupta","Sewon Min"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Off-candidate prefixes can leave every teacher vote zero","plainExplanation":"Algorithm F normalizes per-digit candidate votes at every sampled rollout state without masking unsupported prefixes or defining a fallback. If a sampled first digit is absent from the query’s sixteen candidates, every later matching candidate set is empty.","primaryUrl":"https://arxiv.org/html/2607.01538","noteUrl":"/review?paper=review-neurips-2026-153028","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153033","title":"Learning-Augmented Approximation for Unrelated-Machines Makespan Scheduling","authors":["Kaito Baba","Evripidis Bampis","Georgios Mitropoulos"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning-Augmented Approximation for Unrelated-Machines Makespan Scheduling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.13133","noteUrl":"/review?paper=review-neurips-2026-153033","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153034","title":"Jailbreak susceptibility prediction and mitigation via the behavioral geometry of models","authors":["Hayden Helm","Xiaodong Liu","Weiwei Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Ordinary k-medoids minimizes a sum rather than the largest nearest-center distance","plainExplanation":"Ordinary k-medoids minimizes a sum rather than the largest nearest-center distance.","primaryUrl":"https://arxiv.org/html/2605.26409","noteUrl":"/review?paper=review-neurips-2026-153034","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153035","title":"Auction-Based Online Policy Adaptation for Evolving Objectives","authors":["Guru Shabadi","Kaushik Mallik"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Auction-Based Online Policy Adaptation for Evolving Objectives","plainExplanation":"The auction guarantees require attention to the actual bid grid, cap, tie rule and winner-dependent rewards. The report gives terminal-game counterexamples and identifies restricted statements that can be proved; shared progress toward goals defeats the separate slack-based argument.","primaryUrl":"https://arxiv.org/html/2604.02151","noteUrl":"/review?paper=review-neurips-2026-153035","collection":"neurips-2026","severity":"critical","issueTypes":["definition_error","false_statement","missing_assumption","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-153036","title":"SemanticDialect: Semantic-Aware Mixed-Format Quantization for Video Diffusion Transformers","authors":["Wonsuk Jang","Thierry Tambe"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"SemanticDialect: Semantic-Aware Mixed-Format Quantization for Video Diffusion Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.02883","noteUrl":"/review?paper=review-neurips-2026-153036","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153039","title":"EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting","authors":["Junwei Luo","Shuai Yuan","Zhenya YANG","Yansheng Li","Zhe Liu","Hengshuang Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.27277","noteUrl":"/review?paper=review-neurips-2026-153039","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153040","title":"SSR3D-LLM: Structured Spatial Reasoning via Latent Steps for Fine-Grained Grounding in Unified 3D-LLMs","authors":["Jiawei LI","Ziyi Liu","Weijie Shi","Long Chen","Jiajie Xu","Xiaofang Zhou"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SSR3D-LLM: Structured Spatial Reasoning via Latent Steps for Fine-Grained Grounding in Unified 3D-LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28490","noteUrl":"/review?paper=review-neurips-2026-153040","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153042","title":"Efficiently Aligning Draft Models via Parameter- and Data-Efficient Adaptation","authors":["Luxi Lin","Zhihang Lin","Zhanpeng Zeng","Yuhao Chen","Qingyu Zhang","Jixiang Luo","Rongrong Ji"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Direct target fallback does not preserve stochastic speculative decoding","plainExplanation":"Direct target fallback does not preserve stochastic speculative decoding. The stochastic verification description is incomplete; inference-code behavior is not established.","primaryUrl":"https://arxiv.org/html/2603.09527","noteUrl":"/review?paper=review-neurips-2026-153042","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153043","title":"Differentiable Learning of Lifted Action Schemas for Classical Planning","authors":["Jonas Reiter","Jakob Gebler","Hector Geffner"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The displayed binary cross-entropy is missing its minus sign","plainExplanation":"Equation (9) writes \\(y\\log p+(1-y)\\log(1-p)\\) as a BCE loss to minimize. For target one it is \\(\\log p\\), increasing with \\(p\\) and tending to minus infinity as \\(p\\downarrow0\\).","primaryUrl":"https://arxiv.org/html/2605.13282","noteUrl":"/review?paper=review-neurips-2026-153043","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153044","title":"Regret Minimization in Single-Dimensional Contract-Design with Binary Actions","authors":["Riccardo Poiani","Martino Bernasconi","Andrea Celli"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Regret Minimization in Single-Dimensional Contract-Design with Binary Actions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06125","noteUrl":"/review?paper=review-neurips-2026-153044","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153045","title":"LLMs can construct powerful representations and streamline sample-efficient supervised learning","authors":["Ilker Demirel","Lawrence Shi","Zeshan Hussain","David Sontag"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"LLMs can construct powerful representations and streamline sample-efficient supervised learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.11679","noteUrl":"/review?paper=review-neurips-2026-153045","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153049","title":"Sparse All-Layer Connector for Domain Generalised Semantic Segmentation","authors":["Ruoyu Guo","XIN KUN LIN","Maurice Pagnucco","Yang Song"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A globally nonempty dropout mask can still zero an individual token row","plainExplanation":"A globally nonempty dropout mask can still zero an individual token row. The printed dropout/renormalization pipeline omits a per-token boundary rule.","primaryUrl":"https://arxiv.org/html/2610.15660","noteUrl":"/review?paper=review-neurips-2026-153049","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153050","title":"History-Aware Conformal Prediction Sets for Censored Time-to-Event Outcomes","authors":["Yuyao Wang","Alexander W Levis","Shu Yang","Larry Han"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"History-Aware Conformal Prediction Sets for Censored Time-to-Event Outcomes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06581","noteUrl":"/review?paper=review-neurips-2026-153050","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153051","title":"3DSPA: A 3D Semantic Point Autoencoder for Evaluating Video Realism","authors":["Bhavik Chandna","Kelsey Allen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"3DSPA: A 3D Semantic Point Autoencoder for Evaluating Video Realism","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.20354","noteUrl":"/review?paper=review-neurips-2026-153051","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153052","title":"Reflected Schrödinger Bridge Matching","authors":["Marcus Häggbom","Viktor Nilsson","Pierre Nyquist","Joakim Andén"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Reflected Schrödinger Bridge Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.03626","noteUrl":"/review?paper=review-neurips-2026-153052","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153053","title":"Complexity of Classical Acceleration for $\\ell_1$-Regularized PageRank","authors":["Kimon Fountoulakis","David Martínez-Rubio"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Complexity of Classical Acceleration for $\\ell_1$-Regularized PageRank","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.21138","noteUrl":"/review?paper=review-neurips-2026-153053","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153054","title":"dMoE: dLLMs with Learnable Block Experts","authors":["Sicheng Feng","Zigeng Chen","Gongfan Fang","Xinyin Ma","Xinchao Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A top-p coreset need not contain eight experts","plainExplanation":"The routing specification selects eight experts per token from an aggregate top-p coreset. Let all 32 token distributions coincide, with first-expert mass \\(0.99\\) and total mass \\(0.009\\) on the next seven experts.","primaryUrl":"https://arxiv.org/html/2605.30876","noteUrl":"/review?paper=review-neurips-2026-153054","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153055","title":"MoCA: Mixture-of-Components Attention for Scalable Compositional 3D Generation","authors":["Zhiqi Li","Wenhuan Li","Tengfei Wang","Zhenwei Wang","Junta Wu","Haoyuan Wang","Yunhan Yang","Zehuan Huang","YANG LI","Chunchao Guo","Peidong Liu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MoCA: Mixture-of-Components Attention for Scalable Compositional 3D Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.07628","noteUrl":"/review?paper=review-neurips-2026-153055","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153056","title":"PickMoment: Continuous-Time Single-Image-to-Video via Learning Deblurring and Blur-to-Video","authors":["Junseong Shin","Hyeonsu Jo","Daehyun Kim","Tae Hyun Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PickMoment: Continuous-Time Single-Image-to-Video via Learning Deblurring and Blur-to-Video","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01279","noteUrl":"/review?paper=review-neurips-2026-153056","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153060","title":"Online Fair Division Meets Reordering Buffers","authors":["Georgios Amanatidis","Giulio Giaconi","Evangelos Markakis","Nicos Protopapas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Online Fair Division Meets Reordering Buffers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.01159","noteUrl":"/review?paper=review-neurips-2026-153060","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153061","title":"Partially Performative Prediction","authors":["Jaewook Lee","Tijana Zrnic"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Partially Performative Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.07890","noteUrl":"/review?paper=review-neurips-2026-153061","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153065","title":"Towards Financial World Modeling","authors":["Humzah Merchant","Alec Guthrie","Simon Mahns","Randall Balestriero","Bradford Levy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Towards Financial World Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.09048","noteUrl":"/review?paper=review-neurips-2026-153065","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153066","title":"Rep2Text: Decoding Full Text from a Single LLM Token Representation","authors":["Haiyan Zhao","Zirui He","Yiming Tang","Fan Yang","Ali Payani","Dianbo Liu","Mengnan Du"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Rep2Text: Decoding Full Text from a Single LLM Token Representation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2511.06571","noteUrl":"/review?paper=review-neurips-2026-153066","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153071","title":"GRASP: Learning to Ground Social Reasoning in Multi-Person Non-Verbal Interactions","authors":["Junho Kim","Xu Cao","Houze Yang","Bikram Boote","Ana Jojic","Fiona Ryan","Bolin Lai","Sangmin Lee","James Rehg"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"GRASP: Learning to Ground Social Reasoning in Multi-Person Non-Verbal Interactions","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15764","noteUrl":"/review?paper=review-neurips-2026-153071","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153073","title":"The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes","authors":["Siqi Zhu","Xuyan Ye","Hongyu Lu","Weiye Shi","Ge Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A pointwise optimum does not explain interference across different examples","plainExplanation":"Equations (10)–(11) minimize reverse KL separately for each input and use a geometric mixture over privileged information conditional on that same input. If privileged information is deterministic given the input, \\(\\mathcal D(I\\mid x)\\) is a point mass at \\(I(x)\\).","primaryUrl":"https://arxiv.org/html/2605.11182","noteUrl":"/review?paper=review-neurips-2026-153073","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153074","title":"Decomposing how prompting steers behavior","authors":["Fan Cheng","Nikolaus Kriegeskorte"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Exact activation patching is not a universal behavioral ceiling in a different context","plainExplanation":"Exact activation patching is not a universal behavioral ceiling in a different context.","primaryUrl":"https://arxiv.org/html/2606.03093","noteUrl":"/review?paper=review-neurips-2026-153074","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153076","title":"Self-EvolveRec: Self-Evolving Recommender Systems with LLM-based Directional Feedback","authors":["Sein Kim","Sangwu Park","HongSeok Kang","Wonjoong Kim","Jimin Seo","Yeonjun In","Kanghoon Yoon","Hyunsik Jeon","Chanyoung Park"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Self-EvolveRec: Self-Evolving Recommender Systems with LLM-based Directional Feedback","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.12612","noteUrl":"/review?paper=review-neurips-2026-153076","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153077","title":"FedVSSAM: Mitigating Flatness Incompatibility in Sharpness-Aware Federated Learning","authors":["Bingnan Xiao","Yuan Gao","Bingcong Li","Wei Ni","Xin Wang","Tony Quek"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FedVSSAM: Mitigating Flatness Incompatibility in Sharpness-Aware Federated Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09144","noteUrl":"/review?paper=review-neurips-2026-153077","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153078","title":"Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL","authors":["Junyi Wu","Weijian Luo","Haoyang Zheng","Ruizhe Zhang","Guang Lin"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24001","noteUrl":"/review?paper=review-neurips-2026-153078","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153085","title":"Understanding Circulant Permutation to Extend the CMinHash Estimator","authors":["Keegan Kang","Avery Hood","Benedict H Wong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Understanding Circulant Permutation to Extend the CMinHash Estimator","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.18551","noteUrl":"/review?paper=review-neurips-2026-153085","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153086","title":"LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE","authors":["Lei Jin","Yu Shang","Yiding Ma","Xinhao Jin","Xin Zhang","Chen Gao","Wei Wu","Yong Li"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.21790","noteUrl":"/review?paper=review-neurips-2026-153086","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153089","title":"A second order regret bound for NormalHedge","authors":["Yoav S Freund","Nicholas Harvey","Victor S. Portella","Yabing Qi","Yu-Xiang Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A second order regret bound for NormalHedge","plainExplanation":"The counting step overstates the comparator mass between quantile lattice points.","primaryUrl":"https://arxiv.org/html/2602.08151","noteUrl":"/review?paper=review-neurips-2026-153089","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153090","title":"RepFusion: Leveraging Multimodal Priors for Denoising in Representation Space","authors":["Xichen Pan","Satya Narayan Shukla","Aashu Singh","Shlok K Mishra","Saining Xie"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"RepFusion: Leveraging Multimodal Priors for Denoising in Representation Space","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.14700","noteUrl":"/review?paper=review-neurips-2026-153090","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153091","title":"VidVec: Unlocking Video MLLM Embeddings for Video-Text Retrieval","authors":["Issar Tzachor","Dvir Samuel","Rami Ben-Ari"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"VidVec: Unlocking Video MLLM Embeddings for Video-Text Retrieval","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.08099","noteUrl":"/review?paper=review-neurips-2026-153091","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153094","title":"INFUSER: Influence-Guided Self-Evolution Improves Reasoning","authors":["Siyu Chen","Miao Lu","Beining Wu","Heejune Sheen","Fengzhuo Zhang","Shuangning Li","Zhiyuan Li","Jose Blanchet","Tianhao Wang","Zhuoran Yang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"INFUSER: Influence-Guided Self-Evolution Improves Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.09052","noteUrl":"/review?paper=review-neurips-2026-153094","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153096","title":"Barycentric Guidance: Turning Foundation Image Editors into Continuous Affective Controllers","authors":["Harvey Mannering","Zhiwu Huang","Adam Prugel-Bennett"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Barycentric Guidance: Turning Foundation Image Editors into Continuous Affective Controllers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.16892","noteUrl":"/review?paper=review-neurips-2026-153096","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153098","title":"FlowBank: Query-Adaptive Agentic Workflows Optimization through Precompute-and-Reuse","authors":["Lingzhi Yuan","Chenghao Deng","Fangxu Yu","Souradip Chakraborty","Mohammad Rostami","Furong Huang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The saturation test occurs after exponential subset enumeration","plainExplanation":"The saturation test occurs after exponential subset enumeration. The pseudocode and early-saturation efficiency rationale disagree; actual runtime is not inferred.","primaryUrl":"https://arxiv.org/html/2606.11290","noteUrl":"/review?paper=review-neurips-2026-153098","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153100","title":"Direct Acceleration of Stochastic Root-Finding Without Variance Reduction and Regularization","authors":["TaeHo Yoon","Nicolas Loizou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Direct Acceleration of Stochastic Root-Finding Without Variance Reduction and Regularization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.12043","noteUrl":"/review?paper=review-neurips-2026-153100","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153104","title":"RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models","authors":["Weijia Liufu","Xiaoyu Guo","Ruiyi Chen","Jingzhi Liu","Kaidong Zhang","Xiwen Liang","Jianqi Lin","Dawei Sun","Yuze Wang","Rongtao Xu","Bingqian Lin","Bowen Yang","Tongtong Cao","Bowen Peng","Dongyu Zhang","Guangrun Wang","Min Wang","Liang Lin","Xiaodan Liang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models","plainExplanation":"The cosine-based failure score can still be negative, so the displayed labeling rule does not guarantee its stated nonnegative range.","primaryUrl":"https://arxiv.org/html/2605.09410","noteUrl":"/review?paper=review-neurips-2026-153104","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153106","title":"AME-TS: Anchored Mixture-of-Experts for Time Series Forecasting","authors":["Rui Wang","Renhao Xue","Ray Razi","Huan Song","Hannah Marlowe"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The analytic descriptors are not defined on an all-zero window","plainExplanation":"The analytic descriptors are not defined on an all-zero window. The analytic-label domain is incomplete; no observed training NaN is established.","primaryUrl":"https://arxiv.org/html/2605.25166","noteUrl":"/review?paper=review-neurips-2026-153106","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153107","title":"AgenticOCR: Parsing Only What You Need for Efficient Retrieval-Augmented Generation","authors":["Zhengren Wang","Dongsheng Ma","Huaping Zhong","Jiayu Li","Yijie Wang","Tianyao He","Wentao Zhang","Bin Wang","Conghui He"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The reward expression adds two of its stated penalties","plainExplanation":"The reward expression adds two of its stated penalties. This is a local reward-sign error; the executable RL reward need not share it.","primaryUrl":"https://arxiv.org/html/2602.24134","noteUrl":"/review?paper=review-neurips-2026-153107","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153108","title":"Exact power indices for plurality-voting ensembles","authors":["Ilie Sarpe","Theofanis Georgakopoulos","Aristides Gionis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The multiclass Shapley pseudocode omits incorrect-class work and accumulation","plainExplanation":"The multiclass Shapley pseudocode omits incorrect-class work and accumulation. The printed algorithm is incomplete; the intended exact formulas and a corrected implementation can remain valid.","primaryUrl":"https://arxiv.org/html/2610.17980","noteUrl":"/review?paper=review-neurips-2026-153108","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153109","title":"Safeguarding Mutual Correction in Source-Free Domain Adaptation via Cut Statistics","authors":["Seongjun Lee","Changhee Lee"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"A larger accuracy difference does not imply a larger correction ratio","plainExplanation":"Separate decreasing calibration curves do not order two different models. A larger accuracy difference does not imply a larger correction ratio. Marginal calibration does not imply calibration inside joint cut regions.","primaryUrl":"https://arxiv.org/html/2610.02981","noteUrl":"/review?paper=review-neurips-2026-153109","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153110","title":"CofactVLA: Deconfounding Vision-Language-Action Models via Counterfactual Intervention","authors":["Yan Zhang","Yinan Wu","Haoran Duan","Jungong Han"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"CofactVLA: Deconfounding Vision-Language-Action Models via Counterfactual Intervention","plainExplanation":"Even a perfect orthogonal decomposition leaves the original nuisance component untouched.","primaryUrl":"https://arxiv.org/html/2608.04396","noteUrl":"/review?paper=review-neurips-2026-153110","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","mathematical_error"],"classifications":["major"]},{"slug":"review-neurips-2026-153111","title":"Beyond SFT-to-RL: Pre-alignment via Black-box On-policy Distillation for Multimodal RL","authors":["Sudong Wang","Weiquan Huang","Xiaomin Yu","Zuhao Yang","Hehai Lin","Keming Wu","Chaojun Xiao","CHEN CHEN","Wenxuan Wang","Beier Zhu","Yunjian Zhang","Chengwei Qin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Bradley–Terry training is not the displayed raw-score maximization","plainExplanation":"The paper identifies its Bradley–Terry discriminator update with maximizing the expected raw score difference in Equation (4). Take two responses with \\(P(a)=0.75\\), \\(Q(a)=0.5\\), and discriminator scores \\(D(a)=t,D(b)=0\\).","primaryUrl":"https://arxiv.org/html/2604.28123","noteUrl":"/review?paper=review-neurips-2026-153111","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153114","title":"ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection","authors":["Qingwen Zeng","Zehao Fu","Shuyu Meng","Linghan Huang","Jiayi Zhang","Chenglin Wu","Ling Chen","Huaming Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Shared tokenwise standardization cannot change the layer argmax","plainExplanation":"A decoder intervention does not uniquely identify subspace selection as the cause. The toy is a fixed-code conditional decoder optimum, not a global SAE optimum or Llama reproduction. The evaluated fixed-layer controls and observed negative-CE results remain credited; unique causal identification does not follow.","primaryUrl":"https://arxiv.org/html/2610.04905","noteUrl":"/review?paper=review-neurips-2026-153114","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153115","title":"Distill to Think, Foresee to Act: Cognitive-Physical Reinforcement Learning for Autonomous Driving","authors":["Yang Wu","Qiang Meng","Zhaojiang Liu","Youquan Liu","Jian Yang","Jin Xie"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Reward penalties and trajectory ranking do not impose hard safety constraints","plainExplanation":"The paper describes collision and drivable-area terms in an additive reward, followed by cognitive filtering and physical-expert ranking, as enforcing hard physical constraints. A finite penalty changes an objective but does not exclude infeasible trajectories.","primaryUrl":"https://arxiv.org/html/2605.21139","noteUrl":"/review?paper=review-neurips-2026-153115","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153117","title":"SEED: Self-Speculative Decoding via Implicit Encoder–Decoder","authors":["Hankun Lin","Patrick Pynadath","Ruqi Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SEED: Self-Speculative Decoding via Implicit Encoder–Decoder","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36590","noteUrl":"/review?paper=review-neurips-2026-153117","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153119","title":"Non-Colliding Biometric Identities for Digital Entities: Geometry, Capacity, and Million-Scale Virtual Identity Provisioning","authors":["Yuyang Ji","Yixuan Shen","Anil Jain","Xiaoming Liu","Feng Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Independent real identities do not make their shared-identity collision indicators independent","plainExplanation":"Independent real identities do not make their shared-identity collision indicators independent. The interval algebra is valid conditional on a Poisson model; its asserted applicability and unconditional coverage need additional support.","primaryUrl":"https://arxiv.org/html/2605.18238","noteUrl":"/review?paper=review-neurips-2026-153119","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153122","title":"HyperVAttention: Efficient Sparse Attention with Spatio-Temporal Clustering for Video Diffusion","authors":["Dongyeun Lee","Amir Zandieh","Vahab Mirrokni","Junmo Kim","Insu Han"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"HyperVAttention: Efficient Sparse Attention with Spatio-Temporal Clustering for Video Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.03012","noteUrl":"/review?paper=review-neurips-2026-153122","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153124","title":"Steer2Edit: From Activation Steering to Component-Level Editing","authors":["Chung-En Sun","Ge Yan","Zimo Wang","Lily Weng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Steer2Edit: From Activation Steering to Component-Level Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09870","noteUrl":"/review?paper=review-neurips-2026-153124","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153128","title":"Fisher information and the geometry of memorization in neural networks","authors":["Daniel Bernstein","Chase Goddard","Luca Di Carlo","Vudtiwat Ngampruetikorn","David Schwab"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The exact barcode-risk formula needs positive signal and the actual residual rank","plainExplanation":"The exact barcode-risk formula needs positive signal and the actual residual rank. The exact unconditional identity has a domain gap. Rare rank failure at large experimental sample size does not imply empirical failure.","primaryUrl":"https://arxiv.org/html/2610.16161","noteUrl":"/review?paper=review-neurips-2026-153128","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153129","title":"Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models","authors":["Kun Feng","Shaocheng Lan","Yuchen Fang","Wenchao He","Sihan Lu","Shuqi Gu","Lintao Ma","Xingyu Lu","Kan Ren"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A near-zero ARCH-LM statistic does not indicate strong ARCH effects","plainExplanation":"Appendix I.1 interprets near-zero ARCH-LM statistics as evidence of high time-varying volatility. The LM statistic is \\(NR^2\\) from regressing squared residuals on their lags.","primaryUrl":"https://arxiv.org/html/2509.25826","noteUrl":"/review?paper=review-neurips-2026-153129","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153131","title":"Total Variation Distance Estimation in Autoregressive Models","authors":["Eric Price","Kevin Tian","Zhiyang Xun","Yusong Zhu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Total Variation Distance Estimation in Autoregressive Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.19510","noteUrl":"/review?paper=review-neurips-2026-153131","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153132","title":"Finding Simple Proofs for First-Order Optimization","authors":["Daniel Berg Thomsen","Manu Upadhyaya","Baptiste Goujaud","Aymeric Dieuleveut","Adrien Taylor"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Finding Simple Proofs for First-Order Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.08753","noteUrl":"/review?paper=review-neurips-2026-153132","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153134","title":"Neglected Free Lunch from Post-training: Progress Advantage for LLM Agents","authors":["Changdae Oh","Wendi Li","Seongheon Park","Samuel (Min-Hsuan) Yeh","Tanwi Mallick","Sharon Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Clipping a surrogate does not place the final policy in the assumed reference-policy ball","plainExplanation":"Clipping a surrogate does not place the final policy in the assumed reference-policy ball. Proposition 1 for KL-regularized optima remains valid; its universal coverage of clipping-only methods is unsupported.","primaryUrl":"https://arxiv.org/html/2606.26080","noteUrl":"/review?paper=review-neurips-2026-153134","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153135","title":"DialectLLM: A Dialect-Aware Dialog[ue] Generation Framework Beyond Standard American English","authors":["Jio Oh","Paul Vicinanza","Thomas Butler","Steven Whang","Dezhi Hong","Amani Namboori"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"DialectLLM: A Dialect-Aware Dialog[ue] Generation Framework Beyond Standard American English","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.22888","noteUrl":"/review?paper=review-neurips-2026-153135","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153138","title":"DAWN: Dependency-Aware Fast Inference for Diffusion LLMs","authors":["Lizhuo Luo","Zhuoran Shi","Jiajun Luo","Zhi Wang","Shen Ren","Wenya Wang","Tianwei Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The threshold-only update rule can leave the decoder permanently masked","plainExplanation":"The threshold-only update rule can leave the decoder permanently masked. The printed decoder is not a complete unmasking procedure on all admitted states; this does not establish an implementation hang.","primaryUrl":"https://arxiv.org/html/2602.06953","noteUrl":"/review?paper=review-neurips-2026-153138","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153139","title":"Your Language Model is Its Own Critic: Reinforcement Learning with Value Estimation from Actor’s Internal States","authors":["Yunho Choi","Jongwon Lim","woojin Ahn","Minjae Oh","Jeonghoon Shim","Yohan Jo"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Length normalization changes the baseline-cancellation argument","plainExplanation":"The unbiasedness argument uses an independent cross-rollout baseline for an unnormalized policy gradient, while Equation (13) weights the update by inverse response length. Consider equiprobable responses EOS and A–EOS, lengths one and two, rewards one and zero, and sequence scores \\(0.5\\) and \\(-0.5\\).","primaryUrl":"https://arxiv.org/html/2605.07579","noteUrl":"/review?paper=review-neurips-2026-153139","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153142","title":"Flash PD-SSM: Memory-Optimized Structured Sparse State-Space Models","authors":["Aleksandar Terzic","Francesco S. Carzaniga","Nicolas Menet","Yannick Biehl","Michael Hersche","Thomas Hofmann","Abbas Rahimi"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A column-one-hot transition requires scatter-add rather than gather","plainExplanation":"A column-one-hot transition requires scatter-add rather than gather. The matrix and kernel recurrences disagree; the result does not assess a compiled kernel.","primaryUrl":"https://arxiv.org/html/2605.19150","noteUrl":"/review?paper=review-neurips-2026-153142","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153144","title":"Test-Time Gradient Guidance of Flow Policies in Reinforcement Learning","authors":["Zhiyuan (Paul) Zhou","Andy Peng","Charles Xu","Qiyang Li","Jost Springenberg","Kevin Frans","Sergey Levine"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Removing singular-value scaling can change the guidance direction","plainExplanation":"A fixed-critic Gibbs tilt need not solve the trajectory-regularized control problem. The chunk target must bootstrap at the chunk endpoint. Removing singular-value scaling can change the guidance direction.","primaryUrl":"https://arxiv.org/html/2606.11087","noteUrl":"/review?paper=review-neurips-2026-153144","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153145","title":"SimSD: Simple Speculative Decoding in Diffusion Language Models","authors":["Junxia Cui","Haotian Ye","Runchu Tian","Hongcan Guo","Jinya Jiang","Haoru Li","Chaojie Ren","Yiming Huang","Kaijie Zhu","Zhongkai Yu","Kun Zhou","Jingbo Shang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The verifier mask can make the target conditional depend on the proposal","plainExplanation":"The exact speculative-correction claim requires a fixed target conditional, but the mask permits a two-hop path from the proposal being checked to its verifier representation. The allowed path \\(M_1\\leftarrow M_2\\leftarrow D_1\\) lets a verifier score depend on the drafted token it is meant to test.","primaryUrl":"https://arxiv.org/html/2606.02544","noteUrl":"/review?paper=review-neurips-2026-153145","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153147","title":"DirectUV: Image-Conditioned UV Texture Generation with Surface-Aware Positional Encoding","authors":["Jiantao Lin","Yingjie Xu","Mingzhi Sheng","Yangkai Wei","Hao CHEN","Yingcong Chen"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Rotary position encoding does not remove attention edges between charts","plainExplanation":"The no-SAPE explanation attributes absent direct cross-chart interaction to two-dimensional UV-grid RoPE. RoPE rotates queries and keys; it does not mask their attention edge.","primaryUrl":"https://arxiv.org/html/2609.34651","noteUrl":"/review?paper=review-neurips-2026-153147","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153150","title":"A general kernel framework for non-CND distance measures using $|\\mathcal{D}|$-dimensional sparse landmark embeddings","authors":["Marcus Noack","Maher B Alghalayini","Mark D Risser"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Bounded support overlap does not bound Gram condition number","plainExplanation":"Bounded support overlap does not bound Gram conditioning; sparsity alone does not prevent concentrated distances or uncovered-input collisions; sparse embeddings do not make explicit Gaussian Gram assembly linear; and separation by one radius does not ensure disjoint supports. The positive-semidefinite kernel construction and conditional support identities remain intact.","primaryUrl":"https://arxiv.org/html/2609.19083","noteUrl":"/review?paper=review-neurips-2026-153150","collection":"neurips-2026","severity":"major","issueTypes":["complexity_error","false_statement","incorrect_bound","invalid_inference","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-153152","title":"A mathematical theory of balancing relational generalization and memorization","authors":["Luke Cheng","Samuel Lippl"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"A mathematical theory of balancing relational generalization and memorization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22972","noteUrl":"/review?paper=review-neurips-2026-153152","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153153","title":"PRISM: Programming Interactive Scenes from Monocular Images for Embodied Simulation","authors":["Yumeng He","Yichen Song","Xiaotian Yang","Weijia Zhang","Zanwei Zhou","俊儒 宫","Xiaokang Yang","Yunbo Wang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"PRISM: Programming Interactive Scenes from Monocular Images for Embodied Simulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.24212","noteUrl":"/review?paper=review-neurips-2026-153153","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153154","title":"From “Weak” Signals to Strong Models: Preference Delta Aggregation with LoRA Merging","authors":["Qi Sun","Siyue Zhang","Yulin Chen","Yuxiang Xue","Ru Peng","Chen Zhao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"From “Weak” Signals to Strong Models: Preference Delta Aggregation with LoRA Merging","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.00357","noteUrl":"/review?paper=review-neurips-2026-153154","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153155","title":"Embedding Foundation Model Predictions in Discrete-Choice Models with Structural Guarantees","authors":["Yingshuo Wang","Xian Sun","Yanhang Li","Zhichao Fan","Zexin Zhuang"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"A nonzero cost derivative does not recover cost from predictions alone","plainExplanation":"Main-text Proposition 2 explicitly assumes that costs can be recovered from foundation predictions. Appendix A.2 cannot replace that assumption by a nonzero cost derivative: other varying inputs can give the same prediction at different costs. Even with recoverability, density of the correction class guarantees equal loss infima, not an open interval of exact minimizers. The appendix already states the valid limiting conclusion. Exact compensation requires closure under cost shifts or exact interpolation of the required utility differences on the training inputs.","primaryUrl":"https://arxiv.org/html/2606.26432","noteUrl":"/review?paper=review-neurips-2026-153155","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153157","title":"CAVE: A Structured Credit Assignment Approach for Fragmented Visual Evidence Reasoning","authors":["Tengda Guo","Jie Leng","Hanlei Li","Yaoyuan Liang","Qingyue Zhang","Dian Yang","Mingyu Zhang","Yuhua Fu","Shao-Lun Huang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"CAVE: A Structured Credit Assignment Approach for Fragmented Visual Evidence Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16416","noteUrl":"/review?paper=review-neurips-2026-153157","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153160","title":"Sequential Probabilistic Uncertainty Estimation for Parallel Multi-Agent Reasoning Systems","authors":["Tunyu Zhang","Zihao Zhao","Yusong Zhao","Haizhou Shi","Zhuohang Li","Haoxian Chen","Hao Wang","Dimitris Metaxas"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Sequential Probabilistic Uncertainty Estimation for Parallel Multi-Agent Reasoning Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.08901","noteUrl":"/review?paper=review-neurips-2026-153160","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153161","title":"Sink vs. diagonal patterns as mechanisms for attention switch and oversmoothing prevention","authors":["Peter Súkeník","Cristina Lopez Amado","Christoph Lampert","Marco Mondelli"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Sink vs. diagonal patterns as mechanisms for attention switch and oversmoothing prevention","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08453","noteUrl":"/review?paper=review-neurips-2026-153161","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153165","title":"Predicting directional flexibility in proteins","authors":["Vsevolod Viliuga","Leif Seute","Matteo Tadiello","Nicolas Wolf","Frauke Gräter","Arne Elofsson"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The coupling head lacks the constraints needed for a real correlation normalization","plainExplanation":"The coupling head lacks the constraints needed for a real correlation normalization. The printed coupling normalization lacks domain and correlation constraints; actual repository behavior is not inferred.","primaryUrl":"https://arxiv.org/html/2609.08474","noteUrl":"/review?paper=review-neurips-2026-153165","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153166","title":"BrainVista: Modeling Naturalistic Brain Dynamics as Multimodal Next-Token Prediction","authors":["Xuanhua Yin","Ted Zhao","Lina Yao","Weidong Cai"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The parcel-map description needs a distinct metric","plainExplanation":"Each correlation already aggregates over all parcels. Averaging over time leaves one scalar, with no remaining parcel index to map onto the cortical surface.","primaryUrl":"https://arxiv.org/html/2602.04512","noteUrl":"/review?paper=review-neurips-2026-153166","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-153168","title":"Flow-Based Conformal Predictive Distributions","authors":["Trevor Harris"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Continuous score functions can still produce tied empirical scores","plainExplanation":"Exact empirical-grid calibration needs atomless score laws or tie handling. Boundary mixing also lacks a probability kernel on the outer-quantile interval where the required order statistic is undefined and the usual infinite-threshold completion has an empty boundary.","primaryUrl":"https://arxiv.org/html/2602.07633","noteUrl":"/review?paper=review-neurips-2026-153168","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153170","title":"Select-then-differentiate: Solving Bilevel Optimization with Manifold Lower-level Solution Sets","authors":["Saeed Masiha","Zebang Shen","Negar Kiyavash","Niao He"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Select-then-differentiate: Solving Bilevel Optimization with Manifold Lower-level Solution Sets","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09209","noteUrl":"/review?paper=review-neurips-2026-153170","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153174","title":"CANDO: Cooperative Agentic Network for Layout Design Optimization","authors":["Athanasios Masouris","ZHENG JING","Benjamin S Chandler","Hadi Jamali-Rad"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"CANDO: Cooperative Agentic Network for Layout Design Optimization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.10044","noteUrl":"/review?paper=review-neurips-2026-153174","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153175","title":"Augmented Equivariant Mesh Networks for Anatomical Segmentation","authors":["Daniel Saragih"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Identically initialized virtual nodes remain identical under the stated updates","plainExplanation":"All virtual nodes start with the same learned feature seed and centroid coordinate, then use shared message and coordinate updates. At initialization every virtual node has the same state.","primaryUrl":"https://arxiv.org/html/2605.08172","noteUrl":"/review?paper=review-neurips-2026-153175","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153179","title":"Measure-to-measure Regression with Transformers","authors":["Matthew Vandergrift","Martha White","Yury Polyanskiy","Philippe Rigollet","Lazar Atanackovic"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Regular scalar current-state ODE flows cannot represent reflection","plainExplanation":"Regular scalar current-state ODE flows cannot represent reflection. The maximal-expressivity transfer has a scope gap; unrestricted or source-conditioned transport representations are not refuted. An isolated path collision alone is not asserted to force positive integrated training loss.","primaryUrl":"https://arxiv.org/html/2605.28075","noteUrl":"/review?paper=review-neurips-2026-153179","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153182","title":"Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies","authors":["Yi Wen","Derong Xu","Pengyue Jia","Yichao Wang","Yingyi Zhang","Maolin Wang","Junyi Li","Wenlin Zhang","Xiaopeng Li","Yong Liu","Xiangyu Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An independent-pair cosine ceiling does not impose a retrieval ranking tradeoff","plainExplanation":"The independent-pair cosine bound does not imply the claimed shared-encoder retrieval tradeoff: separated categories can still have perfect relevance ranking. An intermediate centroid inequality also fails for unequal lengths with antipodal directions.","primaryUrl":"https://arxiv.org/html/2610.11573","noteUrl":"/review?paper=review-neurips-2026-153182","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153185","title":"The Design Space of Tri-Modal Masked Diffusion Models","authors":["Louis Bethune","Victor Guilherme Turrisi da Costa","Bruno Mlodozeniec","Pau Rodriguez","Lokesh Boominathan","Nikhil Bhendawade","Amitis Shidani","Joris Pelemans","Theo X. Olausson","R Devon Hjelm","Paul Dixon","Joao Monteiro","Pierre Ablin","Vishnu Banna","Arno Blaas","Nick Henderson","Kari Noriy","Dan Busbridge","Marco Cuturi","Joshua Susskind","Irina Belousova","Luca Zappella","Russell Webb","Jason Ramapuram"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Random masked-count normalization changes the ELBO time weighting","plainExplanation":"Random masked-count normalization changes the ELBO time weighting. The displayed unbiased-ELBO interpretation is false; the result does not determine which normalization was executed.","primaryUrl":"https://arxiv.org/html/2602.21472","noteUrl":"/review?paper=review-neurips-2026-153185","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153186","title":"Covariate-Adjusted Deep Causal Learning for Heterogeneous Panel Data Models","authors":["Guanhao Zhou","Yuefeng Han","Xiufan Yu"],"year":2026,"category":"Statistical learning","status":"unresolved","issueTitle":"Covariate-Adjusted Deep Causal Learning for Heterogeneous Panel Data Models","plainExplanation":"Observed pre-treatment data need not identify loadings on unseen post-treatment factors. The explicit two-world construction defeats the printed v1 consistency rate even when the noiseless control inverse is exact.","primaryUrl":"https://arxiv.org/html/2505.20536","noteUrl":"/review?paper=review-neurips-2026-153186","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-153187","title":"Mixture of Layers: Dynamic Layer Routing for Visual Reasoning","authors":["Jeonghwan Kim","Sofia Stoica","Jiwan Chung","Ansel Blume","Hyeonjeong Ha","Zhenhailong Wang","Xin Dong","Heng Ji"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Mixture of Layers: Dynamic Layer Routing for Visual Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.09440","noteUrl":"/review?paper=review-neurips-2026-153187","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153189","title":"Backdoor Purification for LoRA-Tuned LLMs via Null-Space Projection","authors":["Jianwei Li","Jung-Eun Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Backdoor Purification for LoRA-Tuned LLMs via Null-Space Projection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.00685","noteUrl":"/review?paper=review-neurips-2026-153189","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153191","title":"PEIRA: Learning Predictive Encoders through Inter-View Regressor Alignment","authors":["Michael Arbel","Basile Terver","Jean Ponce"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The recommended window includes an endpoint that removes every mode","plainExplanation":"The recommended regularization interval includes $\\lambda=1$, where all canonical modes—including the constant mode—can vanish. The strict-window result already excludes that endpoint.","primaryUrl":"https://arxiv.org/html/2605.17671","noteUrl":"/review?paper=review-neurips-2026-153191","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153192","title":"Revisiting Diffusion Model Predictions Through Dimensionality","authors":["Qing Jin","Chaoyang Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Revisiting Diffusion Model Predictions Through Dimensionality","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.21419","noteUrl":"/review?paper=review-neurips-2026-153192","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153193","title":"BFS-PO: Best-First Search for Large Reasoning Models","authors":["Fiorenzo Parascandolo","Wenhui Tan","Enver Sangineto","Xiaoyi Yu","Ruihua Song","Rita Cucchiara"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The best-correct selector compares against incorrect short outputs","plainExplanation":"The best-correct selector compares against incorrect short outputs. The formal selector is inconsistent with its intended shortest-correct operation; empirical search is not disproved.","primaryUrl":"https://arxiv.org/html/2602.14917","noteUrl":"/review?paper=review-neurips-2026-153193","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153196","title":"DUET: Optimize Token-Budget Allocation for Reinforcement Learning with Verifiable Rewards","authors":["Haoyu Hu","Xuandong Zhao","Nori Jacoby","Xuhai &amp;quot;Orson&amp;quot; Xu"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"DUET: Optimize Token-Budget Allocation for Reinforcement Learning with Verifiable Rewards","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08441","noteUrl":"/review?paper=review-neurips-2026-153196","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153198","title":"Scalable Minimal-Change Learning for Controllable Image Editing","authors":["Shuo Chen","Fengming Huang","Yu Yao","Mingming Gong","Tongliang Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Scalable Minimal-Change Learning for Controllable Image Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06021","noteUrl":"/review?paper=review-neurips-2026-153198","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153201","title":"Effective Biological Representation Learning by Masking Gene Expression","authors":["Kian Kenyon-Dean","Alina Selega","Ihab Bendidi","Jordan M Sorokin","Luca Bertinetto","David Errington","Hayley Donnella","Oren Kraus"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Effective Biological Representation Learning by Masking Gene Expression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31562","noteUrl":"/review?paper=review-neurips-2026-153201","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153203","title":"When and Why Does Multi-Agent Debate Fail and Does It Really Underperform?","authors":["Yongqiang Chen","Gang Niu","James Cheng","Bo Han","Masashi Sugiyama"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Positive mutual information need not strictly improve classification risk","plainExplanation":"The judge statistic mixes prior-inclusive and prior-free likelihood conventions and can count prior odds twice. Separately, positive information does not guarantee strict classification improvement when all updated posteriors stay on the same side of the decision threshold.","primaryUrl":"https://arxiv.org/html/2510.20963","noteUrl":"/review?paper=review-neurips-2026-153203","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153205","title":"Coupled Integral PINN for Discontinuity","authors":["Yeping Wang","Shihao Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Coupled Integral PINN for Discontinuity","plainExplanation":"The universal strong-residual barrier is false for Euler contacts. The full resolved CI-PINN objective can still diverge through its entropy penalty.","primaryUrl":"https://arxiv.org/html/2411.11276","noteUrl":"/review?paper=review-neurips-2026-153205","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","invalid_inference"],"classifications":["critical"]},{"slug":"review-neurips-2026-153206","title":"Stochastic Optimization with Random Search","authors":["El Mahdi Chayti","Taha EL BAKKALI EL KADI","Omar Saadi","Martin Jaggi"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Stochastic Optimization with Random Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.15610","noteUrl":"/review?paper=review-neurips-2026-153206","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153207","title":"Safe Evolution with Circuit Anchors","authors":["Yan Liu","Jie Fu","Tsung-Yi Ho"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Safe Evolution with Circuit Anchors","plainExplanation":"The safety theorem permits fixed circuit activations while trainable output weights reverse refusal probabilities. A valid one-step example changes refusal from 99% to 1% with zero anchor loss.","primaryUrl":"https://arxiv.org/html/2608.05158","noteUrl":"/review?paper=review-neurips-2026-153207","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-153208","title":"Point Cloud Sequence Encoding for Material-conditioned Graph Network Simulators","authors":["Philipp Dahlinger","Balázs Gyenes","Niklas Freymuth","Luca Geminiani","Tobias Würth","Johannes Mitsch","Nadja Klein","Luise Kärger","Gerhard Neumann"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Point Cloud Sequence Encoding for Material-conditioned Graph Network Simulators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20978","noteUrl":"/review?paper=review-neurips-2026-153208","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153210","title":"Divide et Calibra: Multiclass Local Calibration via Vector Quantization","authors":["Cesare Barbera","Lorenzo Perini","Giovanni De Toni","Andrea Passerini","Andrea Pugnana"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Divide et Calibra: Multiclass Local Calibration via Vector Quantization","plainExplanation":"No issue was found in the checked claims. The invalid shared-point vector mean-value step has been removed. The replacement integral identity is exact under the current differentiability assumptions.","primaryUrl":"https://arxiv.org/html/2605.21060","noteUrl":"/review?paper=review-neurips-2026-153210","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153215","title":"X-AVDD: Cross-Attentive Audio-Visual Dataset Distillation","authors":["Saumyaranjan Mohanty","Aravind Reddy","Konda Reddy Mopuri"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"X-AVDD: Cross-Attentive Audio-Visual Dataset Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.13687","noteUrl":"/review?paper=review-neurips-2026-153215","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153217","title":"Budgeting Discretion: Theory and Evidence on Street-Level Decision-Making","authors":["Gaurab Pokharel","Sanmay Das","Patrick J Fowler"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Budgeting Discretion: Theory and Evidence on Street-Level Decision-Making","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.10039","noteUrl":"/review?paper=review-neurips-2026-153217","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153220","title":"Graph Mamba Operator: A Latent Simulator for Interacting Particle Systems","authors":["Karn Tiwari","Niladri Dutta","Prathosh AP","N M Anoop Krishnan"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The graph factors belong inside the exact flow exponential","plainExplanation":"The exact-flow derivation puts graph factors outside an exponential, and the rollout formulas separate noncommuting gates and mixers. Input dependence also does not preclude finite memory, while the infinite-series representation needs a prehistory boundary condition. A generic Hadamard norm bound is false, although the actual diagonal-gating stability bound has a direct repair.","primaryUrl":"https://arxiv.org/html/2606.09432","noteUrl":"/review?paper=review-neurips-2026-153220","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153221","title":"Rethinking Latency Denial-of-Service: Attack the LLM Serving Framework, Not the Model","authors":["Tianyi Wang","Huawei Fan","Yuanchao Shu","Peng Cheng","Cong Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Average utilization does not guarantee an unused admission slot","plainExplanation":"The blocking-time statement omits head cancellation, which can release followers earlier. The utilization-based TTFT bound also fails under permitted bursts: low average load does not ensure an available admission slot.","primaryUrl":"https://arxiv.org/html/2602.07878","noteUrl":"/review?paper=review-neurips-2026-153221","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153223","title":"Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders","authors":["Chungpa Lee","Jihoon Kwon","Kyle Min","Jy-yong Sohn"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Nonnegative codes invalidate the unconstrained eigenvector optimum","plainExplanation":"The eigenvector construction can be suboptimal under nonnegative codes when its feature projections have opposite signs. The regularized global-optimum claims also omit the decoder-column normalization imposed by their proofs.","primaryUrl":"https://arxiv.org/html/2606.29888","noteUrl":"/review?paper=review-neurips-2026-153223","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153224","title":"LoRAtorio: An intrinsic approach to LoRA Skill Composition","authors":["Niki Foteinopoulou","Ignas Budvytis","Stephan Liwicki"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The right LoRA factor acts first on the input","plainExplanation":"Appendix B sets \\(A\\in\\mathbb R^{d\\times r}\\), \\(B\\in\\mathbb R^{r\\times k}\\), but treats \\(ABx\\) as applying \\(A\\) first and identifies the input-active subspace with its row space. For \\(x\\in\\mathbb R^k\\), the defined update is \\(ABx=A(Bx)\\).","primaryUrl":"https://arxiv.org/html/2508.11624","noteUrl":"/review?paper=review-neurips-2026-153224","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153225","title":"Stability and Diversity of Networked Self-Consuming Generative Ecosystems","authors":["Xiukun Wei","Yang Zhang","Xueru Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Parent-to-data gaps do not control a model’s own representation residual","plainExplanation":"The global ecosystem update can jump between maximizers despite the stated local spectral condition. The feature-residual estimate also needs an own-model mismatch term or self-loop assumption. Correcting the Hessian signs restores the intended local derivative estimate, while global convergence still requires an isolating condition on the chosen maximizer.","primaryUrl":"https://arxiv.org/html/2610.09409","noteUrl":"/review?paper=review-neurips-2026-153225","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153228","title":"EvoCUA: Evolving Computer Use Agents via Learning from Scalable Synthetic Experience","authors":["Taofeng Xue","Chong Peng","Mianqiu Huang","Linsen Guo","Tiancheng Han","Haozhe Wang","Xiaocheng Zhang","Xin Yang","Dengchang Zhao","Jinrui Ding","Xiandi Ma","Yuchen Xie","Peng Pei","Xunliang Cai","Xipeng Qiu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EvoCUA: Evolving Computer Use Agents via Learning from Scalable Synthetic Experience","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.15876","noteUrl":"/review?paper=review-neurips-2026-153228","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153229","title":"Mutable Transcripts: Mitigating Context Pollution through Editable Conversation State","authors":["Dan Barry","Andrew Hines"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mutable Transcripts: Mitigating Context Pollution through Editable Conversation State","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.31354","noteUrl":"/review?paper=review-neurips-2026-153229","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153230","title":"Reasoning with Sampling: Cutting at Decision Points","authors":["Felix Zhou","Anay Mehrotra","Quanquan C Liu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The first branch lies outside the proposal-comparability hypothesis","plainExplanation":"The entropy-cut mixing proof uses proposal comparability at the first branch, where the stated condition does not apply. A two-leaf early-branch example has arbitrarily slow mixing despite satisfying all later comparisons.","primaryUrl":"https://arxiv.org/html/2605.30327","noteUrl":"/review?paper=review-neurips-2026-153230","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153231","title":"Advancing Affordance-Grounded Creative Tool Use in Large Multimodal Models","authors":["Cheng Qian","Hyeonjeong Ha","Jiayu Liu","Jeonghwan Kim","Emre Can Acikgoz","Bingxuan Li","Kunlun Zhu","Jiateng Liu","Aditi Tiwari","Zhenhailong Wang","Xiusi Chen","Heng Ji"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Advancing Affordance-Grounded Creative Tool Use in Large Multimodal Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26396","noteUrl":"/review?paper=review-neurips-2026-153231","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153233","title":"SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation","authors":["Jie Sun","Mao Zheng","Mingyang Song","Qiyong Zhong","Yilin Cheng","Bichuan Feng","Pengfei Liu","Junfeng Fang","Xiang Wang"],"year":2026,"category":"Transformers & sequence models","status":"unresolved","issueTitle":"SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation","plainExplanation":"No issue was found in the checked claims. The overlap-graph components end exactly at shared token boundaries, so no finer jointly tokenizable partition exists. The KL claim follows by factoring each unit's probability into its group mass and conditional probability. The square-table example correctly gives 8!/4=10080 and identifies the baselines' combinatorial errors.","primaryUrl":"https://arxiv.org/html/2605.07711","noteUrl":"/review?paper=review-neurips-2026-153233","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153234","title":"Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries","authors":["Kun Zeng","Yu Huo","Siyu Zhang","Zi Ye","Yuecheng Zhuo","haoyue liu","YuQuan Lu","Junhao Wen","Xiaoying Tang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06978","noteUrl":"/review?paper=review-neurips-2026-153234","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153235","title":"Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation","authors":["Duong Nguyen","Nghia Hoang","Hang T Nguyen","Thanh Trung Huynh","Phi Le Nguyen","Minh Do"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Independent assignments omit the occupied-column penalty","plainExplanation":"The stated MAP objective penalizes the number of occupied concept columns, but Equation (15) updates assignments independently before pruning only empty columns. Moving the final assignment out of a column changes the penalty by \\(\\lambda\\); this coupling is absent from an independent per-pattern update.","primaryUrl":"https://arxiv.org/html/2610.00678","noteUrl":"/review?paper=review-neurips-2026-153235","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153237","title":"Causal Multi-Task Demand Learning","authors":["Varun Gupta","Vijay Kamble"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Causal Multi-Task Demand Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09969","noteUrl":"/review?paper=review-neurips-2026-153237","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153240","title":"Causal Discovery Under Hard Selection Bias: A New Robust Score-Matching Approach","authors":["Yiwen Qiu","Francesco Montagna","Shimeng Huang","Francesco Locatello"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Restricted nonlinear ANMs need not have the proposed conditional rank gap","plainExplanation":"Smooth restricted nonlinear and linear non-Gaussian additive-noise models can give rank one to both a sink and a nonsink at every conditioning value. The nonlinear rank-gap guarantee therefore needs an additional nondegeneracy condition. In the linear example, pooling conditional covariances restores the rank gap, so the practical averaged test is not refuted. The estimation appendix also needs the standard conversion from a noise predictor to the smoothed score.","primaryUrl":"https://arxiv.org/html/2610.18075","noteUrl":"/review?paper=review-neurips-2026-153240","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","missing_assumption","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-153242","title":"Quality-Diversity Optimization as Multi-Objective Optimization","authors":["Xi Lin","Ping Guo","Yilu Liu","Bo Xue","Qingfu Zhang","Jianyong Sun"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The smooth Tchebycheff reference offset is scaled incorrectly","plainExplanation":"The smooth Tchebycheff formula loses nonzero reference offsets in its limit. STCH-Set also fails the claimed diminishing-improvement property, and SoM-optimal sets can include redundant dominated members.","primaryUrl":"https://arxiv.org/html/2602.00478","noteUrl":"/review?paper=review-neurips-2026-153242","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153246","title":"Variance Reduction for Expectations with Diffusion Teachers","authors":["Jesse Bettencourt","Xindi Wu","Matan Atzmon","James Lucas","Jonathan Lorraine"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The pair estimator and discrete sampling convention need consistent probability normalization","plainExplanation":"Equation (65) halves this Horvitz–Thompson estimator. Equation (66), centered at the target sum, is consequently a mean-squared error rather than its variance.","primaryUrl":"https://arxiv.org/html/2605.21489","noteUrl":"/review?paper=review-neurips-2026-153246","collection":"neurips-2026","severity":"repairable","issueTypes":["mathematical_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153248","title":"Towards a Universal Causal Reasoner","authors":["Qirun Dai","Xiao Liu","Jiawei Zhang","Dylan Zhang","Hao Peng","Chenhao Tan"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Towards a Universal Causal Reasoner","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.24873","noteUrl":"/review?paper=review-neurips-2026-153248","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153249","title":"Consistent Geometric Deep Learning via Hilbert Bundles and Cellular Sheaves","authors":["Kartik Tandon","Julian J Gould","Tanishq Bhatia","Francesca Dominici","Alejandro Ribeiro","Claudio Battiloro"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Section-dependent approximation sequences do not establish one common core","plainExplanation":"The common-core convergence argument fixes one approximation sequence after constructing section-dependent sequences, leaving a gap before the resolvent step. Other local proof corrections concern reciprocal normalization, a concentration sign, an unsupported diagonal rate and the dimension symbol in a heat-kernel statement; the latter is correct in later uses.","primaryUrl":"https://arxiv.org/html/2605.06395","noteUrl":"/review?paper=review-neurips-2026-153249","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153250","title":"Rank Is Not Capacity: Spectral Occupancy for Latent Graph Models","authors":["Nikolaos Nakis","Panagiotis Promponas","Konstantinos Tsirkas","Katerina Mamali","Eftychia Makri","Leandros Tassiulas","Nicholas A Christakis"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Rank Is Not Capacity: Spectral Occupancy for Latent Graph Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11142","noteUrl":"/review?paper=review-neurips-2026-153250","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153251","title":"A Matter of Interest: Understanding Interestingness Judgments of Math Problems in Humans and Language Models","authors":["Shubhra Mishra","Yuka Machino","Gabriel Poesia","Albert Q. Jiang","Joy Hsu","Adrian Weller","Challenger Mishra","David Broman","Josh Tenenbaum","Mateja Jamnik","Cedegao (Ced) Zhang","Katie Collins"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"A Matter of Interest: Understanding Interestingness Judgments of Math Problems in Humans and Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2511.08548","noteUrl":"/review?paper=review-neurips-2026-153251","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153255","title":"Gradient Routing Localizes and Removes Unintended Behaviors in RL","authors":["Jake Ward","Shawn Hu","Aria Wong","Nathan Hu","Bryce Woodworth","Alexander Turner","Alex Cloud"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Gradient Routing Localizes and Removes Unintended Behaviors in RL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.17062","noteUrl":"/review?paper=review-neurips-2026-153255","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153258","title":"A2Eval: Agentic and Automated Evaluation for Embodied Brain","authors":["Shuai Zhang","Jiayu Hu","Zijie Chen","Zeyuan Ding","Yi Zhang","Yingji Zhang","Ziyi Zhou","Junwei Liao","Shengjie Zhou","Yong Dai","Zhenzhong Lan","Xiaozhu Ju"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"A2Eval: Agentic and Automated Evaluation for Embodied Brain","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.01640","noteUrl":"/review?paper=review-neurips-2026-153258","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153259","title":"Ultra Flash: Scaling Real-Time Streaming Video Generation to High Resolutions","authors":["Luxury","Jie Huang","Zihao Fan","Xiaoxiao Ma","Yuming Li","Siming Fu","Jun-hao Zhuang","Zeyue Xue","Haoran Li","haoyang huang","Nan Duan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Four mathematical definitions or justifications need clarification","plainExplanation":"Causality of both variants does not make them the same causal network.","primaryUrl":"https://arxiv.org/html/2606.09150","noteUrl":"/review?paper=review-neurips-2026-153259","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_ambiguity","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153260","title":"Profit Maximization in Bilateral Trade against a Smooth Adversary","authors":["Simone Di Gregorio","Paul Duetting","Federico Fusco","Chris Schwiegelshohn"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Profit Maximization in Bilateral Trade against a Smooth Adversary","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12664","noteUrl":"/review?paper=review-neurips-2026-153260","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153262","title":"Beyond Normal References: Discriminative Few-Shot Anomaly Detection","authors":["Huan Wang","Jun Shen","Jun Yan","Guansong Pang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Classifier VC dimension does not control the full real-valued loss class","plainExplanation":"Theorem B.1 uses the ordinary classifier VC dimension \\(d\\) in a quantitative Rademacher bound for the composite episode-loss class. On \\(N\\) episode identifiers, allow independently chosen scores \\(0.6\\) or \\(0.9\\), plus the constant score \\(0.1\\).","primaryUrl":"https://arxiv.org/html/2605.23231","noteUrl":"/review?paper=review-neurips-2026-153262","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153264","title":"CompilerKV: Risk-Adaptive KV Cache Compression via Offline Experience Compilation","authors":["Ning Yang","Chengzhi Wang","Yibo Liu","Baoliang Tian","Haijun Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CompilerKV: Risk-Adaptive KV Cache Compression via Offline Experience Compilation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.08686v3","noteUrl":"/review?paper=review-neurips-2026-153264","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153267","title":"EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation","authors":["Ruozhen He","Meng Wei","Ziyan Yang","Vicente Ordonez"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15199","noteUrl":"/review?paper=review-neurips-2026-153267","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153268","title":"Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety Routing","authors":["Maryam Hashemzadeh Barvarz","Jerry Huang","Minseon Kim","Marc-Alexandre Côté","Sarath Chandar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety Routing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.00686","noteUrl":"/review?paper=review-neurips-2026-153268","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153269","title":"Permit: Permission-Aware Representation Intervention for Controlled Generation in Large Language Models","authors":["Pengcheng Sun","Lan Zhang","Zhaopeng Zhang","Jiewei Lai","Chen Tang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Permit: Permission-Aware Representation Intervention for Controlled Generation in Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09480","noteUrl":"/review?paper=review-neurips-2026-153269","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153273","title":"EMA-Nesterov: Stabilizing Nesterov's Lookahead for Accelerated Deep Learning Optimization","authors":["Chung-Yiu Yau","Dawei Li","Athanasios Glentis","Valentyn Boreiko","Hoi-To Wai","Mingyi Hong"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"EMA-Nesterov: Stabilizing Nesterov's Lookahead for Accelerated Deep Learning Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25395","noteUrl":"/review?paper=review-neurips-2026-153273","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153275","title":"Reading Between the Dots: Decoding Hidden Computation across Filler Tokens","authors":["Kaley Brauer","Claudio Mayrink Verdun","Samuel Marks"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Reading Between the Dots: Decoding Hidden Computation across Filler Tokens","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.03502","noteUrl":"/review?paper=review-neurips-2026-153275","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153276","title":"Seirênes: Adversarial Self-Play with Evolving Distractions for LLM Reasoning","authors":["Chi Zhang","Haibo Qiu","Qiming ZHANG","Yufei Xu","Xinbo Gao","Jing Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Filtering zero-advantage hint groups makes the mastery test vacuous or impossible","plainExplanation":"Filtering zero-advantage hint groups makes the mastery test vacuous or impossible. The printed predicate does not implement the stated hint-robust retirement criterion.","primaryUrl":"https://arxiv.org/html/2605.11636","noteUrl":"/review?paper=review-neurips-2026-153276","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153277","title":"Ordinary Least Squares as an Attention Mechanism","authors":["Philippe Goulet Coulombe"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Ordinary Least Squares as an Attention Mechanism","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2504.09663","noteUrl":"/review?paper=review-neurips-2026-153277","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153278","title":"Partition Tree: Conditional Density Estimation over General Outcome Spaces","authors":["Felipe Angelim","Alessandro Leite"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Partition Tree: Conditional Density Estimation over General Outcome Spaces","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04042","noteUrl":"/review?paper=review-neurips-2026-153278","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153279","title":"AnyMo: Scaling Any-Modality Conditional Motion Generation with Masked Modeling","authors":["Yiheng Li","Zhuo Li","RuiBing Hou","Yingjie Chen","Hong Chang","Hao Liu","Shiguang Shan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Quantizer indices cannot be subtracted directly as latent approximations","plainExplanation":"Equation (1) returns integer grid indices from zero to \\(L-1\\), and Equation (2) subtracts those outputs directly in recursive latent residual refinement. For \\(z=1,L=4\\), the printed map outputs two, leaving residual minus one.","primaryUrl":"https://arxiv.org/html/2605.29488","noteUrl":"/review?paper=review-neurips-2026-153279","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153280","title":"Sublinear Time Quantum Sensitivity Sampling","authors":["Zhao Song","David Woodruff","Lichen Zhang"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The leverage oracle uses the wrong spectral inverse factor","plainExplanation":"The coreset sampler never uses its computed precision-dependent budget. The printed sampling rule lacks its accuracy mechanism; corrected coreset existence is not ruled out.","primaryUrl":"https://arxiv.org/html/2509.16801","noteUrl":"/review?paper=review-neurips-2026-153280","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153281","title":"When Does Non-Uniform Replay Matter in Reinforcement Learning?","authors":["Michal Korniak","Mikołaj Czarnecki","Yarden As","Piotr Miłoś","Pieter Abbeel","Michal Nauman"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The printed inverse CDF excludes the newest buffer element","plainExplanation":"The inverse-CDF formula subtracts one after flooring, then clips, while claiming full support on the geometrically weighted buffer indices. For \\(q=2,n=4\\), let \\(X=\\log_q(1+U(q^n-1))\\), with \\(U\\in[0,1)\\).","primaryUrl":"https://arxiv.org/html/2605.10236","noteUrl":"/review?paper=review-neurips-2026-153281","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153282","title":"Assign and Add: A Mechanistic Study of Compositional Arithmetic","authors":["Brady Exoo","Alberto Bietti","John Sous"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The attention initialization cannot satisfy all three stated weight orders","plainExplanation":"The attention initialization cannot satisfy all three stated weight orders. The theorem’s printed premise is inconsistent; its one-step conclusion admits the stated upper-bound-only repair.","primaryUrl":"https://arxiv.org/html/2605.31497","noteUrl":"/review?paper=review-neurips-2026-153282","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153283","title":"Rescaled Asynchronous SGD: Optimal Distributed Optimization under Data and System Heterogeneity","authors":["Ammar Mahran","Artavazd Maranjyan","Peter Richtarik"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Rescaled Asynchronous SGD: Optimal Distributed Optimization under Data and System Heterogeneity","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13434","noteUrl":"/review?paper=review-neurips-2026-153283","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153284","title":"The Reciprocity Gradient","authors":["Yue Lin","Pascal Poupart","Shuhui Zhu","Dan Qiao","Wenhao Li","Yuan Liu","Hongyuan Zha","Baoxiang Wang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Balanced schedules do not make the current pair independent of reputation","plainExplanation":"The expected-payoff calculation incorrectly separates balanced-schedule interaction indicators from reputation history. The virtual-gradient guarantee also needs derivative control rather than output error alone, and an undirected-pair coefficient misses a factor of two.","primaryUrl":"https://arxiv.org/html/2605.08323","noteUrl":"/review?paper=review-neurips-2026-153284","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153286","title":"BAT3R: Robust Online 3D Reconstruction with Bayesian Adaptive State Updates","authors":["Lijian Li","Yuanpeng He","Yudian Zheng","Chung-ju Huang","Linyu Li","Dongming Jin","Wenpin Jiao","Zhi Jin","Chi-Man Pun"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"BAT3R: Robust Online 3D Reconstruction with Bayesian Adaptive State Updates","plainExplanation":"The valid triangle inequality only bounds the output by its two endpoint radii. It does not constrain an unrestricted candidate.","primaryUrl":"https://arxiv.org/html/2610.16330","noteUrl":"/review?paper=review-neurips-2026-153286","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","mathematical_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153290","title":"A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression","authors":["JinCheng Ren","Siwei Wu","Yizhi Li","Zhu","Shu XU","Boyu Feng","Ruibin Yuan","Wei Zhang","Riza Batista-Navarro","Min Gao","Yan Bai","Jian Yang","Chenghua Lin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.19572","noteUrl":"/review?paper=review-neurips-2026-153290","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153292","title":"Recursive Multi-Agent Systems","authors":["Jiaru Zou","Rui Pan","Ruizhong Qiu","Pan Lu","Shizhe Diao","Jindong Jiang","Hanghang Tong","Tong Zhang","Markus Buehler","Jingrui He","James Zou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The end-to-end complexity omits final vocabulary decoding","plainExplanation":"The reverse triangle inequality yields only \\(\\|J\\|\\geq1-C(1+\\sqrt{\\log(1/\\delta)/d_h})\\). An unspecified constant \\(C\\) need not leave a positive lower bound. The admissible deterministic norm example \\(A=-I\\) cancels the identity exactly, showing the quoted norm control alone is insufficient. This is a proof gap. The deterministic cancellation example is not a high-probability counterexample to independent Kaiming initialization or the final gradient theorem.","primaryUrl":"https://arxiv.org/html/2604.25917","noteUrl":"/review?paper=review-neurips-2026-153292","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153293","title":"A Comprehensive View of Fairness through Distributional Stability","authors":["Gayane Taturyan","Charlotte Laclau","Stephan Clémençon"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The fairness equivalence cannot divide by zero perturbation variance","plainExplanation":"The fairness equivalence cannot divide by zero perturbation variance. The boundary-law equivalence is false; later uniform perturbation choices avoid this degeneracy.","primaryUrl":"https://arxiv.org/html/2609.37061","noteUrl":"/review?paper=review-neurips-2026-153293","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153294","title":"VIGOR: Zero-Shot Visual Generalization via Latent-Space Consistency in Model-Based Reinforcement Learning","authors":["Mingyu Park","Samyeul Noh","Hyun Myung","Donghwan Lee"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Disjoint one-view batches do not contain the same-frame encoding pair","plainExplanation":"Disjoint one-view batches do not contain the same-frame encoding pair. The printed dataflow and equal-sample efficiency claim are incompatible; paired encoding itself is feasible.","primaryUrl":"https://arxiv.org/html/2610.02801","noteUrl":"/review?paper=review-neurips-2026-153294","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153295","title":"TimeClaw: A Time-Series AI Agent with Exploratory Execution Learning","authors":["Hangchen Liu","Dongyuan Li","Renhe Jiang","Jiewen Deng","Weiwei Ye","Yoshihide Sekimoto"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TimeClaw: A Time-Series AI Agent with Exploratory Execution Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10038","noteUrl":"/review?paper=review-neurips-2026-153295","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153297","title":"LEVDA: Latent Ensemble Variational Data Assimilation via Differentiable Dynamics","authors":["Phillip Si","Peng Chen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"LEVDA: Latent Ensemble Variational Data Assimilation via Differentiable Dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.19406","noteUrl":"/review?paper=review-neurips-2026-153297","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153300","title":"Unifying Goal-Conditioned RL and Unsupervised Skill Learning via Control-Maximization","authors":["Alireza Modirshanechi","Benjamin Eysenbach","Peter Dayan","Eric Schulz"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Unifying Goal-Conditioned RL and Unsupervised Skill Learning via Control-Maximization","plainExplanation":"No issue was found in the checked claims. The current deterministic hypothesis excludes the stochastic-feedback counterexample. The general stochastic claim has been narrowed rather than retained.","primaryUrl":"https://arxiv.org/html/2605.06145","noteUrl":"/review?paper=review-neurips-2026-153300","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153303","title":"Outbidding and Outbluffing Elite Humans: Mastering Liar’s Poker via Self-Play and Reinforcement Learning","authors":["Richard Dewey","Janos Botyanszki","Ciamac C Moallemi","Andrew Zheng"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A legal 66-action round exceeds the stated maximum of 53","plainExplanation":"A legal three-player 3x3 round contains 27 increasing bids and 39 challenges, for 66 actions. Normal raises reset rebid rights exactly as Appendix B permits. This contradicts the maximum of 53 in Table 1. The construction gives a lower bound; it does not independently certify an exact maximum or imply a defect in agent performance.","primaryUrl":"https://arxiv.org/html/2511.03724","noteUrl":"/review?paper=review-neurips-2026-153303","collection":"neurips-2026","severity":"repairable","issueTypes":["complexity_error","false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153304","title":"One World, Dual Timeline: Decoupled Spatio-Temporal Gaussian Scene Graph for 4D Cooperative Driving Reconstruction","authors":["Yulong Chen","Xiaoyun Dong","Haoyu Zhang","Zongxian Yang","Lewei Xie","Xinke Li","Yifan Zhang","Kai Wang","Jianping Wang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Different capture times can still give parallel gradients","plainExplanation":"The universal capture-time interference claim fails when both captures lie on the same side of the anchor. The per-source Fisher application also assumes positive definiteness despite a common two-dimensional projection Jacobian that forces a null direction.","primaryUrl":"https://arxiv.org/html/2605.07910","noteUrl":"/review?paper=review-neurips-2026-153304","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153308","title":"Omni-Safety under Cross-Modality Conflict: Vulnerabilities, Dynamic Mechanisms and Efficient Alignment","authors":["Kun Wang","Zherui Li","Zhenhong Zhou","Jie Zhang","Yitong Zhang","Yan Mi","Kun Yang","Yiming Zhang","Zhongxiang Sun","Qiankun Li","Yang Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Omni-Safety under Cross-Modality Conflict: Vulnerabilities, Dynamic Mechanisms and Efficient Alignment","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.10161","noteUrl":"/review?paper=review-neurips-2026-153308","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153310","title":"D2D: Detector-to-Differentiable Critic for Improved Numeracy in Text-to-Image Generation","authors":["Nobline Yoo","Olga Russakovsky","Ye Zhu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A radial norm objective does not identify a Gaussian output distribution","plainExplanation":"A radial norm objective does not identify a Gaussian output distribution. The Gaussian-alignment guarantee does not follow; empirical initialization or counting improvements are not refuted.","primaryUrl":"https://arxiv.org/html/2510.19278","noteUrl":"/review?paper=review-neurips-2026-153310","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153311","title":"AVO: Agentic Variation Operators for Autonomous Evolutionary Search","authors":["Terry Chen","Zhifan Ye","Bing Xu","Zihao Ye","Timmy Liu","Ali Hassani","Tianqi Chen","Andrew Kerr","Haicheng Wu","Vedaanta Agarwalla","Fengzhe Zhou","Shangda Li","Yang Xu","Yu-Jung Chen","Hanfeng Chen","Aditya Kane","Ronny Krashinsky","Ming-Yu Liu","Vinod Grover","Luis Ceze","Roger Bringmann","John Tran","Wei Liu","Feng Xie","Michael Lightstone","Humphrey Shi"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"AVO: Agentic Variation Operators for Autonomous Evolutionary Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.24517","noteUrl":"/review?paper=review-neurips-2026-153311","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153314","title":"Hidden Forgetting in Continual Multimodal Learning: When Accuracy Survives but Grounding Fails","authors":["Qianyu Chen","Canran Xiao","Runxuan Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Masking question selectors cannot uniquely identify shortcut-prior reliance","plainExplanation":"Masking question selectors cannot uniquely identify shortcut-prior reliance. The shortcut-versus-grounding identification is unsupported; the sensitivity statistic, training loss and measured benchmark gains remain separate valid objects.","primaryUrl":"https://arxiv.org/html/2607.02020","noteUrl":"/review?paper=review-neurips-2026-153314","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153315","title":"Delta-Adapter: Scalable Exemplar-Based Image Editing with Single-Pair Supervision","authors":["Jiacheng Chen","Songze Li","Han Fu","Baoquan Zhao","Wei Liu","Yanyan Liang","Qing Li","Xudong Mao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Delta-Adapter: Scalable Exemplar-Based Image Editing with Single-Pair Supervision","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07940","noteUrl":"/review?paper=review-neurips-2026-153315","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153317","title":"Geometry-Aware Post-Hoc Uncertainty Quantification in Operator Learning","authors":["Oriol Vendrell-Gallart","Nima Negarandeh","Ramin Bostanabad"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Spectral normalization controls expansion but does not prevent collapse","plainExplanation":"Spectral normalization gives an upper distance bound, not injectivity. Nonzero unit-norm weights in the stated separate ReLU projection can collapse an entire positive ray.","primaryUrl":"https://arxiv.org/html/2606.17513","noteUrl":"/review?paper=review-neurips-2026-153317","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153323","title":"Federate the Router: Learning Language Model Routers with Sparse and Decentralized Evaluations","authors":["Baris Askin","Shivam Patel","Anupam Nayak","Andrea Vigano","Jiin Woo","Gauri Joshi","Carlee Joe-Wong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The routing bounds need a cost-independent accuracy scale and cell coverage","plainExplanation":"Shrinking monetary cost units cannot eliminate the noise in observed correctness.","primaryUrl":"https://arxiv.org/html/2601.22318","noteUrl":"/review?paper=review-neurips-2026-153323","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153325","title":"Adaptive Mass-Segmented KV Compression for Long-Context Reasoning","authors":["Junzhe Yang","Xiaoyu Shen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The global overflow trim can erase a segment floor","plainExplanation":"Use three equal-mass segments of length four and keep budget six, giving quota two per segment. Adding the first sink and the last four positions creates an overflow. The global score trim protects must-keeps but not segment floors, so three removals can erase a low-scoring middle segment even though a six-token set preserving all segments exists. Allocation mass and final scores are explicitly decoupled, so their differing order is allowed. The printed algorithm does not implement its preservation guarantee. A reserve-first code path could repair it; actual code and measured preservation are not judged.","primaryUrl":"https://arxiv.org/html/2605.23200","noteUrl":"/review?paper=review-neurips-2026-153325","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153327","title":"LogicTree-RAG: Logic Tree-guided Retrieval-Augmented Generation for Long-form Patent Drafting","authors":["Jiaqi Zhu","Naili Xing","Pan Hexiang","Haotian Gao","Jianwei Yin","Xiaokui Xiao","Beng Chin Ooi"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"LogicTree-RAG: Logic Tree-guided Retrieval-Augmented Generation for Long-form Patent Drafting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30943","noteUrl":"/review?paper=review-neurips-2026-153327","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153329","title":"Cost Efficient Fairness Audit Under Partial Feedback","authors":["Nirjhar Das","mohit sharma","Praharsh Nanavati","Kirankumar Shiragur","Amit Jayant Deshpande"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The likelihood guarantee needs a stronger separation condition","plainExplanation":"The midpoint Taylor identity has a factor-of-four error. With the printed separation threshold, the resulting likelihood bound fails even for a Gaussian family and an exact parameter estimate. Correcting the identity and strengthening separation recovers the stated likelihood and MAP error bounds under the remaining assumptions.","primaryUrl":"https://arxiv.org/html/2510.03734","noteUrl":"/review?paper=review-neurips-2026-153329","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","incorrect_bound","lemma_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153333","title":"Hierarchical Task Network Planning with LLM-Generated Heuristics","authors":["Felipe Meneguzzi","Alexandre Buchweitz","Augusto B. Corrêa","Victor S Putrich","André G. Pereira"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A positive additive tie penalty can destroy admissibility","plainExplanation":"Adding a positive numerical tie-breaking penalty need not preserve an admissible heuristic: any tight lower bound is immediately pushed above the true remaining cost.","primaryUrl":"https://arxiv.org/html/2605.07707","noteUrl":"/review?paper=review-neurips-2026-153333","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153334","title":"Cluster with Auctions for Vector Search","authors":["Swann BESSA","Pierre Fernandez","Gergely Szilvasy","Matthijs Douze","Herve Jegou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The maximization assignment selects larger cross entropy","plainExplanation":"The maximization assignment selects larger cross entropy. The printed alternating-objective sign is wrong; actual solver inputs are not determined by it.","primaryUrl":"https://arxiv.org/html/2607.13728","noteUrl":"/review?paper=review-neurips-2026-153334","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153336","title":"MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens","authors":["Subhojyoti Mukherjee","Mehrab Tanjim"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens","plainExplanation":"No issue was found in the checked claims. Different scalarizations are assigned to distinct script configurations. Their coefficient difference does not establish two incompatible definitions of the same reward.","primaryUrl":"https://arxiv.org/html/2609.30967","noteUrl":"/review?paper=review-neurips-2026-153336","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153337","title":"Grid Games: The Power Of Multiple Grids for Quantizing Large Language Models","authors":["Vage Egiazarian","Erik Schultheis","Andrei Panferov","Earl Killian","Torsten Hoefler","Dan Alistarh"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Grid Games: The Power Of Multiple Grids for Quantizing Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12327","noteUrl":"/review?paper=review-neurips-2026-153337","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153339","title":"Powering Up Zeroth-Order Training via Subspace Gradient Orthogonalization","authors":["Yicheng Lang","Changsheng Wang","Yihua Zhang","Mingyi Hong","Zheng Zhang","Wotao Yin","Sijia Liu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Powering Up Zeroth-Order Training via Subspace Gradient Orthogonalization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.17155","noteUrl":"/review?paper=review-neurips-2026-153339","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153341","title":"Twisted Schrödinger Bridge Matching","authors":["Maxence Noble","Marie Scheid","Yazid Janati","Eric Moulines","Alain Durmus"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Discrete bridge potentials need bridge-time scaling rather than count averaging","plainExplanation":"The discrete-potential bridge drift uses an active-count average where the reference tilt and bridge sensitivity require a time-scaled sum. A smooth nonnegative quadratic example gives a different exact drift, requiring consistent correction of the discrete regression targets.","primaryUrl":"https://arxiv.org/html/2607.16987","noteUrl":"/review?paper=review-neurips-2026-153341","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153345","title":"Test-Time Scaling in the Wild: Why Exploitation, Not Exploration, Is the Bottleneck","authors":["Davide Romano","Kanak Raj","Jerrod Parker","Daniele Giofrè"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Best-of-N capture equals correlation only under additional joint assumptions","plainExplanation":"Best-of-N capture equals correlation only under additional joint assumptions. The universal analytical identity and attenuation transfer are unsupported; an approximate empirical relation between headroom and verifier quality can remain informative.","primaryUrl":"https://arxiv.org/html/2608.18931","noteUrl":"/review?paper=review-neurips-2026-153345","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153346","title":"Token Inflation: How Dishonest Providers Can Overcharge for Large Language Model Usage","authors":["Shahinul Hoque","Jinghuai Zhang","Jinyuan Stella Sun","Fnu Suya"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Token Inflation: How Dishonest Providers Can Overcharge for Large Language Model Usage","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.30040","noteUrl":"/review?paper=review-neurips-2026-153346","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153349","title":"Do Heavy Tails Help Diffusion? On the Subtle Trade-off Between Initialization and Training","authors":["Hamza Cherkaoui","Hélène Halconruy","Antonio Ocello"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Compact-domain approximation does not supply a whole-space uniform rate","plainExplanation":"Convexity gives \\(\\operatorname{TV}(\\int P_a\\,da,\\int Q_a\\,da)\\leq\\int\\operatorname{TV}(P_a,Q_a)\\,da\\), not the reverse. Distinct conditional components can cancel their mismatch after mixing. Even an exactly matched marginal therefore does not force a small conditional initialization term. The conditional initialization estimate is unsupported. The separate marginal stable-prior convergence result can remain valid.","primaryUrl":"https://arxiv.org/html/2605.13175","noteUrl":"/review?paper=review-neurips-2026-153349","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153350","title":"CoMMa: Contribution-Aware Medical Multi-Agents for Decentralized Oncology Decision Support","authors":["Yichen WU","Yujin Oh","Sangjoon Park","Kailong Fan","Yuhan Liu","Zhiyi Shi","Sekeun Kim","Dania Daye","Hana Farzaneh","Wei Liu","Xiang Li","Raul N Uppot","Quanzheng Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A harmful zero-target agent can drive the gate objective to minus infinity","plainExplanation":"If \\(A_i<0\\), then \\(-A_i\\log W_i\\to-\\infty\\) as \\(W_i\\to0\\). The KL target does not oppose this when \\(\\Phi_i=0\\). A coherent two-agent case predicts \\(0.9\\) and \\(0.1\\) for truth one, with empty-coalition prediction \\(0.5\\). The harmful second agent has target zero and centered leave-one-out advantage \\(-\\log3\\); as its gate tends to zero, BCE stays finite and KL tends to zero while the gate term diverges negatively. This is a conditional objective pathology, not a claim that every instance is unbounded or actual training diverged. Positive softmax weights alone permit arbitrarily small values.","primaryUrl":"https://arxiv.org/html/2602.09159","noteUrl":"/review?paper=review-neurips-2026-153350","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153352","title":"UniPath: Adaptive Coordination of Understanding and Generation for Unified Multimodal Reasoning","authors":["Haoyue Bai","Yinyi Luo","Wenwen Wang","Qingsong Wen","Jindong Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"UniPath: Adaptive Coordination of Understanding and Generation for Unified Multimodal Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11400","noteUrl":"/review?paper=review-neurips-2026-153352","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153354","title":"Who Says What: Symbolic Trimodal Binding Mechanisms in Audio-Visual LLMs","authors":["Jihoo Jung","Youngjoon Jang","Joon Son Chung"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Who Says What: Symbolic Trimodal Binding Mechanisms in Audio-Visual LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31193","noteUrl":"/review?paper=review-neurips-2026-153354","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153355","title":"Gen-Searcher: Reinforcing Agentic Search for Image Generation","authors":["Kaituo Feng","Manyuan Zhang","Shuang Chen","Yunlong Lin","Kaixuan Fan","Yilei Jiang","Hongyu Li","Dian Zheng","Chenyang Wang","Xiangyu Yue"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Gen-Searcher: Reinforcing Agentic Search for Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.28767","noteUrl":"/review?paper=review-neurips-2026-153355","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153357","title":"FrED: External Data Influence Estimation via Domain Knowledge Graph Grounding","authors":["Theodoros Aivalis","Iraklis A Klampanos","Antonis Troumpoukis","Joemon M Jose"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"FrED: External Data Influence Estimation via Domain Knowledge Graph Grounding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.21615","noteUrl":"/review?paper=review-neurips-2026-153357","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153358","title":"$f$-GRPO & Beyond: Divergence-Based Reinforcement Learning Algorithms for General LLM Alignment","authors":["Rajdeep Haldar","Lantao Mei","Guang Lin","Yue XING","Qifan Song"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Policy normalization invalidates the unrestricted divergence argument","plainExplanation":"The proof optimizes over arbitrary critics, but a normalized policy restricts its induced critic. A two-response KL example has minimum loss $e^{-1}$ where the theorem gives zero, and weak monotonicity does not ensure strict reward improvement. Two narrower corrections concern the singular-divergence value and a pseudocode sign already handled correctly in the author implementation.","primaryUrl":"https://arxiv.org/html/2602.05946","noteUrl":"/review?paper=review-neurips-2026-153358","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","invalid_reduction","missing_assumption","notation"],"classifications":["critical"]},{"slug":"review-neurips-2026-153361","title":"Variational Trajectory Optimization of Anisotropic Diffusion Schedules","authors":["Pengxi Liu","Zeyu Michael Li","Xiang Cheng"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The trajectory and solver formulas need inverse powers, matrix derivatives and a reverse-step sign correction","plainExplanation":"The final factor in Equation (11) is $M^{1/2}$ instead of $M^{-1/2}$. This changes the time- and schedule-dependent weighting, not merely an overall constant.","primaryUrl":"https://arxiv.org/html/2602.19512","noteUrl":"/review?paper=review-neurips-2026-153361","collection":"neurips-2026","severity":"repairable","issueTypes":["mathematical_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153363","title":"Dimension-Uniform Discretization Analysis of Preconditioned Annealed Langevin Dynamics for Multimodal Gaussian Mixtures","authors":["Lorenzo Baldassari","Josselin Garnier","Knut Solna","Maarten V. de Hoop"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Dimension-Uniform Discretization Analysis of Preconditioned Annealed Langevin Dynamics for Multimodal Gaussian Mixtures","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16473","noteUrl":"/review?paper=review-neurips-2026-153363","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153364","title":"Anchoring Adversarial Trajectories to Data Manifolds: A Bilevel Transfer Optimization Framework","authors":["Yaohua Liu","Yifan Guo","Jiaxin Gao"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Anchoring Adversarial Trajectories to Data Manifolds: A Bilevel Transfer Optimization Framework","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38991","noteUrl":"/review?paper=review-neurips-2026-153364","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153366","title":"Values as Style: Disentangling Values from Semantics with One-Way Mixing for Low-Damage LLM Steering","authors":["Jiale Dai","Hongcan Deng","Liuxian Ma","Xiaoke Niu","Guojie Song"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Values as Style: Disentangling Values from Semantics with One-Way Mixing for Low-Damage LLM Steering","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.39701","noteUrl":"/review?paper=review-neurips-2026-153366","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153369","title":"Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models","authors":["Phongsakon Mark Konrad","Toygar Tanyel","Serkan Ayvaz"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.20436","noteUrl":"/review?paper=review-neurips-2026-153369","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153372","title":"AnyBokeh: Physics-Guided Any-to-Any Bokeh Editing with Optical Fingerprint Transfer","authors":["Xinyu Hou","Xiaoming Li","Zongsheng Yue","Chen Change Loy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AnyBokeh: Physics-Guided Any-to-Any Bokeh Editing with Optical Fingerprint Transfer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.31959","noteUrl":"/review?paper=review-neurips-2026-153372","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153375","title":"D-GAP: Improving Out-of-Domain Robustness via Dataset-Agnostic and Gradient-Guided Augmentation in Amplitude and Pixel Spaces","authors":["Ruoqi Wang","Haitao Wang","Shaojie Guo","Qiong Luo"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The reconstruction uses the opposite Fourier phase sign","plainExplanation":"Under that convention, \\(F=Ae^{i\\arg F}\\). Using \\(Ae^{-i\\arg F}\\) conjugates the spectrum, whose inverse for a real image is its modular spatial reversal. A circularly shifted target has the same amplitude, so an interior amplitude-mixing ratio does not remove the error. Replacing arctangent by a quadrant-aware phase also does not correct the sign. This is a printed reconstruction error. The actual implementation and empirical degradation are not established.","primaryUrl":"https://arxiv.org/html/2511.11286","noteUrl":"/review?paper=review-neurips-2026-153375","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153378","title":"FlexTab: A Flexible Encoder-Decoder Architecture for In-Context Learning Across Diverse Tabular Tasks","authors":["Marek Polewczyk","Maximilian Schambach","Marco Spinaci","Sam Thelin","Johannes Höhne"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"FlexTab: A Flexible Encoder-Decoder Architecture for In-Context Learning Across Diverse Tabular Tasks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30336","noteUrl":"/review?paper=review-neurips-2026-153378","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153382","title":"Breaking Modality Heterogeneity in Low-Bit Quantization for Large Vision-Language Models","authors":["YI ZHONG","Haotong Qin","Xindong Zhang","Lei Zhang","Guolei Sun"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Breaking Modality Heterogeneity in Low-Bit Quantization for Large Vision-Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19929","noteUrl":"/review?paper=review-neurips-2026-153382","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153384","title":"Handwritten Text Recognition Lives in the High-Pixel Variance Subspace","authors":["Carlos Garrido","Jorge Calvo-Zaragoza"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Handwritten Text Recognition Lives in the High-Pixel Variance Subspace","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35473","noteUrl":"/review?paper=review-neurips-2026-153384","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153387","title":"An Efficient Algorithm for Thresholding Monte Carlo Tree Search","authors":["Shoma Nameki","Atsuyoshi Nakamura","Junpei Komiyama","Koji Tabata"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The printed concentration prefactor is negative","plainExplanation":"The efficient tree routines mishandle zero information at threshold ties. The efficient routines are not equivalent at allowed discrete ties; a continuous almost-sure-tie argument does not cover the stated reward family.","primaryUrl":"https://arxiv.org/html/2601.22600","noteUrl":"/review?paper=review-neurips-2026-153387","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153388","title":"EchoPrune: Interpreting Redundancy as Temporal Echoes for Efficient VideoLLMs","authors":["Jiameng Li","Minye Wu","Jiezhang Cao","Aleksei Tiulpin","Matthew Blaschko"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Increasing frame count alone does not establish the Nyquist condition","plainExplanation":"Increasing frame count alone does not establish the Nyquist condition. The general signal-processing guarantee is unsupported; the actual 192-frame setting and observed QA gains are not refuted by this example.","primaryUrl":"https://arxiv.org/html/2605.10050","noteUrl":"/review?paper=review-neurips-2026-153388","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153393","title":"LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling","authors":["Yunjiang Jiang","Ayush Agarwal","Yang Liu","Yihan Wu","Haoran Liu","Bi Xue"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling","plainExplanation":"The printed attention mask and backward derivative formulas retain their local mathematical inconsistencies. The metric-definition row is outside the mathematical review.","primaryUrl":"https://arxiv.org/html/2510.18239","noteUrl":"/review?paper=review-neurips-2026-153393","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153396","title":"SceneBind: Binding What and Where Across Vision, Audio, and Language","authors":["Mingfei Chen","Zijun Cui","Ruoke Zhang","Hyeonggon Ryu","Eli Shlizerman"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SceneBind: Binding What and Where Across Vision, Audio, and Language","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.15265","noteUrl":"/review?paper=review-neurips-2026-153396","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153399","title":"NRF-GS: Neural Residual Fields for Expressive and Compact Gaussian Splatting","authors":["Pratik S. Bisht","Andreas Kolb"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"NRF-GS: Neural Residual Fields for Expressive and Compact Gaussian Splatting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.37115","noteUrl":"/review?paper=review-neurips-2026-153399","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153401","title":"Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness","authors":["Songyuan Sui","Zhen Tan","Mohan Zhang","Rana M Khan","Xia Hu","Tianlong Chen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04792","noteUrl":"/review?paper=review-neurips-2026-153401","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153402","title":"Beyond the Half Approximation: Fair and Efficient Online Class Matching","authors":["Sander Borst","Max Springer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond the Half Approximation: Fair and Efficient Online Class Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23408","noteUrl":"/review?paper=review-neurips-2026-153402","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153403","title":"SLIDERS: Systematic Reviews via Automated Evidence Synthesis and Reconciliation","authors":["Harshit Joshi","Priyank Shethia","Jadelynn Dao","Monica Lam"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SLIDERS: Systematic Reviews via Automated Evidence Synthesis and Reconciliation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.22294","noteUrl":"/review?paper=review-neurips-2026-153403","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153406","title":"S2D: Sparse-To-Dense Keymask Distillation for Unsupervised Video Instance Segmentation","authors":["Leon Sick","Lukas Hoyer","Dominik Engel","Pedro Hermosilla","Timo Ropinski"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"S2D: Sparse-To-Dense Keymask Distillation for Unsupervised Video Instance Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.14440","noteUrl":"/review?paper=review-neurips-2026-153406","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153407","title":"SMI: Statistical Membership Inference for Reliable Unlearned Model Auditing","authors":["Jialong Sun","Zeming Wei","Jiaxuan Zou","Jiacheng Gong","Chengyang Dong","Jie Fu","Heng Xu","Jialong Li","Bo Liu"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"SMI: Statistical Membership Inference for Reliable Unlearned Model Auditing","plainExplanation":"The proof expands a kernel mean discrepancy, not the original covariance residual. The two criteria can have different minimizer sets even when the audit distribution is exactly a mixture of the references.","primaryUrl":"https://arxiv.org/html/2602.01150","noteUrl":"/review?paper=review-neurips-2026-153407","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_reduction","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153410","title":"Why Latent Actions Fail, and How to Prevent It","authors":["Jung Min Lee","Taehyun Cho","Li Zhao","Jungwoo Lee"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Why Latent Actions Fail, and How to Prevent It","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20223","noteUrl":"/review?paper=review-neurips-2026-153410","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153411","title":"VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion","authors":["Hidir Yesiltepe","Jiazhen Hu","Tuna Han Salih Meral","Adil K Akan","Kaan Oktay","Hoda Eldardiry","Pinar Yanardag"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30351","noteUrl":"/review?paper=review-neurips-2026-153411","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153412","title":"LAMP: Look-Ahead Mixed-Precision Inference of Large Language Models","authors":["Stanislav Budzinskiy","Marián Gloser","Tolunay Yilmaz","Ying H Tham","Yuanyi Lin","Wenyi Fang","FAN WU","Philipp Petersen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The softmax condition number omits selected rows from its maximum","plainExplanation":"Let \\(U=\\sum_{j\\notin\\Omega}z_j|y_j|\\). Selected rows have absolute row sum \\(U\\), while unselected row \\(i\\) has \\(U+(1-2z_i)|y_i|\\). If selected rows exist, the correction must therefore be \\(\\max\\{0,\\max_{i\\notin\\Omega}(1-2z_i)|y_i|\\}\\). Omitting zero loses the selected-row maximum when an unselected probability exceeds one half. The general row-maximum formula is incomplete. Particular examples with nonnegative corrections and the separate softmax proposition are not refuted.","primaryUrl":"https://arxiv.org/html/2601.21623","noteUrl":"/review?paper=review-neurips-2026-153412","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153413","title":"Predicting Quantization Price for Selecting PTQ Configurations Before Deployment","authors":["Junbin Qiu","Jian Mu","Weitong Zhang","Yao SHU"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Predicting Quantization Price for Selecting PTQ Configurations Before Deployment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.28270","noteUrl":"/review?paper=review-neurips-2026-153413","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153416","title":"Tatemae: Detecting Alignment Faking via Tool Selection in LLMs","authors":["Matteo Leonesi","Francesco Belardinelli","Flavio Corradini","Marco Piangerelli"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Tatemae: Detecting Alignment Faking via Tool Selection in LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.26511","noteUrl":"/review?paper=review-neurips-2026-153416","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153418","title":"Video Models Can Reason with Verifiable Rewards","authors":["Tinghui Zhu","Sheng Zhang","James Yipeng Huang","Selena Song","Xiaofei Wen","Yuankai Li","Hoifung Poon","Muhao Chen"],"year":2026,"category":"Reinforcement learning","status":"addressed","issueTitle":"Video Models Can Reason with Verifiable Rewards","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15458","noteUrl":"/review?paper=review-neurips-2026-153418","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153419","title":"Focusable Monocular Depth Estimation","authors":["Yuxin Du","Tao Lin","Zile Zhong","Runting Li","Xiyao Chen","Jiting Liu","Chenglin Liu","Yingcong Chen","Yuqian Fu","Bo Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Per-target aggregation does not balance images by target count","plainExplanation":"The pooled empirical distribution gives image \\(i\\) total mass \\(n_i/N\\). Computing its median instead of its mean does not give every image equal weight.","primaryUrl":"https://arxiv.org/html/2605.11756","noteUrl":"/review?paper=review-neurips-2026-153419","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-153420","title":"SurGe: Improved Surface Geometry in Point Maps","authors":["Karim Knaebel","Gonzalo M Garcia","Christian Schmidt","Ilya Fradlin","Lucas Nunes","Daan de Geus","Bastian Leibe"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SurGe: Improved Surface Geometry in Point Maps","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31577","noteUrl":"/review?paper=review-neurips-2026-153420","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153421","title":"Events as Triggers for Behavioral Diversity in Multi-Agent Reinforcement Learning","authors":["Hannes M Büchi","Manon Flageat","Eduardo Sebastián","Amanda Prorok"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Events as Triggers for Behavioral Diversity in Multi-Agent Reinforcement Learning","plainExplanation":"Scaling every deviation by $c>0$ multiplies each pairwise distance by $c$. Scaling only $u_m$ does not: $\\|u_m-u_n\\|$ is a distance from the fixed point $u_n$, and pairs excluding $m$ do not change at all. The joint Euler identity therefore cannot justify Equation (6)'s separate identity for each behavior, or the claimed idempotency of each $P_{u_m}$.","primaryUrl":"https://arxiv.org/html/2605.12388","noteUrl":"/review?paper=review-neurips-2026-153421","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153423","title":"Proactive Instance Navigation with Comparative Judgment for Ambiguous User Queries","authors":["Junhyuk Kwon","Seungjoon Lee","Hyejin Park","Kyle Min","Jungseul Ok"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Proactive Instance Navigation with Comparative Judgment for Ambiguous User Queries","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06223","noteUrl":"/review?paper=review-neurips-2026-153423","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153425","title":"Polynomial-Time Robust Multiclass Linear Classification under Gaussian Marginals","authors":["Ilias Diakonikolas","Giannis Iakovidis","Mingchen Ma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Polynomial-Time Robust Multiclass Linear Classification under Gaussian Marginals","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21428","noteUrl":"/review?paper=review-neurips-2026-153425","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153428","title":"Diagnosing and Repairing Citation Failures in Generative Engine Optimization","authors":["Zhihua Tian","Yuhan Chen","Yao Tang","Jian Liu","Ruoxi Jia"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Diagnosing and Repairing Citation Failures in Generative Engine Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.09296","noteUrl":"/review?paper=review-neurips-2026-153428","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153430","title":"Which Way Did It Move? Diagnosing and Overcoming Directional Motion Blindness in Video LLMs","authors":["Jongseo Lee","Hyuntak Lee","Sunghun Kim","Sooa Kim","Jihoon Chung","Jinwoo Choi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Which Way Did It Move? Diagnosing and Overcoming Directional Motion Blindness in Video LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22823","noteUrl":"/review?paper=review-neurips-2026-153430","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153431","title":"Stochastic Reconfiguration as Statistical Filtering for Overparameterized Neural Quantum States","authors":["Tak Hur"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Stochastic Reconfiguration as Statistical Filtering for Overparameterized Neural Quantum States","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.23334","noteUrl":"/review?paper=review-neurips-2026-153431","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153433","title":"XBRIDGE: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication","authors":["Wooseong Yang","Wei-Chieh Huang","Weizhi Zhang","Yu Wang","Philip S Yu","Junhyun Lee"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"XBRIDGE: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.11676","noteUrl":"/review?paper=review-neurips-2026-153433","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153435","title":"PAC Learning with Bandit Feedback: Sharp Sample Complexity in the Realizable Setting","authors":["Steve Hanneke","Qinglin Meng","Shay Moran","Amirreza Shaeiri"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"PAC Learning with Bandit Feedback: Sharp Sample Complexity in the Realizable Setting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25678","noteUrl":"/review?paper=review-neurips-2026-153435","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153436","title":"Concurrence of Symmetry Breaking and Nonlocality Phase Transitions in Diffusion Models","authors":["Yifan F Zhang","Fangjun Hu","Guangkuo Liu","Mert Okyay","Xun Gao"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Equal sample counts do not make FID bias harmless for selecting minima","plainExplanation":"Equal sample counts do not make FID bias harmless for selecting minima. The blanket minimizer rationale is false. Bootstrap spread and acknowledged imprecision do not establish a common bias; no actual plotted timing reversal is inferred.","primaryUrl":"https://arxiv.org/html/2605.04830","noteUrl":"/review?paper=review-neurips-2026-153436","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153439","title":"MotionGrounder: Grounded Multi-Object Motion Transfer via Diffusion Transformer","authors":["Samuel Teodoro","Yun Chen","Agus Gunawan","Soo Ye Kim","Jihyong Oh","Munchurl Kim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MotionGrounder: Grounded Multi-Object Motion Transfer via Diffusion Transformer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.00853","noteUrl":"/review?paper=review-neurips-2026-153439","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153441","title":"CRISP: Fixing Flying Pixels in Latent LiDAR Generation via Diffusion Decoding","authors":["Andrea Ceron","Michael Schmidt","Alvaro Marcos-Ramiro","Sebastian Schmidt","Benjamin Busam"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"CRISP: Fixing Flying Pixels in Latent LiDAR Generation via Diffusion Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.11376","noteUrl":"/review?paper=review-neurips-2026-153441","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153443","title":"Don't Pay Attention, PLANT It: Pretraining Attention via Learning-to-Rank","authors":["Debjyoti Saha Roy","Byron Wallace","Javed A Aslam"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Bootstrap metric replicates are not new independent test units","plainExplanation":"All replicates condition on one original test set. Increasing their count reduces Monte Carlo uncertainty about that test set’s noisy statistic, without adding independent patient or document information. A Wilcoxon test across those replicates can therefore produce increasingly small p-values around an accidental observed advantage rather than the intended population comparison. The separate percentile-CI-excludes-zero requirement may protect final significance arrows. The Wilcoxon justification alone fails; not every displayed conclusion or its actual decision branch is independently refuted.","primaryUrl":"https://arxiv.org/html/2410.23066","noteUrl":"/review?paper=review-neurips-2026-153443","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153444","title":"Rethinking Layer-wise Model Merging through Chain of Merges","authors":["Pietro Buzzega","Riccardo Salami","Angelo Porrello","SIMONE CALDERARA"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Scalar activation normalization does not improve the Gram condition number","plainExplanation":"Uniform scaling changes magnitude but not singular-value ratios or rank. It cannot by itself supply the stated conditioning improvement.","primaryUrl":"https://arxiv.org/html/2508.21421","noteUrl":"/review?paper=review-neurips-2026-153444","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153447","title":"Dominant-Layer ZO: A Single Layer Dominates Zeroth-Order Fine-Tuning of LLMs","authors":["Wanhao Yu","Ziyan Wang","Zheng Wang","Abeer M Almalky","Yihang ZUO","Shuteng Niu","Sen Lin","Adnan Siraj Rakin","Deliang Fan","Li Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Dominant-Layer ZO: A Single Layer Dominates Zeroth-Order Fine-Tuning of LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05516","noteUrl":"/review?paper=review-neurips-2026-153447","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153450","title":"Lipschitz Dueling Bandits over Continuous Action Spaces","authors":["Mudit Sharma","Shweta Jain","Vaneet Aggarwal","Ganesh Ghalme"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Counting sampled arms does not bound exact coordinate bit length","plainExplanation":"A fixed finite-bit code has finite support, while an exact uniform cube sample has a nonatomic law. An index among at most \\(T\\) samples is a pointer only if the sample coordinates or random tape are stored elsewhere. The number of samples does not prove the precision needed to reproduce their values exactly. The coordinate-bit proof is incomplete. Object-count memory and the achievable regret order are not refuted.","primaryUrl":"https://arxiv.org/html/2604.00523","noteUrl":"/review?paper=review-neurips-2026-153450","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153452","title":"Dooly: Configuration-Agnostic, Redundancy-Aware Profiling for LLM Inference Simulation","authors":["Joon Ha Kim","Geon-Woo Kim","Anoop Rachakonda","Daehyeok Kim"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Dooly: Configuration-Agnostic, Redundancy-Aware Profiling for LLM Inference Simulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07985","noteUrl":"/review?paper=review-neurips-2026-153452","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153453","title":"HyperNSDE: Personalized Neural SDEs for Joint Static—Longitudinal Clinical Data Generation","authors":["Perrine Chassat","Agathe Guilloux"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"HyperNSDE: Personalized Neural SDEs for Joint Static—Longitudinal Clinical Data Generation","plainExplanation":"This minimizes the ELBO, favoring worse reconstruction or larger KL penalties rather than maximizing the stated variational lower bound.","primaryUrl":"https://arxiv.org/html/2610.07383","noteUrl":"/review?paper=review-neurips-2026-153453","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153454","title":"Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks","authors":["Jia Liang","Liangming Pan"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.00449","noteUrl":"/review?paper=review-neurips-2026-153454","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153455","title":"Knowledge Localization in Mixture-of-Experts LLMs Using Cross-Lingual Inconsistency","authors":["Lucas Bandarkar","Alan Ansell","Trevor Cohn"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Unselected logits cancel from renormalized top-k weights","plainExplanation":"Top-k renormalization cancels every unselected logit, contrary to the stated flattening rationale. The minimum-gap threshold also has the opposite count monotonicity to the prose, and the model-specific threshold assignments disagree.","primaryUrl":"https://arxiv.org/html/2603.17102","noteUrl":"/review?paper=review-neurips-2026-153455","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153456","title":"Consolidating Rewarded Perturbations for LLM Post-Training","authors":["Zheyu Zhang","Shuo Yang","Gjergji Kasneci"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Consolidating Rewarded Perturbations for LLM Post-Training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.31494","noteUrl":"/review?paper=review-neurips-2026-153456","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153457","title":"Secure Seed-Based Multi-bit Watermarking for Diffusion Models from First Principles","authors":["Enoal Gesny","Eva Giboulot"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The actual fixed-message law has infinite reverse KL from the Gaussian cover","plainExplanation":"The key definition lacks the orthonormality required for decoding, the flip formula can be negative, and the fixed-message law has a support obstruction to the claimed reverse KL. The formal robustness condition also permits erasure rather than a BSC, and uncentered PCA retains a mean spike omitted from its covariance spectrum.","primaryUrl":"https://arxiv.org/html/2605.06153","noteUrl":"/review?paper=review-neurips-2026-153457","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153458","title":"Closed-Form Linear-Probe Dataset Distillation for Pre-trained Vision Models","authors":["Bincheng Peng","Miki Haseyama","Guang Li","Ping Liu","Takahiro Ogawa"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Closed-Form Linear-Probe Dataset Distillation for Pre-trained Vision Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07194","noteUrl":"/review?paper=review-neurips-2026-153458","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153460","title":"Where, What, Why, and Importance: Structured Defect Grounding for Text-to-Image Feedback","authors":["Huaisong Zhang","Hao Yu","Yuxuan Zhang","Jiahe Wang","Xinrui Chen","Haoxiang Cao","Feng Lu","Wendong Zhang","Changqian Yu","Chun Yuan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Where, What, Why, and Importance: Structured Defect Grounding for Text-to-Image Feedback","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06113","noteUrl":"/review?paper=review-neurips-2026-153460","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153462","title":"ReLoop: Structured Modeling and Behavioral Verification for Reliable LLM-Based Optimization","authors":["Junbo Jacob Lian","Yujun Sun","Huiling Chen","Chaoyu Zhang","Hanzhang Qin","Chung-Piaw Teo"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"ReLoop: Structured Modeling and Behavioral Verification for Reliable LLM-Based Optimization","plainExplanation":"The behavioral verification method depends on its stated assumption and has the disclosed detection limits. No mathematical error was identified in the material examined. Equivalence with the accepted conference manuscript has not been established.","primaryUrl":"https://arxiv.org/html/2602.15983","noteUrl":"/review?paper=review-neurips-2026-153462","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153463","title":"AgentVista: Evaluating Multimodal Agent in Ultra-Challenging Realistic Visual Scenarios","authors":["Zhaochen Su","Jincheng Gao","Hangyu Guo","Zhenhua Liu","Lueyang Zhang","Xinyu Geng","Shijue Huang","Peng Xia","Guanyu Jiang","Cheng Wang","Yue Zhang","Yi R. (May) Fung","Junxian He"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AgentVista: Evaluating Multimodal Agent in Ultra-Challenging Realistic Visual Scenarios","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.23166","noteUrl":"/review?paper=review-neurips-2026-153463","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153466","title":"Federated Concept-Based Models: Interpretable models with distributed supervision","authors":["Dario Fenoglio","Arianna Casanova Flores","Francesco De Santis","Gabriele Dominici","Johannes Schneider","Pietro Barbiero","Giovanni De Felice","Marc Langheinrich","Martin Gjoreski"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Federated Concept-Based Models: Interpretable models with distributed supervision","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04093","noteUrl":"/review?paper=review-neurips-2026-153466","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153467","title":"Unveiling Implicit Advantage Symmetry: Why GRPO Struggles with Exploration and Difficulty Adaptation","authors":["zhiqi yu","Zhangquan Chen","Mengting Liu","Heye Zhang","Liangqiong Qu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"An unsampled logit increase need not increase its probability","plainExplanation":"The full-objective KL correction drops old/current-policy factors and is valid only on-policy. The exploration argument also confuses logit movement with probability movement, omitting the softmax normalization term.","primaryUrl":"https://arxiv.org/html/2602.05548","noteUrl":"/review?paper=review-neurips-2026-153467","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153468","title":"Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification","authors":["Xabier de Juan","Santiago Mazuelas","Yilun Zhu","Clay Scott"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39829","noteUrl":"/review?paper=review-neurips-2026-153468","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153469","title":"STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation","authors":["Ying Shen","Tianrong Chen","Yuan Gao","Yizhe Zhang","Yuyang Wang","Miguel Angel Bautista","Shuangfei Zhai","Joshua Susskind","Jiatao Gu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08029","noteUrl":"/review?paper=review-neurips-2026-153469","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153477","title":"Inertia-1: An Open Exploration of Wearable Motion Foundation Models","authors":["Zongzhe Xu","Aakarsh Anand","Sarah Jiang","Chuntung Zhuang","Zitao Shuai","Sriram Sankararaman","Yuzhe Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Inertia-1: An Open Exploration of Wearable Motion Foundation Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.06617","noteUrl":"/review?paper=review-neurips-2026-153477","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153479","title":"FlashMol: High-Quality Molecule Generation in as Few as Four Steps","authors":["Xinyuan Wei","Zian Li","Shaoheng Yan","Cai Zhou","Muhan Zhang"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The noise-level inverse reverses the indexed grid","plainExplanation":"The noise-level inverse reverses the indexed grid. This is a local inversion/indexing mismatch; it does not establish wrong implemented traversal.","primaryUrl":"https://arxiv.org/html/2605.07020","noteUrl":"/review?paper=review-neurips-2026-153479","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153481","title":"Autoregressive Diffusion World Models for Off-Policy Evaluation of LLM Agents","authors":["Kaixuan Liu","GUOJUN XIONG","Weinan Zhang","Shengpu Tang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The trajectory tilt omits behavior-policy likelihood ratios","plainExplanation":"The exact evaluation-law tilt omits behavior likelihood ratios, and the recorded cosine statistic differs from the stated value estimator. Other theoretical mismatches concern deterministic-policy Q targets, score/noise denoising conventions, and the claim that independent denoising prevents rollout error accumulation.","primaryUrl":"https://arxiv.org/html/2606.05558","noteUrl":"/review?paper=review-neurips-2026-153481","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153485","title":"Scratchpad Patching: Decoupling Compute from Patch Size in Byte-Level Language Models","authors":["Lin Zheng","Vasilisa Bashlovkina","Timothy Dozat","Dan Garrette","Laura Rimell","Joshua Maynez"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Scratchpad Patching: Decoupling Compute from Patch Size in Byte-Level Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09630","noteUrl":"/review?paper=review-neurips-2026-153485","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153486","title":"Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics","authors":["Liang Shi","Jiarui Lu","Junqi Liu","Chence Shi","Zhi Yang","Jian Tang"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The single-track and pair-track channel dimensions cannot be directly added","plainExplanation":"The single-track and pair-track channel dimensions cannot be directly added. The printed operation is dimensionally undefined; an implicit projection in code would be a documentation repair.","primaryUrl":"https://arxiv.org/html/2603.17633","noteUrl":"/review?paper=review-neurips-2026-153486","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153487","title":"EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration","authors":["WUYANG LI","Yang Gao","Mariam Hassan","Lan Feng","Wentao Pan","Po-Chien Luan","Alexandre Alahi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15042","noteUrl":"/review?paper=review-neurips-2026-153487","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153488","title":"Imperfect Influence, Reliable Rankings: A Theory of TRAK for Data Attribution","authors":["Han Tong","Shubhangi Ghosh","Haolin Zou","Arian Maleki"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The full gradient Gram is singular under the allowed linear activation","plainExplanation":"The claimed neural example violates gradient-Gram nondegeneracy through a scaling symmetry. The inverse-perturbation lemma also needs a uniform norm margin, and an absolute derivative bound cannot imply a positive signed self-influence.","primaryUrl":"https://arxiv.org/html/2602.01312","noteUrl":"/review?paper=review-neurips-2026-153488","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153491","title":"INSPO : Unlocking Intrinsic Self-Reflection for LLM Preference Optimization","authors":["Yu Li","Tian Lan","Zhengling Qi"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Nondecreasing scalarizations can add optimal actions through ties","plainExplanation":"The equation implies \\(e^{r/\\beta}/Z=\\operatorname{odds}(P)^{1/(2\\beta)}\\). Since \\(Z\\) is the reward’s own partition function, the reference-weighted sum over each comparator column must equal one. Reciprocal two-action preferences \\(0.8/0.2\\), self ties and a uniform reference instead give a column sum \\(1.5\\) when \\(\\beta=1\\). The reward cannot vary \\(\\log Z\\) independently to remove this normalization constraint. The arbitrary-preference claim is too broad. The conditional Gibbs algebra within the compatible class remains valid.","primaryUrl":"https://arxiv.org/html/2512.23126","noteUrl":"/review?paper=review-neurips-2026-153491","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153493","title":"Group-Aware Matrix Estimation and Latent Subspace Recovery","authors":["Hamza Golubovic","Matthew Shen","Genevera Allen","Tarek M Zikry"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The accelerated iteration count scales with distance rather than squared distance","plainExplanation":"The accelerated stopping count has incorrect distance dependence. The observation model also combines incompatible multiplicity/mask/noise conventions, the global cone does not imply the claimed blockwise RSC thresholds, and the main-text regime for dropping the radius correction reverses its signal-to-noise dependence.","primaryUrl":"https://arxiv.org/html/2605.20559","noteUrl":"/review?paper=review-neurips-2026-153493","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153494","title":"DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation","authors":["Jaehun Jung","Hyunwoo Kim","Brandon Cui","Ximing Lu","David Acuna","Prithviraj Ammanabrolu","Yejin Choi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Zero empirical KL does not identify equal population policies","plainExplanation":"Distinct Bernoulli laws with parameters \\(0.99\\) and \\(0.90\\) have positive population KL, but independently sampled sixteen-rollout histograms can coincide with probability about \\(0.206\\). Applying the same Laplace smoothing preserves their equality and produces zero empirical KL. A finite observed equality cannot identify the underlying laws exactly. The population-equality inference fails. This does not establish that every zero-delta prompt differs or that a measured empirical histogram is wrong.","primaryUrl":"https://arxiv.org/html/2605.15532","noteUrl":"/review?paper=review-neurips-2026-153494","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153495","title":"Co-Evolving Interpolants and Flows via Path-Flow Alignment","authors":["Zeyu Michael Li","William X Chen","Xiang Cheng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Co-Evolving Interpolants and Flows via Path-Flow Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.08717v1","noteUrl":"/review?paper=review-neurips-2026-153495","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153498","title":"Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO","authors":["Jiahua Yang","Chuangchuang Wang","Zhiwei Yang","Xianpeng Zhang","Dongyu Chen","Xing Chen","Tianhuang Su","Haonan Lu","Quanlong Guan","Kai Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO","plainExplanation":"A homogeneous reward group still makes the printed standardized advantage undefined; the mathematical specification needs a zero-variance convention.","primaryUrl":"https://arxiv.org/html/2609.36932","noteUrl":"/review?paper=review-neurips-2026-153498","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153502","title":"WorldAct: Activating Monolithic 3D Worlds into Interactive-Ready Object-Centric Scenes","authors":["Jichen Hu","Jiawei Guo","Jiazhong Cen","Chen Yang","Sikuang Li","Wei Shen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"WorldAct: Activating Monolithic 3D Worlds into Interactive-Ready Object-Centric Scenes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15843","noteUrl":"/review?paper=review-neurips-2026-153502","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153503","title":"MemDLM: Memory-Enhanced DLM Training","authors":["Zehua Pei","Hui-Ling Zhen","Weizhe Lin","Sinno Pan","Yunhe Wang","Mingxuan Yuan","Bei Yu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MemDLM: Memory-Enhanced DLM Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.22241","noteUrl":"/review?paper=review-neurips-2026-153503","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153505","title":"PANDA: Prior-guided Attentional Dual-path Architecture","authors":["Rishabh Jain","Federico Belotti","Stefano Coniglio","Pietro Lió","Stefano Fiorini"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"PANDA: Prior-guided Attentional Dual-path Architecture","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.13635","noteUrl":"/review?paper=review-neurips-2026-153505","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153510","title":"Rare Events, Real Signals: Functional Ensembles as Units of Computation in Deep Spiking Networks","authors":["Aditi Aravind","Konstantinos Ladakis","Mario Alexios Savaglio","Stelios Smirnakis","Maria Papadopouli"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Rare Events, Real Signals: Functional Ensembles as Units of Computation in Deep Spiking Networks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.00073","noteUrl":"/review?paper=review-neurips-2026-153510","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153513","title":"SageSched: Efficient LLM Scheduling Confronting Demand Uncertainty and Hybridity","authors":["Zhenghao Gan","Yichen Bao","Yifei Liu","Chen Chen","Quan Chen","Minyi Guo"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"A resource-cost Gittins index is not optimal for the batched backend","plainExplanation":"Use a constant-duration decode backend with capacity \\(202\\) KV tokens. One job has prompt/output lengths \\((2,10)\\), and four have \\((100,1)\\), all known exactly. Their surrogate indices are \\(70\\) and \\(100.5\\), so cost-priority batching keeps the long-output job while processing only one short job per step. Completion times are \\(1,2,3,4,10\\), mean \\(4\\). Two short jobs fit together at each of the first two steps; running the long job afterward gives \\(1,1,2,2,12\\), mean \\(3.6\\). Both schedules satisfy the stated capacity regime. The universal optimality transfer fails even with exact length predictions. The classical single-server theorem, communication predictor and measured speedups remain valid within their scopes.","primaryUrl":"https://arxiv.org/html/2603.07917","noteUrl":"/review?paper=review-neurips-2026-153513","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153514","title":"Pathways of Visual Information Flow in Vision-Language Models","authors":["Israfel Salazar","Stella Christina Frank","Dan Oneata","Desmond Elliott","Constanza Fierro"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The knockout does not uniquely identify query-text mediation","plainExplanation":"Preserved performance after removing direct image-to-final edges establishes that those edges are unnecessary for that intervention, but does not by itself identify the specific surviving mediator.","primaryUrl":"https://arxiv.org/html/2607.03358","noteUrl":"/review?paper=review-neurips-2026-153514","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153517","title":"DivMoE: Fine-Grained MoE Upcycling via Cross-Domain Expert Composition","authors":["Yuxuan Lou","Kai Yang","Geng Zhang","Yong Liu","Yang You"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DivMoE: Fine-Grained MoE Upcycling via Cross-Domain Expert Composition","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.11317","noteUrl":"/review?paper=review-neurips-2026-153517","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153519","title":"Managing Self-Learning Experts under Per-Round Budget Constraints","authors":["Ilgam Latypov","Alexandra Suvorikova","Alexey Kroshnin","Alexander Gasnikov","Iurii Dorn"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Managing Self-Learning Experts under Per-Round Budget Constraints","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.22654","noteUrl":"/review?paper=review-neurips-2026-153519","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153522","title":"Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings","authors":["Marie Generali Lince","Vincent Divol","Rémi Flamary","Solenne Gaucher","Patrick Loiseau"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The aware transport bound does not bound the unaware parity cost","plainExplanation":"The Jordan-component normalization uses a total-variation convention inconsistent with its declared OT cost. Separately, a group-aware Wasserstein transport bound does not control the much larger exact-parity cost of a single unaware predictor.","primaryUrl":"https://arxiv.org/html/2605.28233","noteUrl":"/review?paper=review-neurips-2026-153522","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153523","title":"Fast Gauss-Newton for Multiclass Cross-Entropy","authors":["Mikalai Korbit","Mario Zanon"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Fast Gauss-Newton for Multiclass Cross-Entropy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06081","noteUrl":"/review?paper=review-neurips-2026-153523","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153525","title":"Minimally Invasive Steering of Language Models","authors":["Taha Entesari","Jingyu (Jack) Zhang","Daniel Khashabi","Mahyar Fazlyab"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Minimally Invasive Steering of Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30218","noteUrl":"/review?paper=review-neurips-2026-153525","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153527","title":"PGID: Progressive Guided Inversion and Denoising for Robust Watermark Detection","authors":["Minh Quoc Duong","Chun Tong Lei","Chun Pong Lau"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PGID: Progressive Guided Inversion and Denoising for Robust Watermark Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09319","noteUrl":"/review?paper=review-neurips-2026-153527","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153529","title":"Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs","authors":["Hoigi Seo","Byung Hyun Lee","Minjun Kim","Dohyun Mah","Jongho Lee","Se Young Chun"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The mutual-information bound needs the projected-mean condition","plainExplanation":"Uncentered projected energy can come from a modality mean even when modalities are independent. A Gaussian example meets the theorem’s listed conditions and gives zero mutual information against a positive claimed bound. Carrying Lemma 1’s projected-mean orthogonality into the theorem, or using a centered covariance overlap, repairs the argument’s missing step.","primaryUrl":"https://arxiv.org/html/2610.04580","noteUrl":"/review?paper=review-neurips-2026-153529","collection":"neurips-2026","severity":"major","issueTypes":["incorrect_bound","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-153530","title":"Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation","authors":["Yongcan Yu","Jian Liang","Lingxiao He","Kuangpu Guo","Yanbo Wang","Shuo Lu","Meng Wang","Ran He"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.17607","noteUrl":"/review?paper=review-neurips-2026-153530","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153531","title":"Structure Over Scale: Learning Visual Reasoning from Pedagogical Video","authors":["Bishoy Galoaa","Xiangyu Bai","Sarah Ostadabbas"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Structure Over Scale: Learning Visual Reasoning from Pedagogical Video","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.23251","noteUrl":"/review?paper=review-neurips-2026-153531","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153532","title":"Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology","authors":["Jesseba Fernando","Grigori Guitchounts"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Entropy effective rank is not an upper bound on algebraic information rank","plainExplanation":"Entropy effective rank measures concentration rather than a hard count of transmitted independent coordinates. The dense Gaussian control also fails to preserve the trained spectrum, unlike a strictly upper-triangular perturbation in its Schur basis.","primaryUrl":"https://arxiv.org/html/2605.14258","noteUrl":"/review?paper=review-neurips-2026-153532","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153533","title":"Self-Trained Verification for Training- and Test-Time Self-Improvement","authors":["Chen Wu","Aditi Raghunathan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Self-Trained Verification for Training- and Test-Time Self-Improvement","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30290","noteUrl":"/review?paper=review-neurips-2026-153533","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153541","title":"The Limits of AI-Driven Allocation: Optimal Screening under Aleatoric Uncertainty","authors":["Santiago Cortes-Gomez","Mateo Dulce Rubio","Carlos M Patiño","Bryan Wilder"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Full support does not eliminate atoms at a screening boundary","plainExplanation":"Full support does not justify untied screening intervals when the risk law has atoms. The printed tail-mass indexing separately collapses every band boundary to the first sorted entry.","primaryUrl":"https://arxiv.org/html/2605.07979","noteUrl":"/review?paper=review-neurips-2026-153541","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153542","title":"Concept Modulation Models: A Unified Framework for Identifiability and Extrapolation","authors":["Soheun Yi","Yizhou Lu","Chandler Squires","Pradeep Ravikumar"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Concept Modulation Models: A Unified Framework for Identifiability and Extrapolation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.18509","noteUrl":"/review?paper=review-neurips-2026-153542","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153543","title":"Robust Dreamer: Deviation-Aware Latent Gaussian Memory for Action-Controlled AR Video Generation","authors":["Hanlin Chen","Jiaxin Wei","Xibin Song","Yifu Wang","Senbo Wang","Hongdong Li","Pan Ji","Gim Hee Lee"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A normalized temporal Gaussian density can exceed opacity one","plainExplanation":"The density peaks at \\(1/\\sqrt{2\\pi v}\\), exceeding one for small positive variance \\(v\\). At a spatial center with opacity \\(0.5\\), an allowed temporal scale can give \\(\\alpha\\approx1.9947\\), so the next transmittance factor \\(1-\\alpha\\) is negative. A probability density is not a probability bounded by one. The written alpha definition lacks the required range control. The expressly unnormalized spatial kernel suggests a possible local correction, not a verified implementation error.","primaryUrl":"https://arxiv.org/html/2605.30855","noteUrl":"/review?paper=review-neurips-2026-153543","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153545","title":"EventLens: Event-Structure Reinforcement Learning for Video Understanding","authors":["Guowen Zhang","Boshen Xu","Zihao Yue","Ziheng Wang","Xiaokun Liu","Xin Tao","Wenyu Qin","Pengfei Wan","Qin Jin"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"EventLens: Event-Structure Reinforcement Learning for Video Understanding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16172","noteUrl":"/review?paper=review-neurips-2026-153545","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153546","title":"Fine-tuning language encoding models on slow fMRI improves prediction for fast ECoG","authors":["Aditya Vaidya","Richard Antonello","Alexander Huth"],"year":2026,"category":"Machine learning","status":"unresolved","issueTitle":"Fine-tuning language encoding models on slow fMRI improves prediction for fast ECoG","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.19224","noteUrl":"/review?paper=review-neurips-2026-153546","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153547","title":"Evaluating and Understanding Scheming Propensity in LLM Agents","authors":["Mia Hopman","Jannes Elstner","Maria Avramidou","Amritanshu Prasad","David Lindner"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Evaluating and Understanding Scheming Propensity in LLM Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.01608","noteUrl":"/review?paper=review-neurips-2026-153547","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153550","title":"Characterizing the Generalization Error of Random Feature Regression with Arbitrary Data-Augmentation","authors":["Lucas Morisset","Alain Durmus","Adrien Hardy"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Characterizing the Generalization Error of Random Feature Regression with Arbitrary Data-Augmentation","plainExplanation":"A marginal concentration proxy is not an exact variance, and marginal centering does not make the noise orthogonal to feature functions.","primaryUrl":"https://arxiv.org/html/2605.10290","noteUrl":"/review?paper=review-neurips-2026-153550","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","mathematical_error","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153551","title":"OmniSelect: Dynamic Modality-Aware Token Compression for Efficient Omni-modal Large Language Models","authors":["Morunliu Yang","Ruotao Xu","Le Li","Yue Wang","Jianxin Zhang","Siwei Feng","Peifeng Li","Yihang Lou","Juntao Li"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"OmniSelect: Dynamic Modality-Aware Token Compression for Efficient Omni-modal Large Language Models","plainExplanation":"The final formula and pseudocode do not use the normalized allocations the preceding equation already provides. Raw sigmoids have no reason to meet an arbitrary target budget. The relevant v1/v2 definitions remain the same; the existing source-linked code check already found the rescaling in the implementation.","primaryUrl":"https://arxiv.org/html/2605.18041","noteUrl":"/review?paper=review-neurips-2026-153551","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","implementation_mismatch"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153553","title":"Learning to Generate Multiple Objects from Dense and Occluded Layouts","authors":["Bach Hoang Ngo","Ngo Tri","Hieu Minh Le","Trung Nghia Le"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mean-one scalar weights do not preserve global gradient norms","plainExplanation":"Mean-one scalar weights do not preserve global gradient norms. The global gradient-scale guarantee is false; the stated scalar-weight normalization remains valid.","primaryUrl":"https://arxiv.org/html/2607.03488v1","noteUrl":"/review?paper=review-neurips-2026-153553","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153555","title":"Improved Regret Analysis For Parallel Gaussian Process Bandit Optimization","authors":["Shion Takeno","Shogo Iwazaki"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Improved Regret Analysis For Parallel Gaussian Process Bandit Optimization","plainExplanation":"The previous point in a modulo-\\(Q\\) class occurs at \\(t-Q\\), which is not yet in the conditioning set. The sequential information-gain identity therefore uses smaller variances than those in the displayed sum.","primaryUrl":"https://arxiv.org/html/2608.16492","noteUrl":"/review?paper=review-neurips-2026-153555","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153556","title":"Why Muon Outperforms Adam: A Curvature Perspective","authors":["Shuche Wang","Fengzhuo Zhang","Jiaxiang Li","Dirk Bergemann","Zhuoran Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Why Muon Outperforms Adam: A Curvature Perspective","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04662","noteUrl":"/review?paper=review-neurips-2026-153556","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153558","title":"Metaphor Is Not All Attention Needs","authors":["Olga Sorokoletova","Francesco Giarrusso","Giacomo De Luca","Piercosma Bisconti","Matteo Prandi","Federico Pierucci","Marcello Galisai","Vincenzo Suriani","Daniele Nardi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Metaphor Is Not All Attention Needs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12128","noteUrl":"/review?paper=review-neurips-2026-153558","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153559","title":"Revitalizing Medical Time Series with Vision-Informed Retrieval: A Vision-Language Perspective","authors":["Guoqi Yu","Juncheng Wang","Emma, Shujun Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A remainder patch does not fit the shared fixed-size projection","plainExplanation":"A remainder patch does not fit the shared fixed-size projection. The projection specification lacks its nondivisible-length convention; no runtime failure is established.","primaryUrl":"https://arxiv.org/html/2609.34652","noteUrl":"/review?paper=review-neurips-2026-153559","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153564","title":"One-Layer Transformers Provably Learn In-Context K-Nearest Neighbor Prediction with Chain-of-Thought","authors":["Lyumin Wu","Yuan Cao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"One-Layer Transformers Provably Learn In-Context K-Nearest Neighbor Prediction with Chain-of-Thought","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.16798","noteUrl":"/review?paper=review-neurips-2026-153564","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153565","title":"FreshMem: Brain-Inspired Frequency-Space Hybrid Memory for Streaming Video Understanding","authors":["Kangcong Li","Peng Ye","Lin Zhang","Chao Wang","Huafeng Qin","Jiayuan Fan","Tao Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Threshold-only consolidation does not enforce a hard memory cap","plainExplanation":"Choose forty-one unit episode centroids in \\(\\mathbb R^{41}\\), mutually orthogonal, with sixteen repeated frames per episode. Each transition opens a new episode because cosine zero is below \\(0.4\\); the minimum retention rate keeps one frame. Adjacent cosine zero is also below merge threshold \\(0.3\\). When the forty-slot STM fills, no permitted threshold merge frees a slot. The described threshold rule does not prove a hard cap. The separate bounded memory component does not cap STM, and actual overflow is not established.","primaryUrl":"https://arxiv.org/html/2602.01683","noteUrl":"/review?paper=review-neurips-2026-153565","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153572","title":"Temporal Consistency Improves Generalization in Contextual Offline Meta Reinforcement Learning","authors":["Mohammadreza Nakhaeinezhad","Aidan Scannell","Kevin Luck","Joni Pajarinen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The learned-model value comparison needs a learned reward bound","plainExplanation":"The printed expectile loss lacks an absolute weight and can be unbounded below. The learned-MDP value lemma also uses the original reward bound without controlling the models’ shared reward magnitude.","primaryUrl":"https://arxiv.org/html/2603.02935","noteUrl":"/review?paper=review-neurips-2026-153572","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153573","title":"Disentangling generalization and memorization in large language models using chess","authors":["Leonard S. Pleiss","Maximilian Schiffer","Robert K von Weizsäcker"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Disentangling generalization and memorization in large language models using chess","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.16823","noteUrl":"/review?paper=review-neurips-2026-153573","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153577","title":"Stochastic Zeroth-Order Optimization Under Heavy-Tailed Noise","authors":["Taha EL BAKKALI EL KADI","El Mahdi Chayti","Qiuyi (Richard) Zhang","Imane Rahali","Omar Saadi"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Stochastic Zeroth-Order Optimization Under Heavy-Tailed Noise","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17394","noteUrl":"/review?paper=review-neurips-2026-153577","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153580","title":"BlockFormer: Transformer-based inference from interaction maps","authors":["Eloïse Touron","Pedro Rodrigues","Julyan Arbel","Nelle Varoquaux","Michael Arbel"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Reset-to-parent proposals create atoms missing from the importance density","plainExplanation":"For a continuous prior on \\([0,1]\\), resetting out-of-bounds proposals to parent \\(a\\) creates atom mass \\(P(a+\\epsilon\\notin[0,1])>0\\). The continuous target has zero mass at that atom. The printed finite density denominator nevertheless assigns it a positive weight, rather than the zero Radon–Nikodym weight required there. An acceptance score independent of the parameter preserves this defect through top-fraction filtering. The boundary treatment and importance weights are inconsistent. No numerical empirical posterior failure is asserted.","primaryUrl":"https://arxiv.org/html/2605.21617v3","noteUrl":"/review?paper=review-neurips-2026-153580","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153581","title":"Revisiting Decentralized Online Convex Optimization with Compressed Communication","authors":["Hao Zhou","Xiaoyu Wang","Chang Yao","Mingli Song","Yuanyu Wan"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Consensus error depends on earlier increments, not only the latest one","plainExplanation":"The local consensus estimate drops historical increments, and the explicit regret bound loses its first-block contribution at $T=L$. Both errors have local history/initialization repairs without disproving the asymptotic rate.","primaryUrl":"https://arxiv.org/html/2607.01665","noteUrl":"/review?paper=review-neurips-2026-153581","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153582","title":"Model Guides You How to Draw: Adaptive Visual Gating for Unified Multimodal Reasoning","authors":["Wenxi Gao","Guanxi Lu","Didi Zhu","Hao Chen","Quan Deng","Zhican Wang","Jiankang Deng","Hongxiang Fan"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Model Guides You How to Draw: Adaptive Visual Gating for Unified Multimodal Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.10004","noteUrl":"/review?paper=review-neurips-2026-153582","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153588","title":"L2P: Unlocking Latent Potential for Pixel Generation","authors":["Zhennan Chen","Junwei Zhu","Xu Chen","Jiangning Zhang","Jiawei Chen","Zhuoqi Zeng","Wei Zhang","Chengjie Wang","Jian Yang","Ying Tai"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"L2P: Unlocking Latent Potential for Pixel Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12013","noteUrl":"/review?paper=review-neurips-2026-153588","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153590","title":"UniVR: Thinking in Visual Space for Unified Visual Reasoning","authors":["Zhongwei Ren","Yunchao Wei","Yao Zhao","GONG WEIBO","Xiao Liu","Anran Wang","Xiangtai Li","Xiaojie Jin"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"UniVR: Thinking in Visual Space for Unified Visual Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.12800","noteUrl":"/review?paper=review-neurips-2026-153590","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153592","title":"Boosting LLM Reasoning via Human-Inspired Reward Shaping","authors":["Wenze Lin","Zhen Yang","Xitai Jiang","Xiaoteng Ma","Gao Huang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Boosting LLM Reasoning via Human-Inspired Reward Shaping","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.04265","noteUrl":"/review?paper=review-neurips-2026-153592","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153594","title":"DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control","authors":["Teli Ma","Jia Zheng","Zifan Wang","Chunli Jiang","Andy Cui","Junwei Liang","Shuo Yang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The feature-extraction timestep is fixed in prose but random in the algorithm","plainExplanation":"Algorithm 1 separately samples the extraction timestep on a uniform grid of \\(1001\\) points. For a timestep-sensitive hidden feature, that produces nonzero training-time variation, whereas the fixed-time prescription produces none. Random prediction time and random latent inputs are separate variables and do not reconcile this extraction-time conflict. This is a fixed-versus-random specification inconsistency, not a claim of failed experiments.","primaryUrl":"https://arxiv.org/html/2603.10448","noteUrl":"/review?paper=review-neurips-2026-153594","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153596","title":"Fast Organic Crystal Structure Prediction with Unit Cell Flow Matching","authors":["Alston Lo","Luka Mucko","Austin Cheng","Andy Cai","Alastair J Price","Wojciech Matusik","Alan Aspuru-Guzik"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Independent normalized lengths and angles need not match the sampled density","plainExplanation":"Independent normalized lengths and angles need not match the sampled density. The claimed simultaneous source constraints are inconsistent; the implementation’s chosen reconciliation is not established.","primaryUrl":"https://arxiv.org/html/2606.03199","noteUrl":"/review?paper=review-neurips-2026-153596","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153597","title":"Theoretical Limits of Language Model Alignment","authors":["Lucas Monteiro Paes","Natalie Mackraz","Barry-John Theobald","Federico Danieli"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"An exponential normalized tilt does not force exponential gold-reward degradation","plainExplanation":"An exponential normalized tilt does not force exponential gold-reward degradation. The exact tilt identity remains valid; universal degradation does not follow, and worst-case amplification is not ruled out.","primaryUrl":"https://arxiv.org/html/2605.07105","noteUrl":"/review?paper=review-neurips-2026-153597","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153598","title":"STAR: Boosting Time Series Foundation Models for Anomaly Detection Through State-Aware Adapter","authors":["Hanyin Cheng","Ruitong Zhang","Yuning Lu","Yang Shu","Peng Chen","Yuan Jun","Meng Wang","Bin Yang","Chenjuan Guo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The matching factor downweights cross-modal mismatch","plainExplanation":"The matching factor downweights cross-modal mismatch. The matching-score orientation conflicts with its interpretation; reconstruction differences may still dominate actual benchmark rankings.","primaryUrl":"https://arxiv.org/html/2510.16014","noteUrl":"/review?paper=review-neurips-2026-153598","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153599","title":"Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing","authors":["Michał Brzozowski","Zuzanna Dubanowska","Enrico Cassano","Neo Christopher Chung"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.25902","noteUrl":"/review?paper=review-neurips-2026-153599","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153603","title":"Structured Coupling for Flow Matching","authors":["Francisco Sumba Toral","Carles Balsells Rodas","Yingzhen Li"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Structured Coupling for Flow Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07676","noteUrl":"/review?paper=review-neurips-2026-153603","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153604","title":"Controlling for Omitted Variable Bias in Deep Neural Networks","authors":["Manuel Pfeuffer","Roshan P Rane","Kerstin Ritter","Sonja Greven"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Controlling for Omitted Variable Bias in Deep Neural Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.25930","noteUrl":"/review?paper=review-neurips-2026-153604","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"aerochem-open-loop-error-bound","title":"AeroChem: Closed-loop Physics-Informed State Space Modeling for Long-term Chemically-Reactive Air Quality Forecasting","authors":["Le Trung Kien","Tung Kieu","Dinh D Nguyen","Cuong Do","Bin Yang","Phi Long Nguyen"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A contracting system contradicts the error lower bound","plainExplanation":"An upper Lipschitz condition permits the contracting dynamics $F(s)=-s$, so it cannot imply the claimed open-loop error growth. Separately, the Newton-rate lemma omits the positive-definite-Hessian condition that its own proof invokes.","primaryUrl":"https://arxiv.org/abs/2610.16630v1","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","incorrect_bound","missing_assumption"],"classifications":["critical"]},{"slug":"review-neurips-2026-153608","title":"Most ReLU Networks Admit Identifiable Parameters","authors":["Moritz Grillo","Guido Montufar"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Most ReLU Networks Admit Identifiable Parameters","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.03601","noteUrl":"/review?paper=review-neurips-2026-153608","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153609","title":"Automatically Refining Coding Rules for AI Coding Agents","authors":["Zhengyuan Jiang","Reachal Wang","Yuepeng Hu","Yupu Wang","Yuqi Jia","Neil Gong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Independent rounding does not preserve the exact mutation budget","plainExplanation":"Independent rounding does not preserve the exact mutation budget. The printed allocation routine does not guarantee its exact count; this is not an assessment of external code.","primaryUrl":"https://arxiv.org/html/2610.00650","noteUrl":"/review?paper=review-neurips-2026-153609","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153610","title":"Edge of Stability Selectively Shapes Learning Across the Data Distribution","authors":["Shauna Kwag","Anakha Ganesh","Tomaso Poggio","Pierfrancesco Beneventano"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Edge of Stability Selectively Shapes Learning Across the Data Distribution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04212","noteUrl":"/review?paper=review-neurips-2026-153610","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153611","title":"Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization","authors":["Chengjun Yu","Shu XU","Jiaqi Wu","Qianben Chen","Tianrui Qin","Zhu","Qiexiang Wang","Jiayu Zhang","Xinpeng Liu","Xin Gui","Jingyi Cao","Yi Yao","WANG PIAOHONG","Dingfeng Shi","He Zhu","Tiannan Wang","Yuqing Wang","Maojia Song","Tianyu Zheng","Jian Yang","Jiaheng Liu","Minghao Liu","Eleanor Jiang","Wangchunshu Zhou"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.22675","noteUrl":"/review?paper=review-neurips-2026-153611","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153612","title":"Persona-Model Collapse in Emergent Misalignment","authors":["Davi Bastos Costa","Renato Vicente"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Persona-Model Collapse in Emergent Misalignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12850","noteUrl":"/review?paper=review-neurips-2026-153612","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153613","title":"Penalty-Based First-Order Methods for Bilevel Optimization with Minimax and Constrained Lower-Level Problems","authors":["Yiyang Shen","Yutian He","Weiran Wang","Qihang Lin"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The outer theorem uses a fixed dual anchor rather than the algorithm’s moving anchor","plainExplanation":"Compactness and closed convexity do not supply the required finite lower objective bound. The general finiteness premise fails; this does not invalidate every associated maximized value function.","primaryUrl":"https://arxiv.org/html/2605.08006","noteUrl":"/review?paper=review-neurips-2026-153613","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153616","title":"Deep Reasoning in General Purpose Agents via Structured Meta-Cognition","authors":["Dean Light","Michael Theologitis","Kshitish Ghate","Stella Li","Benjamin Newman","Chirag Shah","Aylin Caliskan","Pang Wei Koh","Dan Suciu","Yulia Tsvetkov"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Deep Reasoning in General Purpose Agents via Structured Meta-Cognition","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11388","noteUrl":"/review?paper=review-neurips-2026-153616","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153617","title":"Beyond Pessimism: Offline Learning in KL-regularized Games","authors":["Yuheng Zhang","Claire Chen","Nan Jiang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Moving the player maximum outside expectation loses a factor","plainExplanation":"With two equally likely contexts, zero true payoff and reference probabilities \\((0.1,0.9)\\), estimated payoff matrices \\(\\begin{pmatrix}1&1\\\\0&0\\end{pmatrix}\\) and \\(\\begin{pmatrix}1&0\\\\1&0\\end{pmatrix}\\) give unilateral errors \\((1,0.1)\\) and \\((0.1,1)\\). The expected squared maximum is one. Behavior MSE is \\(0.5\\), while each fixed-player concentrability expectation is \\(1.1\\), making the printed upper bound \\(0.55\\). The worst player changes with context. The missing factor affects a proof step. The stated big-O sample rates already hide a universal factor and are not refuted.","primaryUrl":"https://arxiv.org/html/2604.06738","noteUrl":"/review?paper=review-neurips-2026-153617","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153618","title":"Privately Estimating Monotone Statistics in Polynomial Time","authors":["Gavin R Brown","Ephraim Linder","Mahbod Majid","Vikrant Singhal"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The binomial-ratio upper bound is below an elementary lower bound","plainExplanation":"The binomial-ratio bound is smaller than a direct lower bound on its factors. The median mechanism also becomes uniform on a constant-statistic example because its score direction and tied-rank treatment do not identify the median interval.","primaryUrl":"https://arxiv.org/html/2605.27912","noteUrl":"/review?paper=review-neurips-2026-153618","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153619","title":"Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift","authors":["Kevin Ren","Manish Raghavan","Nikhil Garg"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.25152","noteUrl":"/review?paper=review-neurips-2026-153619","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153620","title":"Likelihood-free inference of phylogenetic tree posterior distributions","authors":["Luc Blassel","Noémie Sauvage","Pierre Barrat-Charlaix","Bastien Boussau","Nicolas Lartillot","Laurent Jacob"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The written tree density needs a terminal branch law and consistent branch transformations","plainExplanation":"The product over $2N-3$ merges, the prose specifying $N-3$ merges, and the algorithms performing $N-2$ merges do not define the same full-tree density. The algorithms leave one connecting edge unspecified.","primaryUrl":"https://arxiv.org/html/2510.12976","noteUrl":"/review?paper=review-neurips-2026-153620","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153622","title":"CaC: Advancing Video Reward Models via Hierarchical Spatiotemporal Concentrating","authors":["Jiyuan Wang","Huan Ouyang","Chunyu Lin","Dewen Fan","Boheng Zhang","Tingting Gao","Fan Yang","Jia Sun","Zijun Li","Yongrui Heng","Huaiqing Wang","Zhenlong Yuan","Yiyang Fan","Honglie Wang","Fei Zuo","Haonan fan","Jiuzhou Lin","Guosheng Lin"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"CaC: Advancing Video Reward Models via Hierarchical Spatiotemporal Concentrating","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11723","noteUrl":"/review?paper=review-neurips-2026-153622","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153623","title":"Flow Matching with In-Context Priors for Out-of-Distribution Brain Dynamics","authors":["Sam Gijsen","Michał Łukomski","Marc-Andre Schulz","Kerstin Ritter"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The end-to-end endpoint loss admits a data-independent zero solution","plainExplanation":"The end-to-end endpoint loss admits a data-independent zero solution. The stated end-to-end objective is degenerate/incomplete; actual training collapse is not inferred.","primaryUrl":"https://arxiv.org/html/2606.11833","noteUrl":"/review?paper=review-neurips-2026-153623","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153625","title":"Cross-Family Universality of Behavioral Axes via Anchor-Projected Representations","authors":["Su-Hyeon Kim","Yo-Sub Han"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Cross-Family Universality of Behavioral Axes via Anchor-Projected Representations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.09875","noteUrl":"/review?paper=review-neurips-2026-153625","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153629","title":"Causal Abstractions, Categorically Unified","authors":["Markus Englberger","Devendra Singh Dhami"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"At-most-once generator use does not specify an intervention normal form","plainExplanation":"At-most-once mechanism use does not enforce the intervention normal form needed for morphism uniqueness. Deterministic natural abstraction maps also need not be surjective, so their claimed stochastic right inverses do not always exist.","primaryUrl":"https://arxiv.org/html/2510.05033","noteUrl":"/review?paper=review-neurips-2026-153629","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-153630","title":"GEOPHYS: The Geometry of Physical Plausibility","authors":["Christian Internò","Alexander Pondaven","Habon Issa","Fabio Pizzati","Francesco Pinto","Markus Olhofer","Ivan Laptev","Philip Torr","Eero Simoncelli","Barbara Hammer","David Klindt"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Autoregressive prediction error is not a current-history orthogonal projection","plainExplanation":"The fitted autoregressive residual is not generally orthogonal to the current history span. Least-squares orthogonality across a training design does not justify the per-example geometric projection used in the interpretation.","primaryUrl":"https://arxiv.org/html/2606.20707","noteUrl":"/review?paper=review-neurips-2026-153630","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153632","title":"From Persistence to Survival: Hypothesis Testing, Effect Sizes and Vectorisation for Topological Features","authors":["Juliette Murris","Bernadette J Stolz","Karsten Borgwardt"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Grid evaluation is discontinuous when a persistence value crosses a threshold","plainExplanation":"A small movement of one feature across a threshold changes its survival indicator from $0$ to $1$, producing a vector jump of $1/N$. Distinct lifetimes prevent features from coinciding with each other, not with the grid. The proof also replaces a cross-diagram matching displacement by a within-diagram separation, although the former can tend to zero while the latter stays positive. The example refutes Corollary 3.2 and the continuity inference following Theorem 3.1 for fixed-grid survival vectors. It satisfies Theorem 3.1's data-dependent numerical bound, whose denominator $\\delta$ shrinks with the perturbation. The repair proves local constancy for a fixed grid, fixed cardinality and positive margin $m$; it does not extend to variable cardinality, arbitrary data-dependent grids or downstream classifier robustness.","primaryUrl":"https://arxiv.org/html/2606.11911","noteUrl":"/review?paper=review-neurips-2026-153632","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153633","title":"AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems","authors":["Jaewon Chu","jinwoo seo","Jaewon Cho","Jeehye Na","Yunyang Xiong","Youngdae Kim","Hyunwoo J. Kim"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.08572","noteUrl":"/review?paper=review-neurips-2026-153633","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153635","title":"Step-by-Step Optimization-like Reasoning in LLMs over Expanding Search Spaces","authors":["Nicolás Astorga","Nabeel Seedat","Mihaela van der Schaar"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Two untested classes do not by themselves make filtering strictly more informative","plainExplanation":"Leaving two feasible classes untested does not ensure a strict filtering gain. An already feasible, nonrepeating proposal gives identical transcripts and information with and without filtering.","primaryUrl":"https://arxiv.org/html/2606.05464","noteUrl":"/review?paper=review-neurips-2026-153635","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153636","title":"A Unified Framework for Adversary-Aware Differential Privacy Bounds","authors":["Marika Swanberg","Meenatchi Sundaram Muthu Selva Annamalai","Jamie Hayes","Borja Balle","Adam Smith"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A mean-optimal prior attack need not maximize success-count tails","plainExplanation":"The prior attack that maximizes expected successes need not maximize every success-count tail. A two-action example under a constant private mechanism violates the tail bound built from the mean-optimal action.","primaryUrl":"https://arxiv.org/html/2507.08158","noteUrl":"/review?paper=review-neurips-2026-153636","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153637","title":"Minimizing Modality Gap from the Input Side: Your Speech LLM Can Be a Prosody-Aware Text LLM","authors":["Wenqian Cui","Xiao-Hui Li","Daxin Tan","Qiyong Zheng","Irwin King"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Minimizing Modality Gap from the Input Side: Your Speech LLM Can Be a Prosody-Aware Text LLM","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05927","noteUrl":"/review?paper=review-neurips-2026-153637","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153641","title":"Differentiable Knapsack and Top-k Operators via Dynamic Programming","authors":["Germain Vivier-Ardisson","Michael E Sander","Axel Parmentier","Mathieu Blondel"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Zero-weight items expose a missing zero-capacity partition state","plainExplanation":"The backward algorithms start with an out-of-range row access. Independently, the fixed zero-capacity initialization loses feasible choices for zero-weight items, violating the stated Gibbs and permutation-equivariance properties.","primaryUrl":"https://arxiv.org/html/2601.21775","noteUrl":"/review?paper=review-neurips-2026-153641","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153642","title":"Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images","authors":["Qiwei Wang","Zhongyao Tuo","Xianghui Ze","Yujiao Shi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07978","noteUrl":"/review?paper=review-neurips-2026-153642","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153645","title":"Deep Research as Rubric","authors":["Wangyi Mei","Zhouhong Gu","zhenhanbai","Yin Cai","Lefan Zhang","Zhenxin Ding","Bo chen","Yan Gao","YIWU","Yao Hu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Deep Research as Rubric","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.01091v1","noteUrl":"/review?paper=review-neurips-2026-153645","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153647","title":"HomeFlow: A Data Flywheel for Smart Home Agent Training with Verifiable Simulation","authors":["Yi Gu","Huacan Wang","Shuo Zhang","Yuqing Hou","Xuelei","weipeng.ming","Chen Liu","Fangzhou Yu","Kuan Li","Ronghao Chen","Sen Hu","Mou X Feng","Yi Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"HomeFlow: A Data Flywheel for Smart Home Agent Training with Verifiable Simulation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.01230","noteUrl":"/review?paper=review-neurips-2026-153647","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153648","title":"Consilience for Verifier-Free Test-Time Scaling","authors":["Lecheng Kong","Like Hui","Haitao Mao","Luke Huan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The unnormalized top-token negative-log score need not increase with concentration","plainExplanation":"The unnormalized top-token negative-log score need not increase with concentration. The universal concentration reading is false; conditional usefulness on actual completion pools remains possible.","primaryUrl":"https://arxiv.org/html/2608.09898","noteUrl":"/review?paper=review-neurips-2026-153648","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153649","title":"ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design","authors":["Yulin Zhang","He CAO","Zihao Jiang","Chenyi Zi","Zhipeng Zhou","Zijing Liu","Yu Li","Jia Li","Ziqi Gao"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10189","noteUrl":"/review?paper=review-neurips-2026-153649","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153650","title":"High-arity Sample Compression","authors":["Leonardo Coregliano","William Opich"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The explicit guarantee retains a nonvanishing header-complexity factor","plainExplanation":"The explicitly defined learning guarantee retains a header-complexity factor in its leading asymptotic. A two-header example satisfies the stated growth assumptions but has limiting ratio $1+\\log2$ instead of one.","primaryUrl":"https://arxiv.org/html/2605.12465","noteUrl":"/review?paper=review-neurips-2026-153650","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153652","title":"Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation","authors":["Jacob Dang","Brian Y Xie","Omar G. Younis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.15559","noteUrl":"/review?paper=review-neurips-2026-153652","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153655","title":"MvFFN: Multi-view Floor-Plan Feed-Forward Network for Unposed Wide-Baseline Panorama Layout Reconstruction","authors":["Yuguang Li","Yichuan (Ethan) Deng","Zixuan Liu","Ivaylo Boyadzhiev","Xiuyuan Wu","Xinyu Zou","Linda Shapiro","Alex Colburn"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Cross-view wall masks need a common coordinate domain","plainExplanation":"Equal column indices in differently oriented panoramas do not denote the same physical rays or walls.","primaryUrl":"https://arxiv.org/html/2610.14315","noteUrl":"/review?paper=review-neurips-2026-153655","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_ambiguity"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153657","title":"Unlocking Compositional Generalization in Continual Few-Shot Learning","authors":["Phu-Quy Nguyen-Lam","Phu-Hoa Pham","Dao S Minh","Chi Nguyen Tran","Trung-Kiet Huynh","Long Tran-Thanh"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Collision-free greedy matching is not a generic property","plainExplanation":"Column-collision avoidance is not generic, shared upstream gradients can become negatively aligned after slot tangent projections, and diagonalizing a singular covariance does not produce a defined zero standardized-correlation loss. The explicitly collision-free argument, upstream rank-one identity, and covariance diagonalization remain intact. The attention-purity comparison is a separate reporting exclusion.","primaryUrl":"https://arxiv.org/html/2605.11710","noteUrl":"/review?paper=review-neurips-2026-153657","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","false_statement","interpretation_error","invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153660","title":"Topology-Aware Representation Alignment for Semi-Supervised Vision-Language Learning","authors":["Junwon You","Mihyun Jang","Sangwoo Mo","Jae-Hun Jung"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Topology-Aware Representation Alignment for Semi-Supervised Vision-Language Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.26370","noteUrl":"/review?paper=review-neurips-2026-153660","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153662","title":"Finite Resources False Discovery Rate Control on Structured Hypothesis Spaces","authors":["Binyamin Perets","Shie Mannor"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An arbitrary spatial gate can destroy Rule 1’s exact FDR control","plainExplanation":"Rule 1 can reject only a null with probability one half at a stated level of 0.4. Its exact arbitrary-estimator guarantee therefore fails. Separately, interior score gaps shrink with growing count budgets, the Fisher saturation formula switches in the wrong direction, and the unconditional latent-p posterior needs both mixture components. These defects have distinct repairs; the fixed-budget gap and flip-invariant folded-rule argument are not refuted.","primaryUrl":"https://arxiv.org/html/2606.15393","noteUrl":"/review?paper=review-neurips-2026-153662","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","formula_error","incorrect_bound","invalid_inference","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-153664","title":"Fingerprinting Inference Systems of Large Language Models","authors":["Anna Wimbauer","Jonas Möller","Erik Imgrund","Konrad Rieck"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Fingerprinting Inference Systems of Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29979","noteUrl":"/review?paper=review-neurips-2026-153664","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153666","title":"Discrete Stochastic Localization for Non-autoregressive Generation","authors":["Yunshu Wu","Jiayi Cheng","Longxuan YU","Partha Thakuria","Rob Brekelmans","Vagelis Papalexakis","Greg Ver Steeg"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Discrete Stochastic Localization for Non-autoregressive Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2602.16169v2","noteUrl":"/review?paper=review-neurips-2026-153666","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153667","title":"A Unifying View of Anchoring via Operator-Side Tikhonov Regularization","authors":["Zihao Chen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Unifying View of Anchoring via Operator-Side Tikhonov Regularization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30905","noteUrl":"/review?paper=review-neurips-2026-153667","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153669","title":"NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution","authors":["Xiangtao Kong","Rongyuan Wu","Shuaizheng Liu","Lingchen Sun","Lei Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The scalar binary cross entropy omits the negative-label term","plainExplanation":"The scalar binary cross entropy omits the negative-label term. The scalar display is incomplete; the inherited training objective could already be implemented correctly.","primaryUrl":"https://arxiv.org/html/2510.00820","noteUrl":"/review?paper=review-neurips-2026-153669","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153670","title":"Hyperagents","authors":["Jenny Zhang","Bingchen Zhao","Wannan Yang","Jakob Foerster","Jeff Clune","Minqi Jiang","Sam Devlin","Tatiana Shavrina"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Hyperagents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.19461","noteUrl":"/review?paper=review-neurips-2026-153670","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153671","title":"Beyond Scalar Distances: Semantic Attribute Gradients from Frozen MLLMs for Visual Embeddings","authors":["Shubhang Bhatnagar","Dheeraj Baiju","Narendra Ahuja"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Equal token probabilities do not imply cancellation of their score gradients","plainExplanation":"Equal token probabilities do not imply cancellation of their score gradients. The probability premise is too weak for the derivative conclusion; empirical performance is not refuted.","primaryUrl":"https://arxiv.org/html/2606.15134","noteUrl":"/review?paper=review-neurips-2026-153671","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153672","title":"Test-Time Personalization: A Diagnostic Framework and Probabilistic Fix for Scaling Failures","authors":["Linhai Zhang","Yulan He"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Heteroscedastic NLL does not force a different optimal conditional mean","plainExplanation":"Heteroscedastic NLL does not force a different optimal conditional mean. The strict population improvement does not follow; empirical personalization gains remain possible.","primaryUrl":"https://arxiv.org/pdf/2605.10991v1","noteUrl":"/review?paper=review-neurips-2026-153672","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153673","title":"Diffusion-State Policy Optimization for Masked Diffusion Language Models","authors":["Daisuke Oba","Hiroki Furuta","Naoaki Okazaki"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A self-including group baseline scales the terminal reward gradient","plainExplanation":"For \\(K\\) iid rollouts, a group mean including rollout \\(i\\) has score covariance \\(K^{-1}\\nabla J\\). Subtracting it therefore leaves \\((K-1)\\nabla J/K\\), rather than \\(\\nabla J\\). This discrepancy already occurs at the behavior-policy point with exact importance ratios and no clipping, because the baseline includes the scored rollout’s own reward. This is a scaling and mixed-objective issue in the unbiasedness claim, not a verified implementation defect.","primaryUrl":"https://arxiv.org/html/2602.06462","noteUrl":"/review?paper=review-neurips-2026-153673","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153674","title":"Tool Verification for Test-Time Reinforcement Learning","authors":["Ruotong Liao","Nikolai Röhrich","Xiaohan Wang","Yuhui Zhang","Yasaman Samadzadeh","Volker Tresp","Serena Yeung-Levy"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Tool Verification for Test-Time Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.02203","noteUrl":"/review?paper=review-neurips-2026-153674","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153677","title":"Tessellations of Semi-Discrete Flow Matching","authors":["Emile Pierret","Johannes Hertrich","Samuel Hurault","Julie Delon"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Tessellations of Semi-Discrete Flow Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07513","noteUrl":"/review?paper=review-neurips-2026-153677","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153679","title":"MOSAIC: Module Discovery via Sparse Additive Identifiable Causal Learning for Scientific Time Series","authors":["Shicheng Fan","Nour Elhendawy","Jianle Sun","Ke Fang","Kun Zhang","Yihang Wang","Lu Cheng"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Three bridges from the model definition to the theory need stronger conditions","plainExplanation":"The change-of-variables expression counts the base density once per preimage. With multiple branches it need not integrate to one.","primaryUrl":"https://arxiv.org/html/2605.05524","noteUrl":"/review?paper=review-neurips-2026-153679","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153680","title":"InduceKV: Fixed-Footprint Continual Adaptation of Multimodal LLMs via Inducing KV Memories","authors":["Qianyu Chen","Ziteng Feng","Canran Xiao","Runxuan Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Dropping coverage requires its nonnegativity","plainExplanation":"Dropping coverage requires its nonnegativity. The unrestricted averaged bound is false; pointwise and experimental-setting bounds remain valid under their conditions.","primaryUrl":"https://arxiv.org/html/2607.02010","noteUrl":"/review?paper=review-neurips-2026-153680","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153681","title":"Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models","authors":["Rishabh Sharma","Stefano Martiniani"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.24000","noteUrl":"/review?paper=review-neurips-2026-153681","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153684","title":"Democratizing Tool Learning with Environments Fully Simulated by a Free 8B Language Model","authors":["Chenming Tang","Hsiu-Yuan Huang","Weijie Liu","Junqiang Zheng","Saiyong Yang","Yunfang Wu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Privileged feedback does not guarantee strict improvement","plainExplanation":"Privileged information can exist without an actionable improvement opportunity. The stated premises do not require a positive achievable return gap or an optimizer that closes it, so strict positivity does not follow.","primaryUrl":"https://arxiv.org/html/2604.17739","noteUrl":"/review?paper=review-neurips-2026-153684","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153687","title":"FloatDoor: Platform triggered Backdoors in LLMs","authors":["Nils Loose","Jonas Sander","Felix Mächtle","Thomas Eisenbarth"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Teacher-forced prefixes qualify the exclusive residual-tag explanation","plainExplanation":"Once an earlier target token differs, the teacher-forced prefix is another platform-dependent input. The exclusive-signal inference therefore does not hold for every attacked position.","primaryUrl":"https://arxiv.org/html/2606.19535","noteUrl":"/review?paper=review-neurips-2026-153687","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153688","title":"LatentRAG: Latent Reasoning and Retrieval for Efficient Agentic RAG","authors":["Yijia Zheng","Marcel Worring"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"LatentRAG: Latent Reasoning and Retrieval for Efficient Agentic RAG","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06285","noteUrl":"/review?paper=review-neurips-2026-153688","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153691","title":"FLASH: Efficient Visuomotor Policy via Sparse Sampling","authors":["Jiaqi Bai","Jindou Jia","Yuxuan Hu","Gen Li","Xiangyu Chen","Tuo An","Kuangji Zuo","Jianfei Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FLASH: Efficient Visuomotor Policy via Sparse Sampling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15492","noteUrl":"/review?paper=review-neurips-2026-153691","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153692","title":"Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex","authors":["Yun Qu","Qi Wang","Yixiu Mao","Heming Zou","Yuhang Jiang","Yingyue Li","Wutong Xu","Lizhou Cai","Weijie Liu","Clive Bai","Kai Yang","Yangkun Chen","Saiyong Yang","Xiangyang Ji"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06139","noteUrl":"/review?paper=review-neurips-2026-153692","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153693","title":"Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting","authors":["Fan Xu","Yuan Gao","Kun Wang","Rui Su","Fenghua Ling","Hao Wu","Wanli Ouyang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The K-squared ensemble formula is biased for population CRPS","plainExplanation":"Marginal CRPS is not an exact joint Wasserstein-one objective and cannot identify joint dependence. The finite-ensemble K-squared formula is also biased for population CRPS, although it is the exact score of the empirical ensemble distribution.","primaryUrl":"https://arxiv.org/html/2605.06916","noteUrl":"/review?paper=review-neurips-2026-153693","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153694","title":"Learning Orthogonal Multi-Index Models Beyond Small Initialization: Incremental Learning, Competitive Dynamics and Symmetry","authors":["Mo Zhou","Weihang Xu","Simon Du","Maryam Fazel"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Fast-forward scaling multiplies the data force but not the fixed decay","plainExplanation":"Fast-forward scaling multiplies the data force but not the fixed decay. The proof vector field differs from the algorithm; the eventual learning guarantee is not independently refuted by this identity check.","primaryUrl":"https://arxiv.org/html/2609.10879","noteUrl":"/review?paper=review-neurips-2026-153694","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153696","title":"Learning What to Remember: Test-Time Training via Context Distillation","authors":["Zixuan Wang","Xingyu Dang","Rui-Jie Zhu","Zixin Wen","Hengyu Fu","Wenhao Chai","Jason Lee"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning What to Remember: Test-Time Training via Context Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.01672","noteUrl":"/review?paper=review-neurips-2026-153696","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153698","title":"Epiplexity Guided Data Selection and Generation for Out-of-Distribution Generalization","authors":["Ellen Su","Andres Potapczynski","Shikai Qiu","Edward Hughes","Andrew Wilson"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A global-step objective does not yield the exposure-weighted selector derivative","plainExplanation":"A global-step objective does not yield the exposure-weighted selector derivative. The displayed objective-to-selector derivation is inconsistent; empirical heuristic utility is not disproved.","primaryUrl":"https://arxiv.org/html/2608.11746","noteUrl":"/review?paper=review-neurips-2026-153698","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153702","title":"MaPP: A Unified Marginalized Posterior-Predictive Framework for Data-Efficient RLVR","authors":["YangYang Ren","Haodong Zhu","Sheng Xu","Yanjing Li","Nikolai Y Zolotykh","Wentao Zhang","Baochang Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Frequentist estimator variance is not the posterior variance of the intrinsic target","plainExplanation":"An unconditional score moment cannot be factored from advantage error. The stated second-moment implication is false even for iid responses and genuine scores. Stronger bounds can repair it, but comparing upper bounds alone does not order actual MaPP and GRPO gradient risks.","primaryUrl":"https://arxiv.org/html/2609.34990","noteUrl":"/review?paper=review-neurips-2026-153702","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153703","title":"Human-AI Teaming Through the Lens of Calibration","authors":["Eric Nalisnick","Chi Zhang","Chengxin Qian","Yixin Wang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Human-AI Teaming Through the Lens of Calibration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.10906","noteUrl":"/review?paper=review-neurips-2026-153703","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153704","title":"Neural Backward Filtering Forward Guiding","authors":["Gefan Yang","Frank van der Meulen","Stefan Sommer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neural Backward Filtering Forward Guiding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.23030","noteUrl":"/review?paper=review-neurips-2026-153704","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153705","title":"PREPING: Building Agent Memory without Tasks","authors":["Yumin Choi","Sangwoo Park","Minki Kang","Jinheon Baek","Sung Ju Hwang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PREPING: Building Agent Memory without Tasks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13880","noteUrl":"/review?paper=review-neurips-2026-153705","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153707","title":"Your Neighbors Know: Leveraging Local Neighborhoods for Backdoor Detection in Decentralized Learning","authors":["Sayan Biswas","Antoine Boutet","Davide Frey","Romaric Gaudel","Rachid Guerraoui","Maxime Jacovella","Anne-marie Kermarrec","Dimitri Lerévérend","Francois Taiani","Martijn De Vos"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The mean-drift bound contains an extra factor of one over the node count","plainExplanation":"The convergence rate loses ordinary stochastic-gradient noise when mixing imbalance vanishes. The mean-drift lemma also has an extra factor of the network size, and its mask moments require spatial independence beyond independent rounds. The repairs retain the variance contribution, correct the drift normalization and state the independence needed for the moment calculation.","primaryUrl":"https://arxiv.org/html/2605.19969","noteUrl":"/review?paper=review-neurips-2026-153707","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153708","title":"Baton: Explicit Semantic Blueprints for Joint Video-Audio Generation","authors":["Shuyuan Tu","Qi Tian","Zihan Yang","Yue Wu","Xintong Han","Weijie Kong","Jiangfeng Xiong","Jian-Wei Zhang","Zhao Zhong","Liefeng Bo","Zuxuan Wu","Yu-Gang Jiang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Baton: Explicit Semantic Blueprints for Joint Video-Audio Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25195","noteUrl":"/review?paper=review-neurips-2026-153708","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153709","title":"Exact Instance Compression for Convex Empirical Risk Minimization via Color Refinement","authors":["Bryan Zhu","Ziang Chen"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Exact Instance Compression for Convex Empirical Risk Minimization via Color Refinement","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.00437","noteUrl":"/review?paper=review-neurips-2026-153709","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153710","title":"Unified Approach for Weakly Supervised Multicalibration","authors":["Futoshi Futami","Takashi Ishida"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Singleton bin witnesses give a maximum rather than summed ECE","plainExplanation":"Sconf now uses pair-label agreement, resolving that finding. The finite bin-indicator witness class still takes a maximum over bin discrepancies rather than their sum. The existing temperature-accuracy reporting exclusion is preserved.","primaryUrl":"https://arxiv.org/html/2605.09857","noteUrl":"/review?paper=review-neurips-2026-153710","collection":"neurips-2026","severity":"major","issueTypes":["definition_error","false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153711","title":"Diagnosing and Mitigating Modality Interference in Multimodal Large Language Models","authors":["Rui Cai","Bangzheng Li","Xiaofei Wen","Muhao Chen","Zhe Zhao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Diagnosing and Mitigating Modality Interference in Multimodal Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2505.19616","noteUrl":"/review?paper=review-neurips-2026-153711","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153713","title":"Inference Time Nash Alignment","authors":["Hadi Hosseini","Debmalya Mandal","Duohan Zhang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Inference Time Nash Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.08082","noteUrl":"/review?paper=review-neurips-2026-153713","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153714","title":"RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution","authors":["Siyonng Jian","Siyuan Li","Luyuan Zhang","Zedong Wang","Xin Jin","Ying Li","Cheng Tan","Huan Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution","plainExplanation":"Normalizing by a prior average does not keep each Bernoulli probability below one. Averaging the printed likelihood under that prior also gives one, rather than its claimed evidence.","primaryUrl":"https://arxiv.org/html/2605.21195","noteUrl":"/review?paper=review-neurips-2026-153714","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","mathematical_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153715","title":"Uniform Diffusion Models revisited: Leave-One-Out Denoiser and Absorbing State Reformulation","authors":["Samson Gourevitch","Yazid Janati","Dario Shariatian","Umut Simsekli","Eric Moulines","Eric Xing","Alain Durmus"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Uniform Diffusion Models revisited: Leave-One-Out Denoiser and Absorbing State Reformulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.22765v1","noteUrl":"/review?paper=review-neurips-2026-153715","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153718","title":"Archimedean Copula Inference via Taylor-Mode AD","authors":["Cambridge Yang","Dongdong Li"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"A fully censored internal subtree needs the identity derivative polynomial","plainExplanation":"A fully censored internal subtree needs the identity derivative polynomial. The printed arbitrary-censoring routine omits this case; no cached-code behavior is inferred.","primaryUrl":"https://arxiv.org/html/2605.23134","noteUrl":"/review?paper=review-neurips-2026-153718","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153720","title":"Runtime Monitoring of Perception-Based Autonomous Systems via Embedding Temporal Logic","authors":["Parv Kapoor","Abigail Hammer","Ashish Kapoor","Karen Leung","Eunsuk Kang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Small error levels request a nonexistent calibration order statistic","plainExplanation":"At a closed atomic boundary \\(\\delta=\\varepsilon\\), the atom is Boolean true and its negation false, but both quantitative robustness values are zero. Assigning zero to true accepts both; assigning zero to false rejects the true atom. Strict atom comparisons merely move the same conflict to negation. The equivalence fails at zero margin. Off-boundary sign semantics and empirical monitoring performance are not refuted.","primaryUrl":"https://arxiv.org/html/2605.12651","noteUrl":"/review?paper=review-neurips-2026-153720","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153722","title":"Learning from the Self-future: On-policy Self-distillation for dLLMs","authors":["Yifu Luo","Zeyu Chen","Haoyu Wang","Xinhao Hu","Yuxuan Zhang","Zhizhou Sha","Shiwei Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Shared parameters do not make input gradients constant or remove activation graphs","plainExplanation":"For shared scalar weight \\(w\\), input \\(x_i\\), and squared error \\((wx_i-y_i)^2\\), the per-input gradient is \\(2(wx_i-y_i)x_i\\). At \\(w=0,y_i=1\\), inputs one and two give \\(-2\\) and \\(-4\\). Shared weights make gradients accumulate into one buffer; they do not equate them. Concatenating \\(B\\) examples over \\(T\\) steps still gives \\(BT\\) rows with input-dependent intermediates for generic reverse-mode differentiation or recomputation. The mathematical explanation is false. Real engineering savings may exist, and neither equation (12) nor benchmark accuracy is refuted.","primaryUrl":"https://arxiv.org/html/2606.18195","noteUrl":"/review?paper=review-neurips-2026-153722","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153723","title":"Universal Approximation Theorems for Dynamical Systems with Infinite-Time Horizon Guarantees","authors":["Ábel Ságodi","Memming Park"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Smooth approximation alone does not supply exact localized period correction","plainExplanation":"The arbitrary $C^1$-dense-class assumption does not ensure exact localized period corrections needed for synchronized infinite-horizon tracking. The attractor proof also uses the exact spectral-gap decay rate despite allowed Jordan-block growth; a smaller rate repairs that local estimate.","primaryUrl":"https://arxiv.org/html/2602.08640","noteUrl":"/review?paper=review-neurips-2026-153723","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153727","title":"EvoGround: Self-Evolving Video Agents for Video Temporal Grounding","authors":["MIN JOON JUNG","Byoung-Tak Zhang","Lorenzo Torresani"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EvoGround: Self-Evolving Video Agents for Video Temporal Grounding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13803","noteUrl":"/review?paper=review-neurips-2026-153727","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153731","title":"Sparse Planning in Visual World Models via Cost Gradients","authors":["Yingchen Xu","Edward Grefenstette"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Sparse Planning in Visual World Models via Cost Gradients","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10274","noteUrl":"/review?paper=review-neurips-2026-153731","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153734","title":"The Hidden Power of Scaling Factor in LoRA Optimization","authors":["Zicheng Zhang","Haoran Li","Jiaxing Wang","Guoqiang Gong","Anqi Li","Yudong Hu","Ting Xiong","Yurong Gao","Junxing Hu","Zhida Jiang","Yifeng Zhang","Pengzhang Liu","Qixia Jiang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The drift Hessian needs a nonzero coupling to be indefinite","plainExplanation":"The drift Hessian is indefinite only when its off-diagonal coupling is nonzero. The expected trace identity also does not imply top-eigenvalue suppression, and the stability lemma confuses global smoothness sufficiency with a local-Hessian necessary-and-sufficient condition.","primaryUrl":"https://arxiv.org/html/2606.12883","noteUrl":"/review?paper=review-neurips-2026-153734","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153742","title":"Embodied Neurocomputation: A Framework for Interfacing Biological Neural Cultures with Scaled Task-Driven Validation","authors":["Johnson Zhou","Daniel Tanneberg","Forough Habibollahi","Alon Loeffler","Kiaran Lawson","Valentina Baccetti","Kwaku D Abu-Bonsrah","Candice Desouza","Finn Doensen","Bradley Watmuff","Daria Kornienko","Azin Azadi","Justin L Bourke","Bernhard Sendhoff","Brett J. Kagan"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Embodied Neurocomputation: A Framework for Interfacing Biological Neural Cultures with Scaled Task-Driven Validation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13315","noteUrl":"/review?paper=review-neurips-2026-153742","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153746","title":"Best Arm Identification for Bandits with Shifting Means","authors":["Lukas Zierahn","Wouter Koolen","Shubhada Agrawal","Christina Katsimerou","Dirk van der Hoeven"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Best Arm Identification for Bandits with Shifting Means","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10488","noteUrl":"/review?paper=review-neurips-2026-153746","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153747","title":"Dynamic Convolutions Improve Transformers","authors":["Oliver Sieberling","Bharat Runwal","Rameswar Panda","Yoon Kim"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Dynamic Convolutions Improve Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03825v1","noteUrl":"/review?paper=review-neurips-2026-153747","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153750","title":"E4GEN: Event-level Explainable Extreme-Enhanced Time-series Generation","authors":["Lin Jiang","Dahai Yu","Ximiao Li","Guang Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Bounded TV errors do not imply small alignment error","plainExplanation":"The inequality gives \\(\\operatorname{TV}(P_d,P_n)\\leq\\operatorname{TV}(P_d,q)+\\operatorname{TV}(P_n,q)\\). Each term is automatically bounded by one; boundedness alone gives no small value. A common short identity transition preserves an initial near-unit mismatch. Finite reconstruction moments also do not imply small TV, as arbitrarily close distinct point masses show. The triangle inequality and a conditional small-error statement remain valid. The unconditional alignment justification is missing, not every empirical alignment effect.","primaryUrl":"https://arxiv.org/html/2606.01634","noteUrl":"/review?paper=review-neurips-2026-153750","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153752","title":"Multidimensional Observer Model and Perceptual Dimensions of Human Image Quality Assessment","authors":["Sheng Zhao","Weikai Lin","Yuhao Zhu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Multidimensional Observer Model and Perceptual Dimensions of Human Image Quality Assessment","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.38487","noteUrl":"/review?paper=review-neurips-2026-153752","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153754","title":"The risk of KV cache compression","authors":["Lukas Haverbeck","Carmen Amo Alonso","Andres Posada Moreno","Sebastian Trimpe","Marco Pavone"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Resetting the level before storage destroys the balanced carry schedule","plainExplanation":"Resetting the level before storage destroys the balanced carry schedule. The printed schedule contradicts its bucket invariant; this is not evidence about an external implementation.","primaryUrl":"https://arxiv.org/html/2607.01520","noteUrl":"/review?paper=review-neurips-2026-153754","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153755","title":"LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling","authors":["Biswajit Banerjee","Claudia Alvarez Carreno","Anton S Petrov"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Maximum trie dropout can still select a multi-residue token","plainExplanation":"Maximum trie dropout can still select a multi-residue token.","primaryUrl":"https://arxiv.org/html/2609.37675","noteUrl":"/review?paper=review-neurips-2026-153755","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153763","title":"When Think-with-Image Meets Safety: What Determines Multimodal Jailbreak Robustness?","authors":["Yuan Tian","Bing Hu","Fang Wu","Xiaomin Li","binghang lu","Neil Gong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"When Think-with-Image Meets Safety: What Determines Multimodal Jailbreak Robustness?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27932","noteUrl":"/review?paper=review-neurips-2026-153763","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153765","title":"Goal-Conditioned Supervised Learning for Multi-Objective Recommendation","authors":["Shijun Li","Hilaf Hasson","Jing Hu","Joydeep Ghosh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Return-conditioned logs are not fresh rollout labels under stochastic dynamics","plainExplanation":"A one-action MDP can transition equally to states with deterministic rewards \\((1,1)\\) or \\((0,0)\\). Exact imitation has no action choice to change and cannot force future randomness to match the logged return. A model receiving that same logged return as input and target can achieve identity reconstruction, although fresh rollout means remain \\((1/2,1/2)\\). Deterministic rewards and perfect action imitation do not remove stochastic outcome-selection bias. The separate theorem and empirical recommendation utility are not refuted.","primaryUrl":"https://arxiv.org/html/2412.08911","noteUrl":"/review?paper=review-neurips-2026-153765","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153767","title":"Reformulating KV Cache Eviction Problem for Long-Context LLM Inference","authors":["Tho Mai","Joo-Young Kim"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Reformulating KV Cache Eviction Problem for Long-Context LLM Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07234","noteUrl":"/review?paper=review-neurips-2026-153767","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153768","title":"Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting","authors":["Haoxin Liu","Yichen Zhou","Rajat Sen","B. Aditya Prakash","Abhimanyu Das"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.29401","noteUrl":"/review?paper=review-neurips-2026-153768","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153769","title":"AdaPaD: Adaptive Parallel Deflation for PEFT with Self-Correcting Rank Discovery","authors":["Barbara Su","Fangshuo Liao","Anastasios Kyrillidis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A vanishing algorithmic term cannot leave a dimension-uniform Gaussian noise floor","plainExplanation":"A vanishing algorithmic term cannot leave a dimension-uniform Gaussian noise floor. Post-update projection does not supply the required initialization bound. The simplified warm-up bound omits a logarithmic divergence near unit contraction. Factor scaling rules out strict Nash uniqueness modulo signs. Clamping the prescribed spectrum invalidates the stated gap formula.","primaryUrl":"https://arxiv.org/html/2605.10741","noteUrl":"/review?paper=review-neurips-2026-153769","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","proof_gap","reporting_inconsistency"],"classifications":["major"]},{"slug":"review-neurips-2026-153771","title":"Flow Matching for Offline Reinforcement Learning with Discrete Actions","authors":["Fairoz Nower Khan","Nabuat Zaman Nahim","Ruiquan Huang","Haibo Yang","Peizhong Ju"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The policy-improvement justification needs a matching proposal, regularized value and finite time objective","plainExplanation":"Endpoint tilting retains its proposal as the reference measure. Current-policy proposals therefore do not automatically recover the fixed behavior-referenced target in the corollary.","primaryUrl":"https://arxiv.org/html/2602.06138","noteUrl":"/review?paper=review-neurips-2026-153771","collection":"neurips-2026","severity":"repairable","issueTypes":["mathematical_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153772","title":"Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse","authors":["Taekyung Heo","Rasoul Shafipour","Ritchie Zhao","Maximilian Golub","Mohammad Mahdi Kamani","Ritika Borkar","Makesh T Chandran","Pantea Zardoshti","Bita Darvish Rouhani"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.03893","noteUrl":"/review?paper=review-neurips-2026-153772","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153774","title":"CSO-LLM: Class Subspace Orthogonalization for Post-Training Backdoor Detection and Trigger Inversion in LLMs","authors":["Zhengxing Li","David J Miller","Guangmingmei Yang","George Kesidis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A signed objective cannot use its raw old value as a relative-improvement denominator","plainExplanation":"A signed objective cannot use its raw old value as a relative-improvement denominator. The printed stopping rule can reverse improvement and deterioration; actual retained tokens are not determined by it.","primaryUrl":"https://arxiv.org/html/2606.31309","noteUrl":"/review?paper=review-neurips-2026-153774","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153776","title":"M$^2$E-UAV: A Benchmark and Analysis for Onboard Motion-on-Motion Event-Based Tiny UAV Detection","authors":["Weiqi Yan","Lixin Chen","Xiangrui Hou","zhipeng cai","Youbiao Wang","Yangyang Shi","Yu Zang","Cheng Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"M$^2$E-UAV: A Benchmark and Analysis for Onboard Motion-on-Motion Event-Based Tiny UAV Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10496","noteUrl":"/review?paper=review-neurips-2026-153776","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153778","title":"Take It or Leave It: Intent-Controlled Partial Optimal Transport","authors":["Parth TRIPATHI","bertrand chapron","Fabrice collard","Nicolas Courty","ronan fablet"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Take It or Leave It: Intent-Controlled Partial Optimal Transport","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20030","noteUrl":"/review?paper=review-neurips-2026-153778","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153781","title":"Jointly Reinforcing Diversity and Quality in Language Model Generations","authors":["Tianjian Li","Yiming Zhang","Ping Yu","Swarnadeep Saha","Daniel Khashabi","Jason Weston","Jack Lanchantin","Tianlu Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The equivalence classifier has the opposite orientation from diversity","plainExplanation":"The equivalence classifier has the opposite orientation from diversity. The formula and example disagree; this does not establish that the actual trained reward promotes repetition.","primaryUrl":"https://arxiv.org/html/2509.02534","noteUrl":"/review?paper=review-neurips-2026-153781","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153785","title":"Dynamic Latent Routing","authors":["Fangyuan Yu","Xin Su","Amir Abdullah"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Dynamic Latent Routing","plainExplanation":"The suffix return must use shifted environment rewards. A reset policy index alone does not justify resetting the reward clock.","primaryUrl":"https://arxiv.org/html/2605.14323","noteUrl":"/review?paper=review-neurips-2026-153785","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","mathematical_error","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-153786","title":"ManifoldCache: Training-Free Diffusion Acceleration via Constraint Manifold Caching","authors":["Prashant Pandey","Sri Venkatraya Chowdary Devineni","Brejesh Lall"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The cacheability proof omits the state drift of variance-preserving diffusion","plainExplanation":"The probability-flow argument omits the linear drift of the stated variance-preserving kernel, leaving its cache-stride bound unproved. The frozen-posterior estimation theorem remains valid, but it does not by itself bound arbitrary block reuse: such a guarantee needs explicit feature-drift and downstream-sensitivity control. The paper already limits its Gaussian illustration, and the findings do not assess the implemented sampler or experiments.","primaryUrl":"https://arxiv.org/html/2610.04510","noteUrl":"/review?paper=review-neurips-2026-153786","collection":"neurips-2026","severity":"major","issueTypes":["missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153787","title":"FIVE-VLA: Fast and EffectIVE Closed-Loop Autonomous Driving with Recurrent Action Memory","authors":["Kemal Oksuz","Alexandru Buburuzan","Yuhan Yao","Puneet Dokania"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"FIVE-VLA: Fast and EffectIVE Closed-Loop Autonomous Driving with Recurrent Action Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.18623v1","noteUrl":"/review?paper=review-neurips-2026-153787","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153789","title":"Posterior Inference in Latent Space for Scalable Constrained Black-box Optimization","authors":["Kiyoung Om","Kyuil Sim","Taeyoung Yun","Hyeongyu Kang","Jinkyoo Park"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Posterior Inference in Latent Space for Scalable Constrained Black-box Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2507.00480","noteUrl":"/review?paper=review-neurips-2026-153789","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153791","title":"Two Speeds of Learning: A Representation-Readout Decomposition of Grokking and Double Descent","authors":["Chi-Ning Chou","Oscar Uzdelewicz","Neng-Chun Chiu","Yao-Yuan Yang","SueYeon Chung"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Squaring before maximization changes the cone-projection value","plainExplanation":"A one-sided cone requires a positive-part or zero choice. Squaring each directional score first rewards both positive and negative inner products, so the displayed expression is not the constrained-distance value. The displayed strong-duality rewrite fails for an admissible, strictly separable pair of manifolds. The primal quadratic program is retained. This calculation does not determine which expression the implementation uses or invalidate the paper’s other geometric claims.","primaryUrl":"https://arxiv.org/html/2605.27078","noteUrl":"/review?paper=review-neurips-2026-153791","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153792","title":"AOT-POT: Adaptive Operator Transformation for Large-Scale PDE Pre-training","authors":["Qitan Lv","Hong Wang","Hao Zhongkai","Wen Wu","Xuenan Xu","Bowen Zhou","Feng Wu","Chao Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Row and column normalization does not make an unconstrained affine matrix nonnegative","plainExplanation":"Row and column normalization does not make an unconstrained affine matrix nonnegative. The printed construction needs its positivity step. An implementation treating Sinkhorn as including softmax could already repair the documentation.","primaryUrl":"https://arxiv.org/html/2605.15793","noteUrl":"/review?paper=review-neurips-2026-153792","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153796","title":"Training Data Attribution in Diffusion Models via Mirrored Unlearning and Noise-Consistent Skew","authors":["Joan Serrà","Dipam Goswami","Fabio Morreale","Wei-Hsiang Liao","Yuki Mitsufuji"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Training Data Attribution in Diffusion Models via Mirrored Unlearning and Noise-Consistent Skew","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.17938","noteUrl":"/review?paper=review-neurips-2026-153796","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153797","title":"Allocentric Perceiver: Disentangling Allocentric Reasoning from Egocentric Visual Priors via Frame Instantiation","authors":["Hengyi Wang","RuiQiang Zhang","Chang Liu","Guanjie Wang","Zehua Ma","Han Fang","Weiming Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"One cross product does not orthogonalize both input axes","plainExplanation":"One cross product does not orthogonalize both input axes. The stated columns need not form a rigid rotation; the finding does not establish the implementation’s fallback choices.","primaryUrl":"https://arxiv.org/html/2602.05789","noteUrl":"/review?paper=review-neurips-2026-153797","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153798","title":"Training-Induced Escape from Token Clustering in a Mean-Field Formulation of Transformers","authors":["Noboru Isobe","Daisuke Inoue","Masaaki Imaizumi"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Training-Induced Escape from Token Clustering in a Mean-Field Formulation of Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.07772","noteUrl":"/review?paper=review-neurips-2026-153798","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153802","title":"CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs","authors":["Han Guo","Jack Zhang","Arjun Menon","Driss Guessous","Vijay Thakkar","Yoon Kim","Tri Dao"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19269","noteUrl":"/review?paper=review-neurips-2026-153802","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153803","title":"Conditional Optimal Bridge for Riemannian Activation Steering","authors":["Seyed Arshan Dalili","Ajay Narayanan Sridhar","Vijaykrishnan Narayanan","Mehrdad Mahdavi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Conditional Optimal Bridge for Riemannian Activation Steering","plainExplanation":"No issue was found in the checked claims. The revised limit has the correct direction: increasing the denominator makes fixed-distance costs vanish. The old small-diffusivity example has been removed.","primaryUrl":"https://arxiv.org/html/2607.10517","noteUrl":"/review?paper=review-neurips-2026-153803","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153805","title":"A Mechanistic Analysis of Looped Reasoning Language Models","authors":["Hugh Blayney","Alvaro Arroyo","Johan Obando Ceron","Pablo Samuel Castro","Aaron Courville","Michael Bronstein","Xiaowen Dong"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"A Mechanistic Analysis of Looped Reasoning Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.11791","noteUrl":"/review?paper=review-neurips-2026-153805","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153806","title":"Single Shot HDR Recovery via a Video Diffusion Prior","authors":["Chinmay Talegaonkar","Jinshi He","Nicholas Antipa"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Single Shot HDR Recovery via a Video Diffusion Prior","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11628","noteUrl":"/review?paper=review-neurips-2026-153806","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153807","title":"Optimal Recalibration of an Online Predictor","authors":["Lunjia Hu","Kevin Tian","Chutong Yang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Optimal Recalibration of an Online Predictor","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.19689","noteUrl":"/review?paper=review-neurips-2026-153807","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153809","title":"Geometry-Constrained Kolmogorov–Arnold Networks: Learning Edge Geometry via Banach Duality","authors":["Senanayak Sesh Kumar Karri"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The smoothed power map does not have the displayed global Lipschitz constant","plainExplanation":"The smoothed power map does not have the displayed global Lipschitz constant. The local certificate is incorrect; empirical network performance is not independently refuted.","primaryUrl":"https://arxiv.org/html/2608.25807","noteUrl":"/review?paper=review-neurips-2026-153809","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153810","title":"Does Seeing More Mean Knowing More? Mono-Anchored Advantage Normalization for Multi-Source Visual Reasoning","authors":["Fanhu Zeng","Zhicong Luo","Zefan Wang","Li You","Chi Chen","Maosong Sun"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Hybrid normalization does not preserve the claimed gradient identity","plainExplanation":"Using each rollout in its own empirical baseline creates finite-sample bias, and the random standard deviation cannot be discarded. The normalized decomposition also needs two batch scale factors. Individual positive advantages are determined by the hybrid mean; only the average multi-source advantage directly tracks the mono–multi mean reward gap.","primaryUrl":"https://arxiv.org/html/2605.25437v2","noteUrl":"/review?paper=review-neurips-2026-153810","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","invalid_inference","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-153811","title":"Attention-Based Pretraining for Unsupervised Amortized Causal Discovery","authors":["Naiyu Yin","Tian Gao","Yue Yu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The operator orientation does not match the transpose-SEM loss","plainExplanation":"The operator orientation does not match the transpose-SEM loss.","primaryUrl":"https://arxiv.org/html/2506.18285v1","noteUrl":"/review?paper=review-neurips-2026-153811","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-153813","title":"EcoGEO: Trajectory-Aware Evidence Ecosystems for Web-Enabled LLM Search Agents","authors":["Hengwei Ye","Jiasheng Mao","Zhenhan Guan","Zheng Tian"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EcoGEO: Trajectory-Aware Evidence Ecosystems for Web-Enabled LLM Search Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12887","noteUrl":"/review?paper=review-neurips-2026-153813","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153817","title":"SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers","authors":["Javad Rajabi","Kimia Shaban","Koorosh Roohi","David Lindell","Babak Taati"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Centering a nonlinear energy transform can reverse its pointwise sign","plainExplanation":"Centering a nonlinear energy transform can reverse its pointwise sign. Three power-law substitutions lie outside their rounding intervals.","primaryUrl":"https://arxiv.org/html/2605.22668","noteUrl":"/review?paper=review-neurips-2026-153817","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-153818","title":"pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows","authors":["Tong Chen","Maximilian Holsman","Lin Zhao","Yinuo Zhang","Pranam Chatterjee"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01663","noteUrl":"/review?paper=review-neurips-2026-153818","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153820","title":"Deep Double Q-learning","authors":["Prabhat Nagarajan","Martha White","Marlos C. Machado"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Keeping only completed episodes can bias a finite-budget return estimate","plainExplanation":"Keeping only completed episodes can bias a finite-budget return estimate.","primaryUrl":"https://arxiv.org/html/2507.00275","noteUrl":"/review?paper=review-neurips-2026-153820","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153823","title":"When Do Multi-Agent Systems Help? An Information Bottleneck Perspective","authors":["Wendi Yu","Lianhao Zhou","Xiangjue Dong","Sai S Barath","Declan Staunton","Byung-Jun Yoon","Xiaoning Qian","James Caverlee","Shuiwang Ji"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"When Do Multi-Agent Systems Help? An Information Bottleneck Perspective","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.16133","noteUrl":"/review?paper=review-neurips-2026-153823","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153826","title":"Measuring Weak-to-Strong Legibility of Reasoning Models","authors":["Dani Roytburg","Shreya Sridhar","Daphne Ippolito"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Measuring Weak-to-Strong Legibility of Reasoning Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.20508","noteUrl":"/review?paper=review-neurips-2026-153826","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153828","title":"Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time","authors":["Jennifer Wendland","Nicolas Freitag","Maik Kschischo"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The printed treatment kernel can be singular and exceed its claimed bound","plainExplanation":"With one treatment coordinate, \\(w=2,t=3\\) includes index \\(k=2\\), making \\((w-k)^2=0\\) while the numerator can be positive. Even away from that pole, at \\(t=5\\) and treatment probability \\(0.9\\), the allowed terms give \\(0.9(1+1/4+1/9)>1\\) with \\(\\beta=1\\). Reinterpreting the denominator as a positive lag removes one pole but does not normalize the sum. The printed domain and range guarantee fail. The actual simulator implementation and generated datasets are not checked.","primaryUrl":"https://arxiv.org/html/2604.26070","noteUrl":"/review?paper=review-neurips-2026-153828","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153831","title":"Instruct-Particulate: Scaling Feed-Forward 3D Object Articulation with Kinematic Control","authors":["Ruining Li","Yuxin Yao","Matt Zhou","Chuanxia Zheng","Christian Rupprecht","Joan Lasenby","Shangzhe Wu","Andrea Vedaldi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A mean of unit direction votes need not lie on the unit sphere","plainExplanation":"A mean of unit direction votes need not lie on the unit sphere. The axis-domain specification is incomplete; internal forward-kinematics normalization would need to be stated and still needs a zero case.","primaryUrl":"https://arxiv.org/html/2606.14699","noteUrl":"/review?paper=review-neurips-2026-153831","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153838","title":"Strong Teacher Not Needed? On Distillation in LLM Pretraining","authors":["Taiming Lu","Zhuang Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strong Teacher Not Needed? On Distillation in LLM Pretraining","plainExplanation":"The entropy gap from uniform does not determine the true-token effect of label smoothing. The finding is limited to the asserted analytical benefit profile.","primaryUrl":"https://arxiv.org/pdf/2605.23857","noteUrl":"/review?paper=review-neurips-2026-153838","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153839","title":"Calibrating Scientific Foundation Models with Inference-Time Stochastic Attention","authors":["Akash Yadav","Taiwo Adebiyi","Ruda Zhang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Calibrating Scientific Foundation Models with Inference-Time Stochastic Attention","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.19530","noteUrl":"/review?paper=review-neurips-2026-153839","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153840","title":"Inverting the Bellman Equation: From $Q$-Values to World Models","authors":["Alistair Letcher","Mattie Fellows","Alexander D. Goldie","Jonathan Richens","Jakob Foerster","Oliver Richardson"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The recovery threshold places the noise-amplification factor in the numerator","plainExplanation":"Nearest columns in one norm need not minimize residuals in another. The appendix minimizer/uniqueness claim is false as stated; exact-match identification is a different valid property.","primaryUrl":"https://arxiv.org/pdf/2606.21173v1","noteUrl":"/review?paper=review-neurips-2026-153840","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153841","title":"See it to Place it: Evolving Macro Placements with Vision Language Models","authors":["Ikechukwu Uchendu","Swati Goel","Karly Hou","Ebrahim Songhori","Kuang-Huei Lee","Wenjie (Joe) Jiang","Vijay Janapa Reddi","Vincent Zhuang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"See it to Place it: Evolving Macro Placements with Vision Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.28733","noteUrl":"/review?paper=review-neurips-2026-153841","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153844","title":"CoRDS: Coreset-Based Representative and Diverse Selection for Streaming Video Understanding","authors":["Ailar Mahdizadeh","Puria Azadi Moghadam","Muchen Li","Xiangteng He","Leonid Sigal"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The joint-space representative need not minimize value-space distortion","plainExplanation":"A minimum of a sum is not the sum of separate minima. Orthogonal residual ranking is not exact greedy ranking for the regularized log determinant. The seeded selection loop returns one token beyond its budget. The joint-space representative need not minimize value-space distortion.","primaryUrl":"https://arxiv.org/html/2605.14310","noteUrl":"/review?paper=review-neurips-2026-153844","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153845","title":"Testable Learning of General Halfspaces under Massart Noise","authors":["Ilias Diakonikolas","Giannis Iakovidis","Daniel Kane","Sihan Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The slice width is too large for the claimed threshold-variation bound","plainExplanation":"The slice width is too large for the claimed threshold-variation bound. The parameter-to-soundness deduction is incomplete; the learner’s existence and qualitative rate are not disproved.","primaryUrl":"https://arxiv.org/html/2602.22300","noteUrl":"/review?paper=review-neurips-2026-153845","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153849","title":"Curriculum Multiple Shooting for Robust Training of Neural and Universal Differential Equations","authors":["Sebastian Persson","Giacomo Fabrini","Branwen Snelling","Fabian Fröhlich"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The final single window still optimizes its initial state","plainExplanation":"The continuity penalty disappears, but the first window’s initial state remains free. For \\(u^{\\prime}=\\theta\\), observations \\(u(1)=u(2)=1\\) fit exactly with free intercept one and \\(\\theta=0\\). With prescribed \\(u(0)=0\\), the squared-error optimum is \\(\\theta=3/5\\) with positive residual. Returning only the dynamics parameter does not make these optimization problems equal. The equality to known-initial-state single shooting fails as written. Applications intentionally estimating an initial state can still use a different legitimate formulation.","primaryUrl":"https://arxiv.org/html/2608.05777","noteUrl":"/review?paper=review-neurips-2026-153849","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153854","title":"Real2Sim in HOI: Toward Physically Plausible HOI Reconstruction from Monocular Videos","authors":["Yubo ZHAO","Yujin Chai","Yunao Dong","Chengfeng Zhao","Zijiao Zeng","Yuan Liu","Chi-Keung Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Frame-index second differences are not acceleration in physical seconds","plainExplanation":"Frame-index second differences are not acceleration in physical seconds. The formula and units disagree; a common conversion preserves method ordering and could already be present in evaluation code.","primaryUrl":"https://arxiv.org/html/2605.14462","noteUrl":"/review?paper=review-neurips-2026-153854","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153855","title":"Structured Masked Diffusion for Joint Multiuser Decoding","authors":["Taekyun Lee","Jiyoung Yun","Jeffrey Andrews","Hyeji Kim"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Structured Masked Diffusion for Joint Multiuser Decoding","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.26580","noteUrl":"/review?paper=review-neurips-2026-153855","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153856","title":"Causal Survival Forests with Negative Controls","authors":["Zijun Gao","Kyoungeui Hong","Leyi Ma","Qianli Wu","Zachary Izzo","Ruishan Liu"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Functional nuisance errors enter the score at first order","plainExplanation":"The score is not orthogonal to all permitted nuisance-function errors. A bounded example has a valid oracle score, yet nuisance error of order $n^{-1/3}$ shifts the root enough to invalidate the claimed centered limit. A separate minor inconsistency concerns two definitions of the centering function; the paper's algorithms already use the bridge mixture required by the oracle proof.","primaryUrl":"https://arxiv.org/html/2608.19749","noteUrl":"/review?paper=review-neurips-2026-153856","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","invalid_inference","notation"],"classifications":["critical"]},{"slug":"review-neurips-2026-153859","title":"Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models","authors":["Shutong Wu","Jiawei zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Full-union edge feasibility does not imply prefix-closed feasibility or optimal recovery","plainExplanation":"Taking the largest passing verification index does not certify every earlier prefix. Full-union edge feasibility does not imply prefix-closed feasibility or optimal recovery.","primaryUrl":"https://arxiv.org/html/2510.00294","noteUrl":"/review?paper=review-neurips-2026-153859","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153861","title":"Rethinking Semantic ID Construction for Generative Recommendation: SimHash with Parallel Decoding and Semantic Alignment","authors":["Yuqing Liu","Huiyuan Chen","Yibo Wang","Wooseong Yang","Philip S Yu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Rethinking Semantic ID Construction for Generative Recommendation: SimHash with Parallel Decoding and Semantic Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07402","noteUrl":"/review?paper=review-neurips-2026-153861","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153863","title":"Attribute-Efficient Learning of Sparse Halfspaces with Constant Malicious Noise Rate","authors":["Shiwei Zeng","Jie Shen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The auxiliary sample and margin hypotheses are weaker than its proof requires","plainExplanation":"The needed uniform quadratic-moment bound is established only with sample size of order \\(s^2\\log^5d\\,\\log^2(1/\\delta^{\\prime})\\); Algorithm 3 also draws a quadratic-in-\\(s\\) sample. The proof uses pancake width \\(\\tau=4(\\log(1/\\varepsilon)+1)/\\sqrt d\\), and its condition \\(\\tau\\leq\\gamma/2\\) requires coefficient eight, not four, in the margin lower bound. The stated auxiliary hypotheses are unsupported. Main Theorem 2 already uses the stronger regime and is unaffected.","primaryUrl":"https://arxiv.org/html/2505.21430","noteUrl":"/review?paper=review-neurips-2026-153863","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153864","title":"TOPPO: Rethinking PPO for Multi-Task Reinforcement Learning with Critic Balancing","authors":["Yuanpeng Li","Rui Miao","Gefei Lin","Annie Qu"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"TOPPO: Rethinking PPO for Multi-Task Reinforcement Learning with Critic Balancing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11473","noteUrl":"/review?paper=review-neurips-2026-153864","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153865","title":"ODDR: One-Step Deshadow Diffusion via Reward Guidance","authors":["Junseong Shin","Kijun Kim","Minseong Kim","Dongjin Kim","Tae Hyun Kim"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"ODDR: One-Step Deshadow Diffusion via Reward Guidance","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.01291","noteUrl":"/review?paper=review-neurips-2026-153865","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153866","title":"TRACE: Structure-Aware Character Encoding for Robust and Generalizable Document Watermarking","authors":["Jiale Meng","Jie Zhang","Runyi Hu","Zhe-Ming Lu","Tianwei Zhang","Yiming Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The target displacement is divided by the stroke length twice","plainExplanation":"The target move can miss the requested decoding threshold because of an extra stroke-length normalization or a reversed selected-axis sign. The inverse-cosine angle also loses the orientation needed by a later branch. A correct geometric target must satisfy the full minimum-gap decoding predicate and the separate handle constraints. The authentication claim needs a document-bound security argument; the mask boundary and reference-selection notation have smaller corrections supplied by nearby definitions and comments.","primaryUrl":"https://arxiv.org/html/2603.12873","noteUrl":"/review?paper=review-neurips-2026-153866","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153869","title":"When Does Online Imitation Learning Help in LLM Post-Training? The Role of (Non-)Realizability Beyond Horizon","authors":["Huaqing Zhang","Jingchu Gai","Juno Kim","Bingbin Liu","Andrej Risteski"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"When Does Online Imitation Learning Help in LLM Post-Training? The Role of (Non-)Realizability Beyond Horizon","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.30445","noteUrl":"/review?paper=review-neurips-2026-153869","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153870","title":"Lookalike3D: Seeing Double in 3D","authors":["Chandan Yeshwanth","Angela Dai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Normalized cosine scores can fall outside all classifier intervals","plainExplanation":"Normalized cosine scores can fall outside all classifier intervals. The printed score domain does not make the classifier exhaustive; no observed code failure is asserted.","primaryUrl":"https://arxiv.org/html/2603.24713","noteUrl":"/review?paper=review-neurips-2026-153870","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153874","title":"Asymptotically Log-Optimal Bayes-Assisted Confidence Sequences for Bounded Means","authors":["Valentin Kilian","Stefano Cortinovis","Francois Caron"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Asymptotically Log-Optimal Bayes-Assisted Confidence Sequences for Bounded Means","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07964","noteUrl":"/review?paper=review-neurips-2026-153874","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153875","title":"From Intent to Evidence: A Categorical Approach for Structural Evaluation of Deep Research Agents","authors":["Shuoling Liu","Zhiquan Tan","Kun Yi","Hui Wu","Yihan Li","Jiangpeng Yan","Liyuan Chen","Kai Chen","Qiang Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An empty cross-hom-set does not prove an object nonexistent","plainExplanation":"The free category on two isolated vertices \\(X,Y\\) has \\(\\operatorname{Hom}(X,Y)=\\operatorname{Hom}(Y,X)=\\varnothing\\), yet both objects exist with identity morphisms. In every category \\(\\operatorname{Hom}(X,X)\\) contains \\(\\mathrm{id}_X\\), so its representable functor is not empty. Yoneda determines objects up to isomorphism through their full representable functors; it supplies no nonexistence predicate from one empty cross-hom-set or a finite unsuccessful search. The formal categorical inference is invalid. The benchmark task and scores need not change, and failure to retrieve an entity is distinct from proving it does not exist.","primaryUrl":"https://arxiv.org/html/2603.25342","noteUrl":"/review?paper=review-neurips-2026-153875","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153878","title":"ReFree-S2V: Towards Realistic Co-Speech Video Generation via Reward-Free RL and Multilevel Speech Guidance","authors":["Salaheldin Mohamed","M. Hamza Mughal","Rishabh Dabral","Christian Theobalt"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Summing over speech levels removes the level axis","plainExplanation":"The algorithm retains an encoder-level axis that Eq. (1) has summed out. The pooled conditioning is a vector, rather than the declared level-by-feature array. This is a shape mismatch between the equation and algorithm input. It does not establish a runtime error in the actual speech-conditioning implementation.","primaryUrl":"https://arxiv.org/html/2606.13304","noteUrl":"/review?paper=review-neurips-2026-153878","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-153879","title":"Provable Robustness against Backdoor Attacks via the Primal-Dual Perspective on Differential Privacy","authors":["Aman Saxena","Jan Schuchardt","Yan Scholten","Stephan Günnemann"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Independent per-record Gaussian perturbations combine in joint Euclidean norm","plainExplanation":"Changing one weighted coordinate does not move the barycenter by the full displacement. Independent per-record Gaussian perturbations combine in joint Euclidean norm. The testing CDF formula needs a consistent privacy-loss sign and tie randomization.","primaryUrl":"https://arxiv.org/html/2605.21780","noteUrl":"/review?paper=review-neurips-2026-153879","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153880","title":"Robust Statistical Estimators with Bounded Empirical Sensitivity","authors":["Valentio Iverson","Gautam Kamath","Argyris Mouzakis","Adam Smith"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Robust Statistical Estimators with Bounded Empirical Sensitivity","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21860","noteUrl":"/review?paper=review-neurips-2026-153880","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153884","title":"TRACER: Verifiable Generative Provenance for Multimodal Tool-Using Agents","authors":["Bihui Yu","caijun jia","Jing Chi","Xiaohan Liu","Yining Wang","Richard H Bai","Yuchen Liu","Jingxuan Wei","Junnan Zhu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"TRACER: Verifiable Generative Provenance for Multimodal Tool-Using Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09934","noteUrl":"/review?paper=review-neurips-2026-153884","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153885","title":"VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation","authors":["Boyao Han","Chen Shi","Jingjing Qian","Zhuotao Tian","Li Jiang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.12451","noteUrl":"/review?paper=review-neurips-2026-153885","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153886","title":"PixelDense: Dense Prediction as Representation Alignment for Pixel Diffusion","authors":["Lehan Yang","Daiqing Qi","Wenhao Zhang","Avery Li","Yiqing Yang","Yifan Li","Yu Kong","Haitian Zheng","Zhifei Zhang","Zhe Lin","Varun Jampani","Sheng Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"PixelDense: Dense Prediction as Representation Alignment for Pixel Diffusion","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.00483","noteUrl":"/review?paper=review-neurips-2026-153886","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153889","title":"Mobility Helps Learning: Unsupervised Model Adaptation for Object Recognition via Movement","authors":["Yanan Ma","Yihang Tao","Zhengru Fang","Senkang Hu","Xianhao Chen","Yuguang Fang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The MSE criterion drops the density-ratio factor in the actual weight","plainExplanation":"Gaussian view marginals do not determine cross-view covariance. The MSE criterion drops the density-ratio factor in the actual weight.","primaryUrl":"https://arxiv.org/html/2610.15846","noteUrl":"/review?paper=review-neurips-2026-153889","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153892","title":"Efficient Analytic Uncertainty Quantification for Multimodal Regression","authors":["Kun Jin","James Harrison","Jiawei Li","Sihan Liu","Jiayi Liu","Randy Linderman","Yuening Li","Arnab Bhadury","Sourabh P Bansod","Liang Liu","Jasper Snoek"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The delta-method decomposition can produce negative aleatoric variance","plainExplanation":"The delta-method decomposition can produce negative aleatoric variance. Histogram consistency conditions do not ensure the optimal histogram rate. Vanishing histogram error does not control KL without lower-tail conditions.","primaryUrl":"https://arxiv.org/html/2606.25188v1","noteUrl":"/review?paper=review-neurips-2026-153892","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153893","title":"SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning","authors":["Philip Schroeder","Thomas Weng","Karl Schmeckpeper","Eric Rosen","Stephen Hart","Ondrej Biza"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Trajectory-specific extrema do not create comparable completion labels","plainExplanation":"Trajectory-specific extrema do not create comparable completion labels.","primaryUrl":"https://arxiv.org/html/2603.28730","noteUrl":"/review?paper=review-neurips-2026-153893","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153895","title":"Learning Causal Orderings for In-Context Tabular Prediction","authors":["Sascha Xu","Sarah Mameche","Jilles Vreeken"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The negative attention penalty is applied to allowed rather than blocked edges","plainExplanation":"The negative attention penalty is applied to allowed rather than blocked edges. The printed conventions contradict one another; implementation attention is not determined by them.","primaryUrl":"https://arxiv.org/pdf/2605.22335v1","noteUrl":"/review?paper=review-neurips-2026-153895","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153896","title":"SMoA: Spectrum Modulation Adapter for Parameter-Efficient Fine-Tuning","authors":["Yongkang Liu","Xing Li","Mengjie Zhao","Shanru Zhang","Zijing Wang","Qian Li","Shi Feng","Feiliang Ren","Daling Wang","Hinrich Schuetze"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The effective-rank bound need not strictly exceed the reference rank","plainExplanation":"The effective-rank bound need not strictly exceed the reference rank. The unconditional strictness is false; the following conditional/full-local-rank repair is already valid.","primaryUrl":"https://arxiv.org/html/2605.21147","noteUrl":"/review?paper=review-neurips-2026-153896","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153905","title":"Does Sparse Connectivity Improve Generalization? Convolutional Networks Below the Edge of Stability","authors":["Tongtong Liang","Esha Singh","Rahul Parhi","Alex Cloninger","Yu-Xiang Wang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"An unrestricted affine component makes the seminorm ball have infinite covering entropy","plainExplanation":"The displayed optimizer has incorrect powers of the problem constants. An unrestricted affine component makes the seminorm ball have infinite covering entropy. The two-dimensional sphere-tail exponent reverses the endpoint inequalities.","primaryUrl":"https://arxiv.org/html/2603.04807","noteUrl":"/review?paper=review-neurips-2026-153905","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153907","title":"Search-Augmented Masked Diffusion Models for Constrained Generation","authors":["Huu Binh Ta","Michael Cardei","Alvaro Velasquez","Ferdinando Fioretto"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Search-Augmented Masked Diffusion Models for Constrained Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.02727","noteUrl":"/review?paper=review-neurips-2026-153907","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153908","title":"Mental-Models for Multi-Agent Systems","authors":["Hanan Gani","Lulu Shao","Manmohan Chandraker"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The scorer serialization includes information outside the stated observable history","plainExplanation":"The scorer serialization includes information outside the stated observable history. The scorer interpretation and serialization disagree; a legitimate privileged critic does not imply deployment leakage.","primaryUrl":"https://arxiv.org/html/2610.12453","noteUrl":"/review?paper=review-neurips-2026-153908","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153911","title":"Horizon-Stream: Long-Horizon Attention for Streaming 3D Reconstruction","authors":["Chong Cheng","Peilin Tao","Nanjie Yao","Guanzhi Ding","Xianda Chen","yuansen du","Xiaoyang Guo","Wei Yin","Weiqiang Ren","Qian Zhang","Zhengqing Chen","Hao Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The readout limit and discounted-gradient interpretation need qualification","plainExplanation":"The geometric factor alone does not force the right side to vanish if the queries grow fast enough.","primaryUrl":"https://arxiv.org/html/2605.23889","noteUrl":"/review?paper=review-neurips-2026-153911","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153912","title":"Dynamic k-center clustering with lifetimes","authors":["Simone Moretti","Paolo Pellizzoni","Andrea Pietracaprina","Geppino Pucci"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Dynamic k-center clustering with lifetimes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.23348","noteUrl":"/review?paper=review-neurips-2026-153912","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153915","title":"Breaking Adversarial Transferability in Fine-Tuned Speech Recognition","authors":["Mojtaba Nafez","Aref Mousavi","Mohammad E Mahdavi","Mobina Poulaei","Kiarash K Feriz","Mohammad Hossein Rohban"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Breaking Adversarial Transferability in Fine-Tuned Speech Recognition","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.09109","noteUrl":"/review?paper=review-neurips-2026-153915","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153916","title":"Steer-to-Detect: Probing Hidden Representations for Detection of LLM-Generated Texts","authors":["Luxu LIANG","Xiang Li"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Steer-to-Detect: Probing Hidden Representations for Detection of LLM-Generated Texts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12890","noteUrl":"/review?paper=review-neurips-2026-153916","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153917","title":"Reinforcing VLAs in Task-Agnostic World Models","authors":["Yucen Wang","Rui Yu","Fengming Zhang","Junjie Lu","Kaixin Wang","Li Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Neutralized groups receive a group-dependent gradient scale","plainExplanation":"For \\(n\\) reliable members in a group of size \\(G\\), mean replacement preserves the reliable mean but multiplies full-group variance by \\(n/G\\). The outer averaging also contributes \\(n/G\\), so the standardized group gradient is scaled by \\(\\sqrt{n/G}\\). This factor varies between groups and can change their aggregate direction, rather than one common learning-rate constant. The harmless-global-scaling justification is false. The method’s usefulness and unavailable implementation choices are not refuted.","primaryUrl":"https://arxiv.org/html/2605.12334","noteUrl":"/review?paper=review-neurips-2026-153917","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153919","title":"OATS: Online Data Augmentation for Time Series Foundation Models","authors":["Junwei Deng","Chang Xu","Jiaqi Ma","Ming Jin","Chenghao Liu","Xu Zhang","Li Zhao","Jiang Bian"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Signed cache updates can leave the exploit sampler without positive mass","plainExplanation":"Train/reference losses \\((w+1)^2/2\\) and \\((w-1)^2/2\\) at zero give gradient dot product \\(-1\\). From cache \\(0.05\\) with update rate \\(0.01\\), repeated scores \\(-1\\) give \\(\\Phi_k=1.05(0.99)^k-1\\), negative by the fifth update. If every subset has such history, all exploit weights are zero. A random batch can also miss a subset entirely, leaving its update mean undefined. Exploration probability alone does not define a selected zero-mass exploit step. This is a manuscript boundary gap. The inspected optional public-code path instead uses positive min–max ratios with an equal-score fallback and epsilon-protected filtering; no released crash or benchmark failure follows.","primaryUrl":"https://arxiv.org/html/2601.19040","noteUrl":"/review?paper=review-neurips-2026-153919","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153921","title":"TKCAM: Text and Keyframe to Camera Trajectory Generation","authors":["Haozhe Yang","Zhiyang Dou","Zekai Gu","Cheng Lin","Wenping Wang","Yuan Liu","Taku Komura"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"TKCAM: Text and Keyframe to Camera Trajectory Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.11105","noteUrl":"/review?paper=review-neurips-2026-153921","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153922","title":"xHC: Expanded Hyper-Connections","authors":["Xiangdong Zhang","Xiaohan Qin","Tuo Dai","Xiaoming Shi","Huaijin Wu","Yebin Yang","Zhuo Xia","Shaofeng Zhang","Yu Wang","Yu Cheng","Junchi Yan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"xHC: Expanded Hyper-Connections","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.14530","noteUrl":"/review?paper=review-neurips-2026-153922","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153925","title":"Prospective Coding Improves Learning in Deep Continuous-Time Recurrent Networks","authors":["Shivang Rawat","Mirko Morello","Flaviano Morone","David Heeger"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Prospective Coding Improves Learning in Deep Continuous-Time Recurrent Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.04134","noteUrl":"/review?paper=review-neurips-2026-153925","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153927","title":"Sharp feature-learning transitions and Bayes-optimal neural scaling laws in extensive-width networks","authors":["Minh Toan Nguyen","Jean Barbier"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The displayed self-consistent system can have multiple error solutions","plainExplanation":"The displayed self-consistent system can have multiple error solutions.","primaryUrl":"https://arxiv.org/html/2605.10395","noteUrl":"/review?paper=review-neurips-2026-153927","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153929","title":"ExpLang: Improved Exploration and Exploitation in LLM Reasoning with On-Policy Thinking Language Selection","authors":["Changjiang Gao","Zixian Huang","Kaichen Yang","Jiajun Chen","Jixing Li","Shujian Huang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ExpLang: Improved Exploration and Exploitation in LLM Reasoning with On-Policy Thinking Language Selection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.21887","noteUrl":"/review?paper=review-neurips-2026-153929","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153931","title":"Rethinking Rubric Generation for Improving LLM Judge and Reward Modeling for Open-ended Tasks","authors":["William Shen","Xinchi Qiu","Chenxi Whitehouse","Lisa Alazraki","Shashwat Goel","Francesco Barbieri","Timon Willi","Akhil Mathur","Ilias Leontiadis"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Covariance is not automatically a sub-Gaussian MGF proxy","plainExplanation":"Covariance is not automatically a sub-Gaussian MGF proxy. The printed covariance-based implication is false; a correctly proxy-qualified Chernoff analysis remains possible.","primaryUrl":"https://arxiv.org/html/2602.05125","noteUrl":"/review?paper=review-neurips-2026-153931","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153932","title":"MAdam: Metric-Aware Multi-Objective Adam","authors":["Fengbei Liu","Rachit Saluja","Sunwoo Kwak","Ruibo Wang","Ruining Deng","Heejong Kim","Johannes C. Paetzold","Mert Sabuncu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Stale pairwise EMA entries need not form a PSD metric","plainExplanation":"A synchronous outer-product EMA preserves positive semidefiniteness; asynchronous entry updates do not. Start with bank \\(a\\begin{pmatrix}1&-1\\\\-1&1\\end{pmatrix}\\). Updating only its first diagonal to \\(\\beta a\\) gives uniform-simplex quadratic value \\(-(1-\\beta)a/4\\), negative. Choosing the scale large enough also defeats the fixed epsilon. This bank is reachable using valid pair updates followed by a zero first-coordinate gradient. The sampled-bank construction lacks the claimed square-root domain guarantee. Warmup alone does not prevent later inconsistency, and no experimental failure is inferred.","primaryUrl":"https://arxiv.org/html/2606.03904","noteUrl":"/review?paper=review-neurips-2026-153932","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153933","title":"Compute Aligned Training: Optimizing for Test Time Inference","authors":["Adam Ousherovitch","Ambuj Tewari"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The Best-of-$N$ sequence probability omits mixed tied batches","plainExplanation":"The Best-of-$N$ sequence probability omits mixed tied batches.","primaryUrl":"https://arxiv.org/html/2604.24957","noteUrl":"/review?paper=review-neurips-2026-153933","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153935","title":"Subliminal Learning Is Steering Vector Distillation","authors":["Camila Blank","Agam Bhatia","Senthooran Rajamanoharan","Arthur Conmy","Neel Nanda"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Residual-gradient alignment does not determine parameter-SGD movement","plainExplanation":"With activation Jacobian \\(J\\), the first-order activation difference is \\(\\eta JJ^\\top(g_{\\mathrm{clean}}-g_{\\mathrm{teacher}})\\), not the gradient difference itself. For \\(J=(1,2)^\\top\\), difference \\((1,-1)\\), and teacher direction \\((1,0)\\), the raw projection is positive but multiplication by \\(JJ^\\top\\) makes it negative. A two-class linear-softmax loss realizes the gradients exactly at zero. The cosine implication fails generally. Actual optimizer outcomes and the observed steering/distillation relationship are not refuted.","primaryUrl":"https://arxiv.org/html/2606.00995","noteUrl":"/review?paper=review-neurips-2026-153935","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153936","title":"ForecastCompass: Guiding Agentic Forecasting with Adaptive Factor Memory","authors":["Yurui Chang","Yongkang Du","Yuanpu Cao","Jinghui Chen","Lu Lin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ForecastCompass: Guiding Agentic Forecasting with Adaptive Factor Memory","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.30858","noteUrl":"/review?paper=review-neurips-2026-153936","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153937","title":"Token Time Continuous Diffusion for Language Modeling","authors":["Parikshit Bansal","Sujay Sanghavi"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Token Time Continuous Diffusion for Language Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.14106","noteUrl":"/review?paper=review-neurips-2026-153937","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153938","title":"Generation Navigator: A State-Aware Agentic Framework for Image Generation","authors":["Jinming Liu","Ruoyu Feng","Yuqi Wang","Wenjun Zeng","Xin Jin"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Generation Navigator: A State-Aware Agentic Framework for Image Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.17969","noteUrl":"/review?paper=review-neurips-2026-153938","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153940","title":"Test-time Multi-agent Coordination by Decomposed Value Gradient Flow","authors":["Dongsu Lee","Haoran Xu","Amy Zhang"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Test-time Multi-agent Coordination by Decomposed Value Gradient Flow","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02554","noteUrl":"/review?paper=review-neurips-2026-153940","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153942","title":"Communication-Efficient LLM Adaptation over Decentralized GPU Meshes","authors":["Sameera Ramasinghe","Shamane Siriwardhana","Thalaiyasingam Ajanthan","Hadi Mohaghegh Dolatabadi","Chamin Hewa Koneputugodage","Gil Avraham","Violetta Shevchenko","James Snewin","Karol Pajak","Harry Xi","Alexander Long"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Changing Top-$K$ support can give full-rank residual covariance","plainExplanation":"Unbiased masked activations do not imply unbiased parameter gradients. Arrival delay does not bound the age of retained EMA gradients. Changing Top-$K$ support can give full-rank residual covariance.","primaryUrl":"https://arxiv.org/html/2609.14339","noteUrl":"/review?paper=review-neurips-2026-153942","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153944","title":"Rethinking Projector Training in Multimodal LLMs","authors":["Fabian Gröger","Naël Ouerghemi","Shuo Wen","Maria Brbic"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rethinking Projector Training in Multimodal LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.18598","noteUrl":"/review?paper=review-neurips-2026-153944","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153945","title":"From Table to Cell: Attention for Better Reasoning with TABALIGN","authors":["Tung Sum Thomas Kwok","Zeyong Zhang","Xinyu Wang","Chunhe Wang","Xiaofeng Lin","Hanwei Wu","Lei Ding","Guang Cheng","Zhijiang Guo"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"From Table to Cell: Attention for Better Reasoning with TABALIGN","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.14465","noteUrl":"/review?paper=review-neurips-2026-153945","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153946","title":"A Minimal Interpretable Architecture for Zero-Shot Reconstruction of Dynamical Systems","authors":["Christoph Jürgen Hemmer","Florian Plaswig","Daniel Durstewitz"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Open contexts allow unbounded orbits at the $\\alpha=1$ endpoint","plainExplanation":"A nearest-neighbor switch inside an ODE step breaks the discrete-map embedding. Open contexts allow unbounded orbits at the $\\alpha=1$ endpoint.","primaryUrl":"https://arxiv.org/html/2607.14937","noteUrl":"/review?paper=review-neurips-2026-153946","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153949","title":"D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models","authors":["Dengyang Jiang","Xin Jin","Dongyang Liu","Zanyi Wang","Mingzhe Zheng","Ruoyi Du","Xiangpeng Yang","Qilong Wu","Zhen Li","Peng Gao","Harry Yang","Steven Hoi"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05204","noteUrl":"/review?paper=review-neurips-2026-153949","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153950","title":"Efficient Generative Transformer Operators for Million-Point PDEs","authors":["Armand Kassaï Koupaï","Lise Le Boudec","Patrick Gallinari"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Efficient Generative Transformer Operators for Million-Point PDEs","plainExplanation":"Incompressibility requires zero divergence, whereas zero time derivative requires a stationary velocity field.","primaryUrl":"https://arxiv.org/html/2512.04974","noteUrl":"/review?paper=review-neurips-2026-153950","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-153952","title":"Cite What You Explore: Budget-Aware LLM Reasoning over Medical KGs with Verifiable Evidence","authors":["Chen Chen","Dongjie Wang","Mei Liu","Zijun Yao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The post-charge minimum-component test does not enforce the hard acquisition budget","plainExplanation":"The post-charge minimum-component test does not enforce the hard acquisition budget. The printed loop does not implement a hard cap; reported values and released implementation overruns are not established by this pseudocode.","primaryUrl":"https://arxiv.org/html/2610.07739","noteUrl":"/review?paper=review-neurips-2026-153952","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153955","title":"From Senses to Decisions: The Information Flow of Auditory and Visual Perception in Multimodal LLMs","authors":["Wish Suharitdamrong","Muhammad Awais","Xiatian Zhu","Sara Atito"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"From Senses to Decisions: The Information Flow of Auditory and Visual Perception in Multimodal LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.10147","noteUrl":"/review?paper=review-neurips-2026-153955","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153956","title":"BitDance: Scaling Autoregressive Generative Models with Binary Tokens","authors":["Yuang Ai","Jiaming Han","Shaobin Zhuang","Weijia Mao","Xuefeng Hu","Ziyan Yang","Zhenheng Yang","Yali Wang","Xiangyu Yue","Hao Chen","Huaibo Huang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"BitDance: Scaling Autoregressive Generative Models with Binary Tokens","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2602.14041v2","noteUrl":"/review?paper=review-neurips-2026-153956","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153957","title":"Elephant in the Fridge: Constant-Memory Frame Packing for Long Video Understanding","authors":["Shuo Gao","Chenhao Zheng","Jason Ren"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An integer token budget cannot retain arbitrarily many independently encoded frames","plainExplanation":"An integer token budget cannot retain arbitrarily many independently encoded frames.","primaryUrl":"https://arxiv.org/html/2610.16118","noteUrl":"/review?paper=review-neurips-2026-153957","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-153959","title":"Anchor PCA","authors":["Benedikt Seiter","Anya Fries","Julius von Kügelgen","Jonas Peters"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Anchor PCA","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06233","noteUrl":"/review?paper=review-neurips-2026-153959","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153960","title":"Simulation-free Unbalanced Dynamic Optimal Transport with General Growth Penalty","authors":["Junda Ying","Yuxuan Wang","Bowen Yang","Peijie Zhou","Lei Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Simulation-free Unbalanced Dynamic Optimal Transport with General Growth Penalty","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.04710","noteUrl":"/review?paper=review-neurips-2026-153960","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153962","title":"From Perception to Punchline: Empowering VLM with the Art of In-the-wild Memes","authors":["Xueyan Li","Yingyi Xue","Mengjie Jiang","Qingzi Zhu","Yazhe Niu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Monotone links and learned-error conditioning need stronger assumptions","plainExplanation":"A general increasing comparison link need not preserve latent rankings under a shared logistic score model with nonuniform pair sampling. The appendix also transfers a margin property of annotation noise to learned-classifier errors without an additional assumption; the main text’s annotation-noise interpretation remains valid.","primaryUrl":"https://arxiv.org/html/2512.24555","noteUrl":"/review?paper=review-neurips-2026-153962","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference"],"classifications":["major"]},{"slug":"review-neurips-2026-153964","title":"Lean4Agent: Formal Modeling and Verification for Agent Workflow and Trajectory","authors":["Ruida Wang","Jerry Huang","Pengcheng Wang","Xuanqing Liu","Chris Kong","Tong Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Lean4Agent: Formal Modeling and Verification for Agent Workflow and Trajectory","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.06523","noteUrl":"/review?paper=review-neurips-2026-153964","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153966","title":"SpikeSSL: A Universal Spike Inference Framework with Dynamics-Informed State-Space Layers","authors":["Chenghao Yue","Siming Xing","shuran liu","Angran Li","Yuanlong Zhang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"SpikeSSL: A Universal Spike Inference Framework with Dynamics-Informed State-Space Layers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.11456","noteUrl":"/review?paper=review-neurips-2026-153966","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153967","title":"When Less is More: The LLM Scaling Paradox in Context Compression","authors":["Ruishan Guo","Yibing Liu","Guoxin Ma","Yan Wang","Yueyang Zhang","Long Xia","Kecheng Chen","Zhiyuan Sun","Daiting Shi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"When Less is More: The LLM Scaling Paradox in Context Compression","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.09789","noteUrl":"/review?paper=review-neurips-2026-153967","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153968","title":"Posterior Contraction Rates for sparse Kolmogorov-Arnold Networks in Anisotropic Besov Spaces","authors":["Jeunghun Oh","Lizhen Lin","Jaeyong Lee","Kyeongwon Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The general-grid Lipschitz constant underestimates SiLU","plainExplanation":"Lemma B.4 bounds each composed KAN tail by a product of layer constants containing the SiLU allowance \\(1\\). Knot spacing is unrestricted in this lemma. A scalar layer with zero spline weights and unit SiLU weight is admissible.","primaryUrl":"https://arxiv.org/html/2605.11652","noteUrl":"/review?paper=review-neurips-2026-153968","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153969","title":"Preconditioned Flow Matching","authors":["Shadab Ahamed","Eshed Gal","Md Shahriar Rahim Siddiqui","Simon Ghyselincks","Moshe Eliasof","Eldad Haber"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Preconditioned Flow Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.02337","noteUrl":"/review?paper=review-neurips-2026-153969","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153971","title":"AlphaPareto: Formulaic Alpha Discovery with LLM-Guided Multi-Objective Reinforcement Learning","authors":["Yingbo Zhao","Zeyu Yang","Zhoufan Zhu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Concavity does not ensure a finite unique dual maximizer","plainExplanation":"The scalarized reward uses a claimed unique maximizer of \\(g(\\lambda)=-\\log\\mathbb E_\\pi e^{\\lambda^\\top r}+\\lambda^\\top c\\) over \\(\\lambda\\geq0\\). The Hessian is minus a tilted reward covariance, which implies concavity but neither strict concavity nor attainment.","primaryUrl":"https://arxiv.org/html/2609.34188","noteUrl":"/review?paper=review-neurips-2026-153971","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153972","title":"Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry","authors":["Zemin Yang","Yaoyu He","Yiming Zhong","Yuhao Zhang","Xinge ZHU","Yao Mu","Qingqiu Huang","Yuexin Ma"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Zero self-displacement does not remove self-normalization","plainExplanation":"The main neighborhood excludes the current sample, while the appendix includes it in the softmax and argues that its zero displacement makes the constructions equivalent. Including the self weight \\(w_{ii}\\) multiplies every off-diagonal second moment by \\(1-w_{ii}\\).","primaryUrl":"https://arxiv.org/html/2606.01098","noteUrl":"/review?paper=review-neurips-2026-153972","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153973","title":"Proxy-Based Approximation of Shapley and Banzhaf Interactions","authors":["Santo Thies","Hubert Baniecki","R. Teal Witter","Eyke Hüllermeier","Maximilian Muschalik","Fabian Fumagalli"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Proxy-Based Approximation of Shapley and Banzhaf Interactions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22738","noteUrl":"/review?paper=review-neurips-2026-153973","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153974","title":"GLUT: 3D Gaussian Lookup Table for Continuous Color Transformation","authors":["Danna Xue","David Serrano-Lozano","Shaolin Su","Javier Vazquez-Corral"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Decay of a Gaussian density does not imply locality of its normalized influence","plainExplanation":"Decay of a Gaussian density does not imply locality of its normalized influence. The normalized hue loss has no continuous extension at a neutral prediction.","primaryUrl":"https://arxiv.org/html/2605.19889","noteUrl":"/review?paper=review-neurips-2026-153974","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-153975","title":"DUET: Unified Dual-Space Emotion Control for Diffusion and Flow-Matching Driven Text-to-Speech","authors":["Xu Zhang","Longbing Cao","zhangkai wu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Raw-logit directions depend on an unidentifiable common mode","plainExplanation":"DUET selects singular directions of a classifier matrix \\(W\\) as directions of emotion sensitivity, using raw-logit changes \\(\\|Wv\\|^2\\). Replacing \\(W\\) by \\(W+\\mathbf1 a^\\top\\) adds the same scalar \\(a^\\top h\\) to every class logit and leaves softmax probabilities unchanged.","primaryUrl":"https://arxiv.org/html/2606.00066","noteUrl":"/review?paper=review-neurips-2026-153975","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153977","title":"The Subjectivity of Monoculture","authors":["Nathanael Jo","Nikhil Garg","Manish Raghavan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Subjectivity of Monoculture","plainExplanation":"No issue was found in the checked claims. The criticized pointwise inference is absent from the revised argument. The new proof controls the sum over item–model pairs on the orthogonal complement of the reparametrization kernel.","primaryUrl":"https://arxiv.org/html/2602.24086","noteUrl":"/review?paper=review-neurips-2026-153977","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153978","title":"Measuring and Decomposing Mode Separation via the Canonical Diffusion","authors":["Shaul Tolkowsky","Ori Meidler","Or Zuk"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Time-step refinement does not remove finite-path correlation bias","plainExplanation":"The mode-separation argument identifies an empirical squared-correlation statistic with its population canonical-diffusion counterpart at fixed physical horizon, claiming the relation for any sample count \\(N\\). The empirical statistic uses a random estimated mean and a random covariance ratio before squaring.","primaryUrl":"https://arxiv.org/html/2605.08777","noteUrl":"/review?paper=review-neurips-2026-153978","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-153980","title":"A Hierarchy of Entropy-Shapley Games for Multivariate Predictive Uncertainty","authors":["Niklas Koenen","Claudia Battistin","Jeriek Van den Abeele","Martin Jullum"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"A Hierarchy of Entropy-Shapley Games for Multivariate Predictive Uncertainty","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.35217","noteUrl":"/review?paper=review-neurips-2026-153980","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153981","title":"Less Supervision, Better Generalization: Weakly Supervised Fake Region Localization in Diffusion-Edited Images","authors":["Junhee Lee","JeonDongHyeon","Taeoh Kim","beomYoung Kim","MyeongAh Cho"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Less Supervision, Better Generalization: Weakly Supervised Fake Region Localization in Diffusion-Edited Images","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.36882","noteUrl":"/review?paper=review-neurips-2026-153981","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153982","title":"Hint Tuning: Less Data Makes Better Reasoners","authors":["Siqi Fan","Minghao Li","Xiaoqian Ma","Xiusheng Huang","Zhuo Chen","Bowen Qin","ZhangLiujie","Shuo Shang","Weihang Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Hint Tuning: Less Data Makes Better Reasoners","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08665","noteUrl":"/review?paper=review-neurips-2026-153982","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153983","title":"Audio-Visual Flamingo: Open Audio-Visual Intelligence for Long and Complex Videos","authors":["Sreyan Ghosh","Arushi Goel","Kaousheik Jayakumar","Lasha Koroshinadze","Nishit Anand","Siddharth Gururani","Hanrong Ye","Pritam Biswas","Yuanhang Su","Ehsan Hosseini-Asl","Sang-gil Lee","Zhifeng Kong","Jaehyeon Kim","Sungwon Kim","S Sakshi","Ramani Duraiswami","Dinesh Manocha","Andrew Tao","Mohammad Shoeybi","Bryan Catanzaro","Ming-Yu Liu","Wei Ping"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Audio-Visual Flamingo: Open Audio-Visual Intelligence for Long and Complex Videos","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.16107","noteUrl":"/review?paper=review-neurips-2026-153983","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153987","title":"FedAdaVR: Adaptive Variance Reduction for Robust Federated Learning under Limited Client Participation","authors":["S M RUHUL KABIR HOWLADER","Xiao Chen","Yifei Xie","Lu Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FedAdaVR: Adaptive Variance Reduction for Robust Federated Learning under Limited Client Participation","plainExplanation":"Identical deterministic clients already expose both problems. Arbitrarily small positive rates leave the model near its nonstationary initialization, while the claimed leading rate is unchanged; the pseudo-gradient also retains a fixed client-rate and local-step scaling factor.","primaryUrl":"https://arxiv.org/html/2601.22204","noteUrl":"/review?paper=review-neurips-2026-153987","collection":"neurips-2026","severity":"critical","issueTypes":["formula_error","incorrect_bound","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-153990","title":"CompactAttention: Accelerating Chunked Prefill with Block-Union KV Selection","authors":["Jiwon Song","Dongwon Jo","Beomseok Kang","jae-joon kim"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"CompactAttention: Accelerating Chunked Prefill with Block-Union KV Selection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16839","noteUrl":"/review?paper=review-neurips-2026-153990","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153992","title":"Conditional Evaluation of Language Models with Cheap Auxiliary Signals","authors":["Zhi Zhang","Lingfeng Lyu","Yue Kang","Doudou Zhou"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Conditional Evaluation of Language Models with Cheap Auxiliary Signals","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.16210","noteUrl":"/review?paper=review-neurips-2026-153992","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153993","title":"Learning with Multiple Correct Answers - Regret Bounds under Different Feedback Models","authors":["Alireza F. Pour","Farnam Mansouri","Shai Ben-David"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Learning with Multiple Correct Answers - Regret Bounds under Different Feedback Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09402","noteUrl":"/review?paper=review-neurips-2026-153993","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153996","title":"Effective Context in Transformers: An Analysis of Fragmentation and Tokenization","authors":["Amirmehdi Jafari Fesharaki","Mohammadamin Rami","Aslan Tchamkerten"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Effective Context in Transformers: An Analysis of Fragmentation and Tokenization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13485","noteUrl":"/review?paper=review-neurips-2026-153996","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153997","title":"Invertible Logits Transformation for Accuracy-Preserving Post-Hoc Uncertainty Calibration","authors":["Lening Zhao","Qipeng Zhan","Li Shen"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Invertible Logits Transformation for Accuracy-Preserving Post-Hoc Uncertainty Calibration","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.10372","noteUrl":"/review?paper=review-neurips-2026-153997","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-153998","title":"Revisiting Subgradient Dominance in Robust MDPs: Counterexamples, Hardness, and Sufficient Conditions","authors":["Toshinori Kitamura","Arnob Ghosh","Alex Ayoub","Thang Chu","Csaba Szepesvari"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Revisiting Subgradient Dominance in Robust MDPs: Counterexamples, Hardness, and Sufficient Conditions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.21177","noteUrl":"/review?paper=review-neurips-2026-153998","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-153999","title":"DisCoMBO: Steering Expert-in-the-Loop Black Box Optimization via Distributional Conformance","authors":["Jonas Seng","Bennet Wittelsbach","Kristian Kersting"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The circuit expansion and regret proof need corrections","plainExplanation":"Recursive Fisher aggregation does not equal the stated flat induced-tree mixture. The regret proof also inverts the Beta shape implied by the bin exponent and treats maximum uncertainty as maximum upper confidence bound. These are distinct algebraic and proof issues; the recursive score remains computable.","primaryUrl":"https://arxiv.org/html/2609.36472","noteUrl":"/review?paper=review-neurips-2026-153999","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound","invalid_inference"],"classifications":["major"]},{"slug":"review-neurips-2026-154001","title":"Model-Free Assessment of Simulator Fidelity via Quantile Curves","authors":["Yu-Shiou Lin","Kaizheng Wang","Garud Iyengar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The non-strict CDF at a quantile need not have the claimed upper bound","plainExplanation":"If every discrepancy equals one, empirical and population CDFs agree exactly, but the non-strict CDF at every positive-index empirical quantile is one. At an interior quantile level this violates the proposed upper bridge even with zero CDF estimation error. A quantile controls counts strictly below and at the threshold differently when atoms occur. The intermediate CDF assertion is false. The final greater-than-or-equal tail theorem may remain true after the correct left-limit proof.","primaryUrl":"https://arxiv.org/html/2512.05024","noteUrl":"/review?paper=review-neurips-2026-154001","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154002","title":"Optimal Dimension-Free Sampling for Regularized Classification","authors":["Meyam Alishahi","Alexander Munteanu","Simon Omlor","Jeff M Phillips"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The regularized-risk lower bounds need truncation by the zero-parameter risk","plainExplanation":"The integrated derivative bound needs an absolute-value exponent. Increment entropy does not control a Gaussian process’s arbitrary common offset. The regularized-risk lower bounds need truncation by the zero-parameter risk. One-dimensional ReLU sampling need not have a support-size logarithmic lower bound.","primaryUrl":"https://arxiv.org/html/2605.23726","noteUrl":"/review?paper=review-neurips-2026-154002","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154004","title":"AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models","authors":["Shouwei Ruan","Bin Wang","Zhenyu Wu","Qihui Zhu","Yuxiang Zhang","Jingzhi Li","Yubin Wang","Xingxing Wei"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.08952","noteUrl":"/review?paper=review-neurips-2026-154004","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154006","title":"GeoX: Mastering Geospatial Reasoning Through Self-Play and Verifiable Rewards","authors":["Kyeongjin Ahn","Seungeon Lee","Krishna Gummadi","Meeyoung Cha"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"GeoX: Mastering Geospatial Reasoning Through Self-Play and Verifiable Rewards","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20006","noteUrl":"/review?paper=review-neurips-2026-154006","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154007","title":"Corrective Diffusion Language Models","authors":["Shuibai Zhang","Fred Peng","Yiheng Zhang","Jin Pan","Grigorios Chrysos"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"One printed iteration cannot replace a visible incorrect token","plainExplanation":"The listing predicts new values only for positions already masked, then may mask low-confidence positions. Starting with a fully visible failing program, one iteration can retain or mask a token but cannot turn an incorrect visible token into a correct visible token. Section 2.2’s all-position argmax is a different update. This is an algorithm/protocol inconsistency. The nonzero one-step measurements are not alleged fabricated or independently recomputed.","primaryUrl":"https://arxiv.org/html/2512.15596","noteUrl":"/review?paper=review-neurips-2026-154007","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154011","title":"Stability and Generalization in Looped Transformers","authors":["Asher Labovich"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Strict spectral contraction is sufficient but not necessary for nonlinear attraction","plainExplanation":"Take \\(J=\\begin{pmatrix}3&3\\\\-3&-3\\end{pmatrix}\\), for which \\(J^2=0\\) and spectral radius zero. The allowed RMSNorm-style map \\(\\phi(v)=\\operatorname{diag}(0.8,0.2)v/\\sqrt{1+\\|v\\|^2/2}\\) is globally \\(0.8\\)-Lipschitz and has derivative \\(G=\\operatorname{diag}(0.8,0.2)\\) at zero. Yet \\(GJ\\) has eigenvalue \\(1.8\\). With recall maps \\(g(x,u)=x+u,h_1=0,h_2(z)=(J-I)z\\), choose \\(\\phi_1(v)=0.8v/\\sqrt{1+\\|v\\|^2/2}\\) and \\(\\phi_2=\\phi\\). Both normalizers are globally contractive and the derivative becomes \\(0.8GJ\\) with eigenvalue \\(1.44>1\\). The general normalization-to-stability implication fails for nonnormal maps. Conditional strict-radius theorems and empirical trends remain valid.","primaryUrl":"https://arxiv.org/html/2604.15259","noteUrl":"/review?paper=review-neurips-2026-154011","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154013","title":"Coupling Models for One-Step Discrete Generation","authors":["Fred Peng","Joey Bose","Anru Zhang","Alexander Tong"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Coupling Models for One-Step Discrete Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07193","noteUrl":"/review?paper=review-neurips-2026-154013","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154014","title":"FUTON: Fourier Tensor Network for Implicit Neural Representations","authors":["Pooya Ashtari","Pourya Behmandpoor","Nikos Deligiannis","Aleksandra Pizurica"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"First-kind Chebyshev polynomials are not orthogonal for the stated unweighted inner product","plainExplanation":"First-kind Chebyshev polynomials are not orthogonal for the stated unweighted inner product.","primaryUrl":"https://arxiv.org/html/2602.13414","noteUrl":"/review?paper=review-neurips-2026-154014","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154015","title":"Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs","authors":["Alexander Panfilov","Peter Romov","Igor Shilov","Yves-Alexandre de Montjoye","Jonas Geiping","Maksym Andriushchenko"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.24511","noteUrl":"/review?paper=review-neurips-2026-154015","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154016","title":"KV-COBRA: KV Cache Compression via Co-Optimized Bit-Rank Allocation","authors":["Sihyeon Ha","Jaeho Lee","Yo-Seb Jeon"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The absolute Gaussian quantization constant violates the rate-distortion lower bound","plainExplanation":"Nearest-integer bit rounding can select an infeasible product budget. The even-rank grid includes dimensions without the required equal-magnitude orthogonal matrix. The absolute Gaussian quantization constant violates the rate-distortion lower bound.","primaryUrl":"https://arxiv.org/html/2609.24298","noteUrl":"/review?paper=review-neurips-2026-154016","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154017","title":"PoseBridge: Bridging the Skeletonization Gap for Zero-Shot Skeleton-Based Action Recognition","authors":["Sanghyeon Lee","Jinwoo Kim","Jong Taek Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"PoseBridge: Bridging the Skeletonization Gap for Zero-Shot Skeleton-Based Action Recognition","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11497","noteUrl":"/review?paper=review-neurips-2026-154017","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154018","title":"One-Bit Clustering for Two Component Sub-Gaussian Mixture Models","authors":["Junren Chen","Yun Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Conditional Gram concentration does not imply concentration about an uncontrolled unconditional mean","plainExplanation":"Conditional Gram concentration does not imply concentration about an uncontrolled unconditional mean.","primaryUrl":"https://arxiv.org/html/2606.21873","noteUrl":"/review?paper=review-neurips-2026-154018","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-154021","title":"HuPER: A Human-Inspired Framework for Phonetic Perception","authors":["Chenxu Guo","Jiachen Lian","Louis Liu","Baihe Huang","Shriyaa Narayanan","Bixing Wu","Zoe Ezzes","Jet Vonk","Zachary Miller","Cheol Jun Cho","Maria Luisa Gorno Tempini","Gopala Anumanchipalli"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"HuPER: A Human-Inspired Framework for Phonetic Perception","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.01634","noteUrl":"/review?paper=review-neurips-2026-154021","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154023","title":"TIGER-FG: Text-Guided Implicit Fine-Grained Grounding for E-commerce Retrieval","authors":["Xinyu Sun","Huangyu Dai","Lingtao Mao","Zexin Zheng","Zihan Liang","Ben Chen","Chenyi Lei","Wenwu Ou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TIGER-FG: Text-Guided Implicit Fine-Grained Grounding for E-commerce Retrieval","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.18434","noteUrl":"/review?paper=review-neurips-2026-154023","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154025","title":"Reward Hacking in Rubric-Based Reinforcement Learning","authors":["Anas Mahmoud","MohammadHossein Rezaei","Zihao Wang","Anisha Gunjal","Bing Liu","Yunzhong He"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Reward Hacking in Rubric-Based Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12474","noteUrl":"/review?paper=review-neurips-2026-154025","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154026","title":"Multi-Token Residual Prediction","authors":["Yufeng Xu","Zishuo Bao","Qian Wang","Zeshen Zhang","Haoqi Zhang","Bowen Peng","Ang Li","Rahul Chalamala","Yucheng Lu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Multi-Token Residual Prediction","plainExplanation":"A full-tensor norm bound can concentrate on one output row. It supplies no averaging factor of inverse sequence length.","primaryUrl":"https://arxiv.org/html/2605.18817","noteUrl":"/review?paper=review-neurips-2026-154026","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","mathematical_error"],"classifications":["major"]},{"slug":"review-neurips-2026-154033","title":"ROLLVERIFY: BRIDGING EFFICIENCY AND ACCURACY IN LONG-TAIL ROLLOUT REINFORCEMENT LEARNING","authors":["Yongqiang Yao","Jingru Tan","Kaihuan Liang","Zixin Yin","Yazhe Niu","Ruihao Gong","Dahua Lin","Ningyi Xu"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"ROLLVERIFY: BRIDGING EFFICIENCY AND ACCURACY IN LONG-TAIL ROLLOUT REINFORCEMENT LEARNING","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.09914","noteUrl":"/review?paper=review-neurips-2026-154033","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154037","title":"Orthros: Phase-Aware Heterogeneous Attention for Efficient Transformers","authors":["Yuhong CHOU","Zehao Liu","Yuqi Pan","Qian Liu","Xianwei Chen","Jibin Wu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Orthros: Phase-Aware Heterogeneous Attention for Efficient Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.13735","noteUrl":"/review?paper=review-neurips-2026-154037","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154038","title":"Leveraging unlabelled data for generalizable neural population decoding","authors":["Ximeng Mao","Nanda H Krishna","Hee-Woon Ryoo","Matthew G Perich","Guillaume Lajoie"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Leveraging unlabelled data for generalizable neural population decoding","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.14086","noteUrl":"/review?paper=review-neurips-2026-154038","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154040","title":"EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields","authors":["Zhaoyang Yang","Yurun Jin","Lizhe Qi","Cong Huang","Kai Chen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06192","noteUrl":"/review?paper=review-neurips-2026-154040","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154043","title":"ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents","authors":["XuHao Hu","Xi Zhang","Haiyang Xu","Kyle Qiao","Jingyi Yang","Xuanjing Huang","Jing Shao","Ming Yan","Jieping Ye"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Equal horizon lengths make the efficiency reward undefined","plainExplanation":"A mixed-success group can have every trajectory terminate at the horizon. For each successful trajectory, \\(s=\\bar s=S_{\\max}\\), selecting the exponential branch with both numerator and denominator zero. Dynamic filtering removes homogeneous success groups, not these equal-length mixed groups. The legal reward boundary is undefined as printed. Actual code guards and runtime behavior are not established.","primaryUrl":"https://arxiv.org/html/2605.12481","noteUrl":"/review?paper=review-neurips-2026-154043","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154044","title":"Shared Modular Recurrence in Contextual MDPs for Universal Morphology Control","authors":["Laurens Engwegen","Max Weltevrede","Caroline Horsch","Daan Brinks","Wendelin Boehmer"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Shared Modular Recurrence in Contextual MDPs for Universal Morphology Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2506.08630","noteUrl":"/review?paper=review-neurips-2026-154044","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154045","title":"Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations","authors":["Hanwei Zhang","Tianma Hu","Gaojie Jin","Xu Cheng","Ronghui Mu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The theorem should state the swap-only domain proved in the appendix","plainExplanation":"Theorem 3.2’s displayed quantifier includes arbitrary continuous vectors with nearby binarizations, while the revised appendix explicitly proves a guarantee for concept-pair swaps. Adding that domain qualifier aligns the theorem with the existing argument.","primaryUrl":"https://arxiv.org/html/2609.38625","noteUrl":"/review?paper=review-neurips-2026-154045","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-154046","title":"From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills","authors":["Zisu Huang","Jingwen Xu","Yifan Yang","Ziyang Gong","Qihao Yang","Muzhao Tian","Xiaohua Wang","Changze Lv","Xuemei Gao","Qi Dai","Bei Liu","Kai Qiu","Xue Yang","Dongdong Chen","Xiaoqing Zheng","Chong Luo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.23899","noteUrl":"/review?paper=review-neurips-2026-154046","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154047","title":"Does Weight Decay Enhance Training Stability?","authors":["Marius Sæther","Amir Kolić","Pierfrancesco Beneventano","Tomaso Poggio"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stability and NTK arguments need discrete and decay-aware conditions","plainExplanation":"The ODE has negative real parts, but the Euler envelope map can have eigenvalues outside the unit disk. Cycle existence and cycle attraction are separate assertions.","primaryUrl":"https://arxiv.org/html/2605.16622","noteUrl":"/review?paper=review-neurips-2026-154047","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154049","title":"Leveraging Error Diversity in Group Rollouts for Reinforcement Learning","authors":["Wenpu Liu","Yuqi Xu","Weichu Xie","Yongfu Zhu","Shuai Dong","Ziyue Wang","Wenqi Shao","Xiaoying Zhang","Tong Yang","Nan Duan","Jiaqi Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Zero-sum advantage changes need not preserve the reward gradient","plainExplanation":"Policy gradients weight each change by its action score: \\(\\sum_i\\Delta_i=0\\) does not imply \\(\\sum_i\\Delta_i\\nabla\\log\\pi(a_i)=0\\). For a three-outcome policy with probabilities \\((1/2,1/4,1/4)\\), rewarding only the third outcome, enumeration of all four-rollout groups gives true and leave-one-out gradients \\((-0.125,-0.0625,0.1875)\\), but an allowed healthy-branch adjustment shifts the first two entries by approximately \\((-0.0013144,0.0013144)\\), with clipping inactive. The separate all-wrong penalty is not needed. The zero-sum-to-unbiasedness implication is false. Intentionally optimizing a useful different objective remains possible.","primaryUrl":"https://arxiv.org/html/2605.17333","noteUrl":"/review?paper=review-neurips-2026-154049","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154051","title":"The Hidden Ratio in Adam: Stable Structure, Compression, and Sign Dynamics","authors":["Yihe Zhou","Tongtian Zhu","Yingxiao Huo","Satya P Dash","Can Wang","Samuel Kaski","Mingfei Sun"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The ratio-state Adam chart breaks at moment cancellation","plainExplanation":"The proposed exact Adam reformulation replaces the second moment \\(v_t\\) by a ratio involving the first moment \\(m_t\\). Its algorithm divides by a previous first moment. With both decay parameters \\(1/2\\) and gradients \\(1,-1/2\\), the second step has \\(m_2=0\\) and \\(v_2=3/8\\).","primaryUrl":"https://arxiv.org/html/2609.35392","noteUrl":"/review?paper=review-neurips-2026-154051","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154053","title":"AI Agents May Always Fall for Prompt Injections","authors":["Sahar Abdelnabi","Eugene Bagdasarian"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AI Agents May Always Fall for Prompt Injections","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17634","noteUrl":"/review?paper=review-neurips-2026-154053","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154059","title":"Uncovering Challenges of Solving the Continuous Gromov-Wasserstein Problem","authors":["Xavier Aramayo-Carrasco","Maksim Nekrashevich","Petr Mokrov","Evgeny Burnaev","Aleksandr Korotin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An independently chosen optimal outer matrix need not recover every optimal transport map","plainExplanation":"An independently chosen optimal outer matrix need not recover every optimal transport map. The critic terms have the same sign instead of forming a marginal discrepancy.","primaryUrl":"https://arxiv.org/html/2303.05978","noteUrl":"/review?paper=review-neurips-2026-154059","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154065","title":"Quantitative Local Convergence of Mean-Field Stein Variational Gradient Flow","authors":["Lénaïc Chizat","Maria Colombo","Roberto Colombo","Xavier Fernández-Real"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Quantitative Local Convergence of Mean-Field Stein Variational Gradient Flow","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09456","noteUrl":"/review?paper=review-neurips-2026-154065","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154070","title":"Survive or Collapse: The Asymmetric Roles of Data Gating and Reward Grounding in Self-Play RL","authors":["Sophia Xiao Pu","Zhaotian Weng","Chengzhi Liu","Jayanth Srinivasa","Gaowen Liu","William Yang Wang","Xin Wang"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Shared proposer updates directly change the solver policy","plainExplanation":"The coupling account treats the proposer as affecting the solver only through the generated task distribution, although both roles share model parameters. For a proposer update \\(\\Delta w\\), the solver log probability changes to first order by \\(\\nabla_w\\log\\pi_S\\cdot\\Delta w\\), even with the task pool and rewards frozen.","primaryUrl":"https://arxiv.org/html/2605.22217","noteUrl":"/review?paper=review-neurips-2026-154070","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154072","title":"Characterizing Memorization in Diffusion Language Models: Generalized Extraction and Sampling Effects","authors":["Xiaoyu Luo","Wenrui Yu","Qiongxiu Li","Johannes Bjerva"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The block-refinement theorem guarantees nondecrease, not strict increase for all schedules","plainExplanation":"The block-refinement theorem guarantees nondecrease, not strict increase for all schedules. Adaptive generation order cannot be averaged using unconditional order frequencies.","primaryUrl":"https://arxiv.org/html/2603.02333","noteUrl":"/review?paper=review-neurips-2026-154072","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154073","title":"Exact Flow Linear Attention: Exact Solution from Continuous-Time Dynamics","authors":["Jingdi Lei","Di Zhang","Soujanya Poria"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Every expanded state contribution must use the current query","plainExplanation":"The expanded output uses each historical query where the current query is required. This is inconsistent with both the state recurrence and the later chunkwise matrix formula. This is a local query-index transcription error. The subsequent chunk-parallel formula uses the correct query interaction and is not refuted.","primaryUrl":"https://arxiv.org/html/2512.12602","noteUrl":"/review?paper=review-neurips-2026-154073","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-154074","title":"SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows","authors":["Albert Gao","Bing Xue","Andrea Zanette"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10563","noteUrl":"/review?paper=review-neurips-2026-154074","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154075","title":"Spectral Transformer Neural Processes","authors":["XIANHE CHEN","Hao Chen","Yingzhen Li"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Spectral Transformer Neural Processes","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09498","noteUrl":"/review?paper=review-neurips-2026-154075","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154076","title":"Isotropic Activation Functions Enable Deindividuated Neurons and Adaptive Topologies","authors":["George Bird"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The diagonal covariance formula has a local transpose error","plainExplanation":"The transpose sends variances to the wrong coordinates under the paragraph’s own convention.","primaryUrl":"https://arxiv.org/html/2602.23405","noteUrl":"/review?paper=review-neurips-2026-154076","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-154077","title":"RefineAny3D: Depth Refinement as Semantic Alignment for Monocular 3D Detection","authors":["Zhihao Zhang","Gengwei Zhang","Tianlong Chen","Xiaoming Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RefineAny3D: Depth Refinement as Semantic Alignment for Monocular 3D Detection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.09147","noteUrl":"/review?paper=review-neurips-2026-154077","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154081","title":"GLINT: Sparsely Gated Vision-Language Alignment for Fine-Grained Radiology Representations","authors":["Jonggwon Park","Seongeun Lee","Junhyun Park","Hannah Yun","Hyunwoong Kim","Sohyun Jeong","Hyewon Kang","Byungmu Yoon","Kyoyun Choi"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"GLINT: Sparsely Gated Vision-Language Alignment for Fine-Grained Radiology Representations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03180","noteUrl":"/review?paper=review-neurips-2026-154081","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154084","title":"Are Your Reasoning Models Reasoning or Guessing? A Mechanistic Analysis of Hierarchical Reasoning Models","authors":["Zirui Ren","Ziming Liu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Are Your Reasoning Models Reasoning or Guessing? A Mechanistic Analysis of Hierarchical Reasoning Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.10679","noteUrl":"/review?paper=review-neurips-2026-154084","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154085","title":"Group Distributionally Robust Optimization with Flexible Sample Queries","authors":["Haomin Bai","Dingzhi Yu","Shuai Li","Haipeng Luo","Lijun Zhang"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Group Distributionally Robust Optimization with Flexible Sample Queries","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.15212","noteUrl":"/review?paper=review-neurips-2026-154085","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154089","title":"Learning Distributions from Multiple Data Providers","authors":["Jon Kleinberg","Amin Saberi","Xizhi Tan","Grigoris Velegkas"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Learning Distributions from Multiple Data Providers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.24732","noteUrl":"/review?paper=review-neurips-2026-154089","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154094","title":"What if Agents Could Imagine? Reinforcing Open-Vocabulary HOI Comprehension through Generation","authors":["Zhenlong Yuan","Yue Wang","Jing Tang","Rui Chen","Kejin Cui","Lei Sun","Dapeng Zhang","Hongwei Yu","Chengxuan Qian","Xiangxiang Chu","Shuo Li","Yuyin Zhou"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The tool reward can increase without improving the prediction","plainExplanation":"The tool reward can increase without improving the prediction.","primaryUrl":"https://arxiv.org/html/2602.11499","noteUrl":"/review?paper=review-neurips-2026-154094","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154096","title":"Leveraging Soft Prompts for Privacy Attacks in Federated Prompt Tuning","authors":["Quan M Nguyen","Min-Seon Kim","Hoang M Ngo","Nghia Hoang","HYUK-YOON KWON","My T. Thai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Leveraging Soft Prompts for Privacy Attacks in Federated Prompt Tuning","plainExplanation":"The sampled cosine can exceed one, making the square-root construction nonreal. If even the lower endpoint exceeds one, the required strict similarity margin is impossible.","primaryUrl":"https://arxiv.org/html/2601.06641","noteUrl":"/review?paper=review-neurips-2026-154096","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","missing_assumption","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154103","title":"OBJECT-UNI: A Unified Model for Object-Centric Spatial Understanding and Controllable Generation","authors":["Mining Tan","yinuo Wang","Ziqi Zhou","Weize Quan","Sifei Li","Jingdong Chen","DanDan Zheng","libin wang","Weiming Dong"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"OBJECT-UNI: A Unified Model for Object-Centric Spatial Understanding and Controllable Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.22757","noteUrl":"/review?paper=review-neurips-2026-154103","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154104","title":"Scaling Generative Foundation Models for Chest Radiography with Rectified Flow Transformers","authors":["Fabio De Sousa Ribeiro","Emma A.M. Stanley","Charles Jones","Tian Xia","Dominic C Marshall","Laurent R Triché","Christopher V Cosgriff","Panagiotis Dimitrakopoulos","Sotirios Tsaftaris","Ben Glocker"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Inverting a one-hot endpoint is not exact discrete abduction","plainExplanation":"The generative SCM obtains a categorical observation by taking the argmax of a continuous flow endpoint, but abduction inverts the flow at the observed one-hot category. Argmax is many-to-one.","primaryUrl":"https://arxiv.org/html/2606.19460","noteUrl":"/review?paper=review-neurips-2026-154104","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154105","title":"PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models","authors":["Bo Peng","Tao Huang","Weijie Kong","Junzhe Li","Yue Wu","Qi Tian","Jiangfeng Xiong","Jian-Wei Zhang","Liefeng Bo","Zhao Zhong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23522","noteUrl":"/review?paper=review-neurips-2026-154105","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154106","title":"Bernini: Latent Semantic Planning for Video Diffusion","authors":["Chenchen Liu","Junyi Chen","Lei Li","Lu Chi","Mingzhen Sun","Zhuoying Li","Yi Fu","Ruoyu Guo","Yiheng Wu","Ge Bai","Zehuan Yuan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Bernini: Latent Semantic Planning for Video Diffusion","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22344","noteUrl":"/review?paper=review-neurips-2026-154106","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154107","title":"RuleSmith: Multi-Agent LLMs for Automated Game Balancing","authors":["Ziyao Zeng","Hao Wang","Chen Liu","Youheng Yao","Jingcheng Ni","Tianyu Liu","Xiatao Sun","Fengyu Yang","Chenyu You","Xiaofeng Liu","Daniel Rakita","Ronald Coifman","Yuval Kluger","Zhiwen Fan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RuleSmith: Multi-Agent LLMs for Automated Game Balancing","plainExplanation":"No issue was found in the checked claims. The current evaluation protocol no longer asserts the false implication from the loss threshold to the two win-rate bounds.","primaryUrl":"https://arxiv.org/html/2602.06232","noteUrl":"/review?paper=review-neurips-2026-154107","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154109","title":"Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning","authors":["Yinghui He","Ling Yang","Jiarui Liu","Yongjin Yang","Lechen Zhang","Mengdi Wang","Sanjeev Arora"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The iterated cubic has three real roots, whose sum is not eighteen","plainExplanation":"The iterated cubic has three real roots, whose sum is not eighteen. A scalar skill-entropy reward cannot identify the correct skill sequence.","primaryUrl":"https://arxiv.org/html/2608.05139","noteUrl":"/review?paper=review-neurips-2026-154109","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154110","title":"Local Sparsity Enables Unsupervised LLM Safety Detection","authors":["Xin Chen, Cynthia","Gil Kur","Aleksandr Shevchenko","Andreas Krause"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Local Sparsity Enables Unsupervised LLM Safety Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.20129","noteUrl":"/review?paper=review-neurips-2026-154110","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154111","title":"Integrating Background Knowledge for Scalable Causal Discovery","authors":["Mátyás Schubert","Theofanis Aslanidis","Tom Claassen","Sara Magliacane"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The shrink test conditions on the variable it tests","plainExplanation":"The printed Grow–Shrink procedure removes \\(V\\) when \\(T\\perp V\\mid MB(T)\\), even though \\(V\\in MB(T)\\). Conditional independence \\(T\\perp V\\mid V\\) is automatic.","primaryUrl":"https://arxiv.org/html/2607.10456","noteUrl":"/review?paper=review-neurips-2026-154111","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154113","title":"Agent MechSuits : Mechanistic Subspace Safety Steering for Multi-Turn CLI Agents","authors":["Weidi Luo","Qiming Zhang","Mingyu Jin","Jie Cai","Yihao Quan","Chaowei Xiao","Jingcheng Niu","Zhen Xiang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Agent MechSuits : Mechanistic Subspace Safety Steering for Multi-Turn CLI Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.22673","noteUrl":"/review?paper=review-neurips-2026-154113","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154116","title":"Learning a Maximum Entropy Model for Visual Textures using Diffusion","authors":["Xinyuan Zhao","Eero Simoncelli"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Learning a Maximum Entropy Model for Visual Textures using Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.17342","noteUrl":"/review?paper=review-neurips-2026-154116","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154117","title":"Learning to Price with Persuasion","authors":["Maria-Florina Balcan","Tejas Pagare","Karan Singh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Moment recovery needs the support locations to be known","plainExplanation":"The persuasion reconstruction claims to identify an atomic value distribution from the signaling moments without explicitly supplying its support locations. With one signaling order, the likelihood \\(v^0=1\\) reveals only the mean.","primaryUrl":"https://arxiv.org/html/2608.16699","noteUrl":"/review?paper=review-neurips-2026-154117","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154119","title":"TIER: Trajectory-Invariant Execution Rewards for Multi-Step Tool Composition","authors":["Anay Kulkarni","Allen Lu","Dheeraj Mekala","Jayanth Srinivasa","Gaowen Liu","Jingbo Shang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"TIER: Trajectory-Invariant Execution Rewards for Multi-Step Tool Composition","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.16790","noteUrl":"/review?paper=review-neurips-2026-154119","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154120","title":"Neural Preconditioned Born Series: A Metric-Matched Framework for Learning-based Preconditioners","authors":["Juntao Wang","Jiwei Jia","Xinliang Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neural Preconditioned Born Series: A Metric-Matched Framework for Learning-based Preconditioners","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.18527","noteUrl":"/review?paper=review-neurips-2026-154120","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154121","title":"Frequency‑Aware Flow Matching for Continuous and Consistent Robotic Action Generation","authors":["Jianing Guo","Fangzheng Chen","Zihao Mao","WONG L Kenny","Zhenhong Wu","Yu Li","Yishuai Cai","Yuanpei Chen","Yikun Ban","Kai Chen","DOU QI","Yaodong Yang","Xianglong Liu","Huijie Zhao","Simin Li"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Frequency‑Aware Flow Matching for Continuous and Consistent Robotic Action Generation","plainExplanation":"Sampling can move a high-frequency tail component into a retained low-frequency coordinate, changing the optimizer.","primaryUrl":"https://arxiv.org/html/2606.20135","noteUrl":"/review?paper=review-neurips-2026-154121","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154123","title":"Unveiling the Value of Motion for Cinematic Camera Trajectories","authors":["Ziqi Zhou","Yujian Yuan","Laura Sevilla-Lara"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Unveiling the Value of Motion for Cinematic Camera Trajectories","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.38683","noteUrl":"/review?paper=review-neurips-2026-154123","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154124","title":"Hyper Input Convex Neural Networks for Shape Constrained Learning and Optimal Transport","authors":["Shayan Hundrieser","Insung Kong","Johannes Schmidt-Hieber"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Several approximation and transport statements need local corrections","plainExplanation":"The piece count is short by one. Substituting it into the approximation lower bounds makes the one-neuron corollaries false, although Theorem 3.2 for an abstract function with a specified number of pieces is unaffected.","primaryUrl":"https://arxiv.org/html/2604.26942","noteUrl":"/review?paper=review-neurips-2026-154124","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","mathematical_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154125","title":"Optimal Experiments for Partial Causal Effect Identification","authors":["Tobias Maringgele","Jalal Etesami"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Optimal Experiments for Partial Causal Effect Identification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06993","noteUrl":"/review?paper=review-neurips-2026-154125","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154127","title":"AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild","authors":["Baiyu Chen","Zechen Li","Wilson Wongso","Lihuan Li","Xiachong LIN","Hao Xue","Benjamin Tag","Flora Salim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22715","noteUrl":"/review?paper=review-neurips-2026-154127","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154131","title":"Evo-Depth: A Lightweight Depth-Enhanced Vision-Language-Action Model","authors":["Tao Lin","Yuxin Du","Jiting Liu","Nuobei Zhu","YUNHE LI","Yuqian Fu","Yinxinyu Chen","Hongyi Cai","Ye Zewei","Bing Cheng","Kai Ye","Yiran Mao","Yilei Zhong","Mingkang Dong","Junchi Yan","Gen Li","Bo Zhao"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Evo-Depth: A Lightweight Depth-Enhanced Vision-Language-Action Model","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14950","noteUrl":"/review?paper=review-neurips-2026-154131","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154132","title":"Proprio: Latent Self-Scoring and Inference-Time Refinement for Physically Plausible Video Generation","authors":["Mariam Hassan","Kaouther Messaoud","WUYANG LI","Alexandre Alahi"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Proprio: Latent Self-Scoring and Inference-Time Refinement for Physically Plausible Video Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.28230","noteUrl":"/review?paper=review-neurips-2026-154132","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154133","title":"How Do Language Models Understand Tables? A Mechanistic Analysis of Cell Location","authors":["Xuanliang Zhang","Dingzirui Wang","Keyan Xu","Qingfu Zhu","Wanxiang Che"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The patch direction reverses the recovery interpretation","plainExplanation":"The effect score is \\((LD_{\\rm patch}-LD_{\\rm corrupt})/(LD_{\\rm clean}-LD_{\\rm corrupt})\\), while the text describes inserting corrupted activations into a clean run. An unchanged clean run scores \\(1\\), and replacing it completely by the corrupted state scores \\(0\\).","primaryUrl":"https://arxiv.org/html/2602.08548","noteUrl":"/review?paper=review-neurips-2026-154133","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154136","title":"Mesh BDF: Barycentric Dominance Field for 3D Native Mesh Generation","authors":["Gaochao Song","Haohan Weng","Luo Zhang","Zibo Zhao","Shenghua Gao"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Chordless edge triangles need not be mesh faces","plainExplanation":"The mesh extraction identifies triangles from chordless three-cycles among true mesh edges, without a geometric face test. A triangular bipyramid is a convex closed triangular manifold.","primaryUrl":"https://arxiv.org/html/2606.31777","noteUrl":"/review?paper=review-neurips-2026-154136","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154137","title":"Offline Materials Optimization with CliqueFlowmer","authors":["Jakub Grudzien Kuba","Benjamin K Miller","Sergey Levine","Pieter Abbeel"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Offline Materials Optimization with CliqueFlowmer","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.06082","noteUrl":"/review?paper=review-neurips-2026-154137","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154139","title":"Controllable Dynamic 3D Shape Generation via 3D Trajectories and Text","authors":["Jaeyeong Kim","Inès H Kim","Jahyeok Koo","Seungryong Kim"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Fixed anchor capacity does not cover arbitrary trajectories","plainExplanation":"The construction allocates \\(64\\) anchors, includes every supplied trajectory start, and fills the remaining slots with \\(64-K\\) shape samples for \\(K\\) trajectories. The claimed arbitrary trajectory count permits \\(K=65\\), but then the fill count is negative and \\(65\\) mandatory starts cannot fit in \\(64\\) anchors.","primaryUrl":"https://arxiv.org/html/2606.05162","noteUrl":"/review?paper=review-neurips-2026-154139","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154140","title":"Discovering What You Can Control: Interventional Boundary Discovery for Reinforcement Learning","authors":["Jiaxin Liu","Anzhe Cheng","Paul Bogdan"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Welch’s test detects mean shifts rather than every distributional action effect","plainExplanation":"A one-step random walk can have baseline absolute action equally \\(0.25\\) or \\(0.75\\), and intervention absolute action uniform on \\([0,1]\\). The summary distributions differ but both means are \\(0.5\\), with positive variances. Welch’s mean test therefore need not reject with probability tending to one; its rejection probability stays at its null level under the usual calibration. The distributional-sensitivity condition is insufficient for this chosen test. Single-dimension construction avoids multiple-testing complications; empirical performance is not refuted.","primaryUrl":"https://arxiv.org/html/2603.18257","noteUrl":"/review?paper=review-neurips-2026-154140","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154142","title":"HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression","authors":["Yi Niu","Tianyi Xu","Mingming Ma","Xinkun Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.07192","noteUrl":"/review?paper=review-neurips-2026-154142","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154147","title":"Kinetic-Optimal Scheduling with Moment Correction for Metric-Induced Discrete Flow Matching in Zero-Shot Text-to-Speech","authors":["Dong Yang","YIYI CAI","Haoyu Zhang","Yuki Saito","Hiroshi Saruwatari"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Kinetic-Optimal Scheduling with Moment Correction for Metric-Induced Discrete Flow Matching in Zero-Shot Text-to-Speech","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09386","noteUrl":"/review?paper=review-neurips-2026-154147","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154148","title":"Self-Supervised Keyframe Discovery for Horizon-Invariant Behavior Cloning","authors":["Prabin Kumar Rath","Omkar Patil","Nakul Gopalan"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Self-Supervised Keyframe Discovery for Horizon-Invariant Behavior Cloning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10857","noteUrl":"/review?paper=review-neurips-2026-154148","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154151","title":"Dual-Pathway Circuits of Object Hallucination in Vision-Language Models","authors":["Jiaxin Liu","Ding Zhong","Yue Wang","Zhidong Yang","Zhaolu Kang","Guangyuan Dong","Qishi Zhan","Pengcheng Fang","Aofan Liu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Dual-Pathway Circuits of Object Hallucination in Vision-Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13156","noteUrl":"/review?paper=review-neurips-2026-154151","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154153","title":"Consistency Regularised Gradient Flows for Inverse Problems","authors":["Alessio Spagnoletti","Tim Wang","O. Deniz Akyildiz","Marcelo Pereyra"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The flow-map score weight does not integrate to one","plainExplanation":"Proposition 1 calls its score weight a convex average and uses the prefactor \\(\\sqrt\\alpha/(1+\\sqrt\\alpha)\\). For \\(\\alpha_s=e^{-s}\\), integrating \\(\\beta_s/(2\\sqrt{\\alpha_s})\\) gives \\(1/\\sqrt\\alpha-1\\).","primaryUrl":"https://arxiv.org/pdf/2605.07907v1","noteUrl":"/review?paper=review-neurips-2026-154153","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154157","title":"PAC-Bayesian Bounds for Learning Partially Observed Stochastic Linear Time-Invariant State-Space Systems with Inputs and Sub-Gaussian Noise","authors":["Mihaly Petreczky","Mohamad Al Ahdab","John Leth"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The bounded-noise exponential bound is missing a square","plainExplanation":"The bounded-innovation concentration estimate is missing a squared system-gain term. The uniform localization posterior can leave the prior support, while the Gaussian localization radius omits the parameter dimension. Three additional local corrections concern the exponential-moment parameter domain, the strong-convexity factor, and a geometric-series coefficient. The general PAC-Bayesian argument remains applicable to valid posteriors; the specialized guarantees require the corrected constants and support conditions.","primaryUrl":"https://arxiv.org/html/2609.08740","noteUrl":"/review?paper=review-neurips-2026-154157","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154158","title":"Heterogeneous Parallelism for Multimodal Large Language Model Training","authors":["Yashaswi Karnati","Kamran J Sadeghi","Akash Mehra","Li Ding","Ali Roshan Ghias","Pranav Prashant Thombre","Shifang Xu","Parth Mannan","Yu Yao","Hao Wu","Eric Harper","Ashwath Aithal","Nima Tajbakhsh"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Heterogeneous Parallelism for Multimodal Large Language Model Training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.27678","noteUrl":"/review?paper=review-neurips-2026-154158","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154160","title":"ArcMark: Distortion-Free Multi-Byte LLM Watermark via Optimal Transport","authors":["Atefeh Gilani","Sajani Vithana","Carol X Long","Oliver Kosut","Lalitha Sankar","Flavio Calmon"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ArcMark: Distortion-Free Multi-Byte LLM Watermark via Optimal Transport","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.07235","noteUrl":"/review?paper=review-neurips-2026-154160","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154161","title":"Backbone-Equated Diffusion OOD via Sparse Internal Snapshots","authors":["Yadang Alexis Rouzoumka","Jean Pinsolle","Eugénie TERREAUX","Christèle Morisseau","Jean-Philippe Ovarlez","Chengfang Ren"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Small remainder energy does not control a growing Jacobian cross term","plainExplanation":"The local expansion writes a noisy feature displacement as \\(bJ_\\lambda\\varepsilon+r_\\lambda\\) and infers a second-moment expansion from \\(\\mathbb E\\|r_\\lambda\\|^2=o(b^2)\\). Take \\(\\phi_\\lambda(x)=(x+x^3)/b\\), base point zero and standard Gaussian \\(\\varepsilon\\).","primaryUrl":"https://arxiv.org/html/2605.11014","noteUrl":"/review?paper=review-neurips-2026-154161","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154163","title":"Distilling LLM Feedback for Lean Theorem Proving","authors":["Gaëtan Narozniak","Gérard Biau","Remi Munos","Ahmad Rammal","Pierre Marion"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Distilling LLM Feedback for Lean Theorem Proving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30861","noteUrl":"/review?paper=review-neurips-2026-154163","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154166","title":"Neural Neural Scaling Laws","authors":["Michael Hu","Jane Pan","Ayush Rajesh Jhaveri","Nicholas Lourie","Kyunghyun Cho"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neural Neural Scaling Laws","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.19831","noteUrl":"/review?paper=review-neurips-2026-154166","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154172","title":"ParetoSlider: Diffusion Models Post-Training for Continuous Reward Control","authors":["Shelly Golan","Michael Finkelson","Ariel Bereslavsky","Yotam Nitzan","Or Patashnik"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"ParetoSlider: Diffusion Models Post-Training for Continuous Reward Control","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.20816","noteUrl":"/review?paper=review-neurips-2026-154172","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154174","title":"Prompt–Activation Duality: Improving Activation Steering via Attention-Level Interventions","authors":["Diancheng Kang","Zheyuan (Frank) Liu","Ningshan Ma","Yue Huang","Zhaoxuan Tan","Meng Jiang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Prompt–Activation Duality: Improving Activation Steering via Attention-Level Interventions","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10664","noteUrl":"/review?paper=review-neurips-2026-154174","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154176","title":"RoPEMover: Depth-Aware Object Relocation via Positional Embeddings","authors":["Ipek Oztas","Duygu Ceylan","Aybars B Aksoy","Aysegul Dundar"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The depth warp omits the inverse scale","plainExplanation":"The target mask is formed by scaling the source object about center \\(c\\) by \\(s\\) and translating by \\(d\\), but the depth lookup only subtracts the translation. The forward map is \\(x_{\\rm tgt}=c+s(x_{\\rm src}-c)+d\\), so its inverse is \\(x_{\\rm src}=c+(x_{\\rm tgt}-d-c)/s\\).","primaryUrl":"https://arxiv.org/html/2606.27332","noteUrl":"/review?paper=review-neurips-2026-154176","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154179","title":"Tempered Guided Diffusion","authors":["Andreas Makris","Paul Fearnhead","Christopher Nemeth"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Tempered Guided Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.03712","noteUrl":"/review?paper=review-neurips-2026-154179","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154180","title":"CoLVR: Enhancing Exploratory Latent Visual Reasoning via Contrastive Optimization","authors":["Ziyang Ding","Linjian Meng","Yiming Wu","YuHan Li","Yuhao Liu","Zhen Zhao"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"CoLVR: Enhancing Exploratory Latent Visual Reasoning via Contrastive Optimization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08802","noteUrl":"/review?paper=review-neurips-2026-154180","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154181","title":"Preference-Guided Adaptation for Open-Vocabulary Semantic Segmentation via Prompt Disagreement","authors":["Hyun-Kurl Jang","Jihun Kim","Kuk-Jin Yoon"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Constant partition offsets still change the DPO derivative","plainExplanation":"The exact preference likelihood contains \\(-\\log\\sigma(z+\\Delta)\\), where \\(\\Delta\\) is a partition-function offset. The derivation drops it on the ground that it is constant with respect to parameters. Even if \\(\\partial_\\theta\\Delta=0\\), differentiation gives \\(-\\sigma(-z-\\Delta)\\partial_\\theta z\\).","primaryUrl":"https://arxiv.org/html/2609.34528","noteUrl":"/review?paper=review-neurips-2026-154181","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154185","title":"On the Tightness and Computational Tractability of Higher-Dimensional Confidence Sequences","authors":["Fabian Denoodt","Sibylle Hess","Joaquin Vanschoren","Christian Andersson Naesseth"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stopping and ratio interpretations need geometric and boundary conditions","plainExplanation":"A small volume can be caused by many narrow coordinates while one coordinate still permits a substantial effect. Anytime coverage does not supply the missing geometric implication.","primaryUrl":"https://arxiv.org/html/2610.03727","noteUrl":"/review?paper=review-neurips-2026-154185","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154186","title":"Don't Lose Focus: Activation Steering via Key-Orthogonal Projections","authors":["Haoyan Luo","Mateo Espinosa Zarlenga","Mateja Jamnik"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Don't Lose Focus: Activation Steering via Key-Orthogonal Projections","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06342","noteUrl":"/review?paper=review-neurips-2026-154186","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154189","title":"The Price of Locality: Why Forward-Forward Underperforms Backpropagation?","authors":["Zhaoxian Wu","Haichuan Liu","Tianyi Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Small Jacobian variation does not preserve exact rank","plainExplanation":"Sample-dependent Jacobians need not preserve exact rank or the claimed effective-rank factor. The appendix already gives valid deviation-dependent alternatives. Linear backpropagation capacity growth also requires positive excess output rank, while the Forward-Forward capacity argument needs a nonzero limiting goodness-gradient energy. In the optimization proof, the invariant-region argument drops the input drift tracked earlier. These conclusions concern the stated mathematical guarantees and proof steps.","primaryUrl":"https://arxiv.org/html/2609.33240","noteUrl":"/review?paper=review-neurips-2026-154189","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154191","title":"AVOC: Enhancing Hour-Level Audio-Video Understanding in Omni-Modal LLMs via Retrieval-Inspired Token Compression","authors":["Yijing Chen","Wenhui Tan","Xiaoyi Yu","Yuyue Wang","Xin Cheng","Kaisi Guan","Hao Jiang","Xiangyang Li","Guojie Zhu","Ruihua Song"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"AVOC: Enhancing Hour-Level Audio-Video Understanding in Omni-Modal LLMs via Retrieval-Inspired Token Compression","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.24286","noteUrl":"/review?paper=review-neurips-2026-154191","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154194","title":"Rethinking Cross-Layer Information Routing in Diffusion Transformer","authors":["Chao Xu","Maohua Li","Qirui Li","Yixuan Xu","Yanke Zhou","Yunhe LI","Cuifeng Shen","Hanlin Tang","Kan Liu","Tao Lan","Lin Qu","Shao-Qun Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The cost interpretation and probe conclusion exceed their supporting formulas","plainExplanation":"The architecture retains all earlier raw sources at both endpoints \\(S=1\\) and \\(S=L\\), but the surrogate charges the latter an extra \\(\\alpha\\log L\\) of compression distortion. The formula therefore does not model actual discarded sources over its stated full domain.","primaryUrl":"https://arxiv.org/abs/2605.20708v4","noteUrl":"/review?paper=review-neurips-2026-154194","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["missing_assumption"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-154196","title":"ESSAM: A Novel Competitive Evolution Strategies Approach to Reinforcement Learning for Memory Efficient LLMs Fine-Tuning","authors":["Zhishen Sun","Sizhe Dang","Guang Dai","Haishan Ye"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The centered Gaussian estimator has a different finite-sample second moment","plainExplanation":"The centered Gaussian estimator has a different finite-sample second moment. Collinearity of two-sample updates does not imply a fixed normalization ratio. The nonlinear Taylor error need not decay with the sample count. The printed perturbation update has an extra sample-count division.","primaryUrl":"https://arxiv.org/html/2602.01003","noteUrl":"/review?paper=review-neurips-2026-154196","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154197","title":"$R^3$: 3D Reconstruction via Relative Regression","authors":["Congrong Xu","Huachen Gao","Xingyu Chen","Yuliang Xiu","Jun Gao","Anpei Chen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"$R^3$: 3D Reconstruction via Relative Regression","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26519","noteUrl":"/review?paper=review-neurips-2026-154197","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154201","title":"Attending on Attention ($A^2$): Smaller Self-Supervised ViTs Localize Better Than Larger Ones","authors":["Sreehari Rammohan","Huy Ha","Carl Vondrick"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Attending on Attention ($A^2$): Smaller Self-Supervised ViTs Localize Better Than Larger Ones","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.03148","noteUrl":"/review?paper=review-neurips-2026-154201","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154203","title":"Information Discernment in Large Language Models","authors":["Joshua Ashkinaze","Laura Kurek","Alina Faisal","Tongyuan Miao","Mariam Joseph","Ceren Budak","Eric Gilbert"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The oracle correlation differs from the benchmark discernment metric","plainExplanation":"The benchmark correlates belief-change magnitude \\(|P_2-P_1|\\) with helpfulness \\(I\\), while the oracle simulation labels \\(\\rho(\\alpha,I)\\) as the same discernment quantity under \\(P_2=P_1+\\alpha(C-P_1)\\). The measured change is \\(|\\alpha|\\,|C-P_1|\\).","primaryUrl":"https://arxiv.org/html/2607.19355","noteUrl":"/review?paper=review-neurips-2026-154203","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154205","title":"A Unified Audio Language Model with Text-Aligned Factorized Audio Tokenization","authors":["Dongchao Yang","Yuanyuan Wang","Songxiang Liu","Dading Chong","Xixin Wu","Helen Meng"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The appendix objective omits earlier streams in the same frame","plainExplanation":"Equation (9) conditions stream \\(k\\) on earlier streams of the current frame; equation (16) conditions only on prior frames and text. Let a fair first bit equal the second bit. Its second-stream conditional entropy given the first is zero, but the marginal entropy used by the appendix is \\(\\log2\\). The objectives therefore have different optimal values even with the stated weights. This is a factorization/objective inconsistency. An unreported shorthand or code convention is not inferred.","primaryUrl":"https://arxiv.org/html/2602.04683","noteUrl":"/review?paper=review-neurips-2026-154205","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154208","title":"Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use","authors":["Yize Cheng","Chenrui Fan","Mahdi JafariRaviz","Keivan Rezaei","Soheil Feizi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Probe orthogonality does not establish cognition-action decoupling","plainExplanation":"Euclidean orthogonality does not make probe scores uncorrelated, statistically independent or causally uncoupled.","primaryUrl":"https://arxiv.org/html/2605.14038","noteUrl":"/review?paper=review-neurips-2026-154208","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154209","title":"Jacobian Descent for Multi-Objective Optimization","authors":["Pierre Quinton","Valérian Rey"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Jacobian Descent for Multi-Objective Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2406.16232","noteUrl":"/review?paper=review-neurips-2026-154209","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154210","title":"Quantum Speedups for Stochastic Optimization with Heavy-Tailed Noise","authors":["Bin Luo","Chengchang Liu","Jonathan Allcock","Shengyu Zhang","John C. S. Lui"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The high-probability guarantee leaves a bad-event expectation term","plainExplanation":"Splitting the expectation gives a good-event contribution plus \\(\\delta\\mathbb E[\\|\\nabla f\\|\\mid\\mathrm{bad}]\\). The latter is not zero merely because its event is rare. Under smoothness and normalized finite-step updates, a deterministic trajectory bound such as \\(\\sqrt{2L\\Delta}+L\\eta(T-1)\\) can control it, but the proof does not choose confidence and accuracy slack to do so. This is an implication gap, not a counterexample to the algorithm or its possible logarithmically adjusted query rate.","primaryUrl":"https://arxiv.org/html/2607.25492","noteUrl":"/review?paper=review-neurips-2026-154210","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154213","title":"Recovery Guarantees for Posterior Sampling of One-Bit Compressed Sensing","authors":["jing ma","Yujia Wu","Zhaoqiang Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The lower-bound proof uses a false covering-number amplification","plainExplanation":"The conditioned distribution can concentrate on the removed heavy atom, even when covering the original distribution to higher mass requires many tiny atoms.","primaryUrl":"https://arxiv.org/html/2610.11834","noteUrl":"/review?paper=review-neurips-2026-154213","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154214","title":"Distributionally Robust Listwise Preference Optimization","authors":["Xudong Wu","Jian Qian","Pangpang Liu","Vaneet Aggarwal","Jiayu Chen"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Distributionally Robust Listwise Preference Optimization","plainExplanation":"Differentiating the expectation also differentiates the unscaled sampling law. Those score-function terms do not all acquire two temperature factors.","primaryUrl":"https://arxiv.org/html/2607.01715","noteUrl":"/review?paper=review-neurips-2026-154214","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154217","title":"Resource-Aware Parameter-Efficient Model Adaptation for Onboard High-Dimensional Data","authors":["Qiyang Zhang","Xinhao Li","Lei Shi","Zheng Lin","Jinfeng Wen","Ao Zhou","Shangguang Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"GELU on constant factors does not enlarge the LoRA function class","plainExplanation":"The auxiliary GELU branches act on trainable parameter matrices, rather than on the input, and are added to the two LoRA factors. Define \\(A_{\\rm eff}=A+\\operatorname{GELU}(A_{\\rm aux})\\) and \\(B_{\\rm eff}=B+\\operatorname{GELU}(B_{\\rm aux})\\).","primaryUrl":"https://arxiv.org/html/2609.33687","noteUrl":"/review?paper=review-neurips-2026-154217","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154223","title":"Look-Before-Move: Narrative-Grounded World Visual Attention in Dynamic 3D Story Worlds","authors":["Jiaming Bian","Bingliang Li","Yuehao Wu","Pichao WANG","Zhi Wang","Hailan Ma","Huadong Mo","Zhenhong Sun"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Look-Before-Move: Narrative-Grounded World Visual Attention in Dynamic 3D Story Worlds","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.26964","noteUrl":"/review?paper=review-neurips-2026-154223","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154225","title":"Autonomous Continual Learning for Environment Adaptation of Computer-Use Agents","authors":["Tianci Xue","Zeyi Liao","Tianneng Shi","Zilu Wang","Kai Zhang","Dawn Song","Yu Su","Huan Sun"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Autonomous Continual Learning for Environment Adaptation of Computer-Use Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.10356","noteUrl":"/review?paper=review-neurips-2026-154225","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154226","title":"Finding Interpretable Prompt-Specific Circuits in Language Models","authors":["Gabriel Franco","Lucas M Tassis","Azalea Rohr","Mark Crovella"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Finding Interpretable Prompt-Specific Circuits in Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.13483","noteUrl":"/review?paper=review-neurips-2026-154226","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154227","title":"You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories","authors":["Zhepei Wei","Xinyu Zhu","Wei-Lin Chen","Chengsong Huang","Jiaxin Huang","Yu Meng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.21468","noteUrl":"/review?paper=review-neurips-2026-154227","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154228","title":"WASD: Wasserstein-based Knowledge Distillation for Large Language Models","authors":["Byeonghu Na","Donghyeok Shin","Yeongmin Kim","Mina Kang","Il-chul Moon"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"WASD: Wasserstein-based Knowledge Distillation for Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07706","noteUrl":"/review?paper=review-neurips-2026-154228","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154229","title":"Approximation in Contrastive Representation Learning","authors":["Yuanfan Li","Zihan Zhang","Yiming Ying","Ding-Xuan Zhou"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Subtracting log priors can change the asserted function span","plainExplanation":"An upper density-ratio bound does not provide the lower overlap used in the proof. The indicator-basis operator includes square roots of the class priors. Coordinate bounds do not ensure that the spherical augmentation is real. Subtracting log priors can change the asserted function span. The joint-support excess-loss identity omits negative pairs outside that support.","primaryUrl":"https://arxiv.org/html/2610.15388","noteUrl":"/review?paper=review-neurips-2026-154229","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154230","title":"FrameVGGT: Coherence-Preserving Memory for Bounded Streaming Geometry","authors":["Zhisong Xu","Takeshi Oishi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Positive mean alignment does not order the full query score gap","plainExplanation":"Scores depend on the whole query, including its orthogonal component. Fix unit query \\(q=(0.9,\\sqrt{0.19})\\), reference key \\(R=(1,0)\\), and alternative key \\(N=(0.6,0.8)\\) or \\((0.8,-0.6)\\). The former has the larger projected concentration gap along the reference direction but the smaller positive attention-score gap, hence larger two-key softmax entropy. Each group is a singleton, so within-group ties impose no extra restriction. The diagnostic-to-gap inference fails under merely positive alignment. The entropy response to a genuinely increasing positive score gap is not challenged.","primaryUrl":"https://arxiv.org/html/2603.07690","noteUrl":"/review?paper=review-neurips-2026-154230","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154231","title":"Neuro-Inspired Inverse Learning for Planning and Control","authors":["Maryna Kapitonova","Tonio Ball"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neuro-Inspired Inverse Learning for Planning and Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24152","noteUrl":"/review?paper=review-neurips-2026-154231","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154234","title":"SceneAligner: 3D-Grounded Floorplan Localization in the Wild","authors":["Junhyeong Cho","Ruojin Cai","Hadar Averbuch-Elor"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SceneAligner: 3D-Grounded Floorplan Localization in the Wild","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22581","noteUrl":"/review?paper=review-neurips-2026-154234","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154236","title":"Slower Generalization, Faster Memorization: A Sweet Spot in Algorithmic Learning","authors":["Shin So","Kyelim Lee","Albert No"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Slower Generalization, Faster Memorization: A Sweet Spot in Algorithmic Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.14659","noteUrl":"/review?paper=review-neurips-2026-154236","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154237","title":"Lifting Biomolecular Data Acquisition","authors":["Eli N. Weinstein","Andrei Slabodkin","Mattia G Gollub","Kerry Dobbs","Xiao-Bing Cui","Fang zhang","Kristina Gurung","Elizabeth B Wood"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The Poisson fourth-moment prefactor has the wrong dimension term","plainExplanation":"Equation 36 raises a single-molecule zero probability to the molecule count, while the data model specifies a negative-binomial count with fixed shape at the whole-cell mean. A fixed-shape negative-binomial zero probability is \\(z(\\mu)=(\\phi/(\\phi+\\mu))^\\phi\\).","primaryUrl":"https://arxiv.org/html/2512.15984","noteUrl":"/review?paper=review-neurips-2026-154237","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154238","title":"FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation","authors":["Yuanzhi Wang","Xuhua Ren","Jiaxiang Cheng","bing ma","Kai Yu","Sen Liang","Wenyue Li","Tianxiang Zheng","Qinglin Lu","Zhen Cui"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Shared view information need not be facial identity information","plainExplanation":"The view-information bound does not specify which shared factor the representations encode.","primaryUrl":"https://arxiv.org/html/2605.04702","noteUrl":"/review?paper=review-neurips-2026-154238","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154239","title":"Policy Gradient Algorithms in Average-Reward Multichain MDPs","authors":["Jongmin Lee","Ernest Ryu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Policy Gradient Algorithms in Average-Reward Multichain MDPs","plainExplanation":"A zero recurrent target-time factor cannot account for estimating a nontrivial transient absorption probability.","primaryUrl":"https://arxiv.org/html/2602.18003","noteUrl":"/review?paper=review-neurips-2026-154239","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154240","title":"Generative Modeling via Drifting","authors":["Mingyang Deng","He Li","Tianhong Li","Yilun Du","Kaiming He"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Generative Modeling via Drifting","plainExplanation":"A probability model cannot equal an extrapolated target that assigns negative mass to a measurable set.","primaryUrl":"https://arxiv.org/html/2602.04770","noteUrl":"/review?paper=review-neurips-2026-154240","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154242","title":"GAE Falls Short in Imperfect-Information Self-Play Reinforcement Learning","authors":["Zhiyuan Fan","Gabriele Farina"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"GAE Falls Short in Imperfect-Information Self-Play Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19235","noteUrl":"/review?paper=review-neurips-2026-154242","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154243","title":"Anytime-Valid Conformal Risk Control","authors":["Bror Hultberg","Dave Zachariah","Antonio Ribeiro"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Anytime-Valid Conformal Risk Control","plainExplanation":"The logarithmic term alone is too small, even when there is no distribution shift and all weights equal one.","primaryUrl":"https://arxiv.org/html/2602.04364","noteUrl":"/review?paper=review-neurips-2026-154243","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","mathematical_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154245","title":"MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory","authors":["Jiaqian Wang","Shengtao Zhang","Ruiwen Zhou","Junwei Liao","Yuchen Feng","Zhuo Li","Yujie Zheng","Weinan Zhang","Ying Wen","Zhiyu li","Feiyu Xiong","Yutao Qi","Bo Tang","Muning Wen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A constant-step reward average does not converge to a deterministic utility","plainExplanation":"An MSE utility update is not an ascent step for the proposed linear surrogate. A constant-step reward average does not converge to a deterministic utility.","primaryUrl":"https://arxiv.org/html/2601.03192","noteUrl":"/review?paper=review-neurips-2026-154245","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154247","title":"FoMEMO: Towards Foundation Models for Expensive Multi-objective Optimization","authors":["Yiming Yao","Fei Liu","Liang Zhao","Xi Lin","Yilu Liu","Qingfu Zhang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"FoMEMO: Towards Foundation Models for Expensive Multi-objective Optimization","plainExplanation":"The radial complement extends beyond the finite reference box near coordinate-axis directions, and its stated moment can diverge.","primaryUrl":"https://arxiv.org/html/2509.03244","noteUrl":"/review?paper=review-neurips-2026-154247","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154251","title":"From Matrix Inversion to Constraints: Provably Tighter Confidence Regions for Importance Weights in Label Shift","authors":["Mushan Li","Kihyun Han","Yanyuan Ma"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Switching to a one-sided boundary interval after observing the count can under-cover","plainExplanation":"A finite random search over a discontinuous calibration quantile need not enclose every feasible weight. Switching to a one-sided boundary interval after observing the count can under-cover.","primaryUrl":"https://arxiv.org/html/2609.14802","noteUrl":"/review?paper=review-neurips-2026-154251","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154252","title":"Very Fast Bayesian Additive Regression Trees on GPU","authors":["Giacomo Petrillo"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Very Fast Bayesian Additive Regression Trees on GPU","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2410.23244","noteUrl":"/review?paper=review-neurips-2026-154252","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154253","title":"SP$^3$: Spherical Priors for Plug-and-Play Restoration","authors":["Sean Man","Ron Raphaeli","Matan Kleiner","Or Ronai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SP$^3$: Spherical Priors for Plug-and-Play Restoration","plainExplanation":"Exact block minimization solves the penalized problem, which need not enforce the original constraint or its stationarity conditions.","primaryUrl":"https://arxiv.org/html/2606.16396","noteUrl":"/review?paper=review-neurips-2026-154253","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154254","title":"From History to State: Constant-Context Skill Learning for LLM Agents","authors":["Haoyang Xie","Xinyuan Wang","Yancheng Wang","Puda Zhao","Feng Ju"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"From History to State: Constant-Context Skill Learning for LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05413","noteUrl":"/review?paper=review-neurips-2026-154254","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154255","title":"Corrected Integrated Laplace Approximation for Bayesian Inference in Latent Gaussian Models","authors":["Jinlin Lai","Charles Margossian","Daniel Sheldon"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The Laplace proposal need not give finite importance-weight variance","plainExplanation":"The Newton loop uses the stopping inequality in the wrong direction. The Laplace-covariance derivative omits the dependence of the mode on the kernel. The Laplace proposal need not give finite importance-weight variance.","primaryUrl":"https://arxiv.org/html/2605.20345","noteUrl":"/review?paper=review-neurips-2026-154255","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154257","title":"Retrieval from Within: An Intrinsic Capability of Attention-Based Models","authors":["Elad Hoffer","Yochai Blau","Edan kinderman","Ron Banner","Daniel Soudry","Boris Ginsburg"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Retrieval from Within: An Intrinsic Capability of Attention-Based Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.05806","noteUrl":"/review?paper=review-neurips-2026-154257","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154258","title":"LaST-R1: Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning","authors":["Hao Chen","Zhonghao Yan","Jiaming Liu","Nuowei Han","Renrui Zhang","Chenyang Gu","Jialin Gao","Peng Jia","Shanghang Zhang","Pheng-Ann Heng"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"LaST-R1: Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.28192","noteUrl":"/review?paper=review-neurips-2026-154258","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154259","title":"Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models","authors":["Zeyang Ding","Xinglin Hu","Zhiyu Chen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The Wasserstein affinity is not guaranteed to have a positive spectrum","plainExplanation":"Gaussian kernelization of these measure distances need not yield a PSD matrix. Consequently the claimed lower bound $K\\succeq\\alpha I$ and real fractional powers of its eigenvalues are not guaranteed.","primaryUrl":"https://arxiv.org/html/2603.22303","noteUrl":"/review?paper=review-neurips-2026-154259","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154260","title":"How Deep Are Deep GPs, Really? A Sharp Threshold and a Non-Gaussian Limit for Compositional GPs","authors":["Mark Kozdoba","Shie Mannor"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Collision configurations provide additional invariant laws","plainExplanation":"Squared distances do not have a density on the full positive orthant. Collision configurations provide additional invariant laws.","primaryUrl":"https://arxiv.org/html/2606.08218","noteUrl":"/review?paper=review-neurips-2026-154260","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154262","title":"Achieving Optimal Static and Dynamic Regret Simultaneously in Bandits with Deterministic Losses","authors":["Jian Qian","Chen-Yu Wei"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Achieving Optimal Static and Dynamic Regret Simultaneously in Bandits with Deterministic Losses","plainExplanation":"The threshold supplies only three quarters of the epochs, and Case 2 epochs can contribute \\(-1/2\\). The displayed constant does not follow from Equation (18).","primaryUrl":"https://arxiv.org/html/2602.07418","noteUrl":"/review?paper=review-neurips-2026-154262","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["incorrect_bound"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-154264","title":"Rebalancing Reference Frame Dominance to Improve Motion in Image-to-Video Models","authors":["Wooseok Jeon","Seungho Park","Seunghyun Shin","Sangeyl Lee","Hyeonho Jeong","Hae-Gon Jeon"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The inference loop confuses elapsed progress with model time","plainExplanation":"Denoising requires evaluating the field along descending model time. An increasing elapsed-progress variable may schedule attention changes, but passing it directly as the model timestamp evaluates a different field. An exact affine Gaussian conditional flow already yields the wrong final mean under this reset clock, even if the scheduler applies a negative step. The printed inference loop has a time-convention mismatch. The actual implementation and empirical results are not independently refuted.","primaryUrl":"https://arxiv.org/html/2605.19398","noteUrl":"/review?paper=review-neurips-2026-154264","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154265","title":"E0: Expressive Fine-Grained Discrete Action Prediction for Vision-Language-Action Models via Tweedie Discrete Diffusion","authors":["Zhihao Zhan","Jiaying Zhou","Likui Zhang","Qinhan Lyu","Hao Liu","Weizheng Li","jusheng zhang","Ziliang Chen","Tianshui Chen","Ruifeng zhai","Keze Wang","Liang Lin","Guangrun Wang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"A conditional mean of several modes need not lie in a low-density gap","plainExplanation":"A conditional mean of several modes need not lie in a low-density gap. The reverse-KL bound needs control of the model probabilities as well as the truth. Two local score identities disagree with the correct boxed Tweedie formula. An arbitrary classifier does not necessarily induce the score of a density. Cross-entropy is not invariant to radial scaling of the logits. Stepwise argmax can collapse an exact posterior sampler to the majority class. Vocabulary-independent sequence length does not make the model cost vocabulary-independent.","primaryUrl":"https://arxiv.org/html/2511.21542v2","noteUrl":"/review?paper=review-neurips-2026-154265","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154266","title":"NIV: Neural Axis Variations for Variable Font Generation","authors":["Nadav Benedek","Ariel Shamir","Ohad Fried"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The tent boundary branch overrides a permitted peak","plainExplanation":"The tuple weight allows \\(\\ell\\leq p\\leq r\\), including default supports whose peak equals an endpoint, but its first branch returns zero at \\(v\\leq\\ell\\) or \\(v\\geq r\\). At an allowed peak \\(p=r\\), the first branch gives \\(w(p)=0\\), although the adjacent increasing branch approaches \\(1\\).","primaryUrl":"https://arxiv.org/pdf/2606.05261v1","noteUrl":"/review?paper=review-neurips-2026-154266","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154267","title":"Kernelized Activation Steering","authors":["Laziz Abdullaev","Minh-Hieu Pham","Bach Do","Khoat Than","Tan Nguyen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Kernelized Activation Steering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.01062","noteUrl":"/review?paper=review-neurips-2026-154267","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154270","title":"Comparing Transformers and Hybrid Models at the Token Level","authors":["Yanhong Li","Will Merrill"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"An average information bound does not imply the stated pointwise KL bound","plainExplanation":"An average information bound does not imply the stated pointwise KL bound.","primaryUrl":"https://arxiv.org/html/2606.20936","noteUrl":"/review?paper=review-neurips-2026-154270","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-154271","title":"Shifting the Gradient: Understanding How Defensive Training Methods Protect Language Model Integrity","authors":["Satchel Grant","Victor Gillioz","Jake Ward","Tom McGrath"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Concentrated PCA variance is not concentrated trait-direction energy","plainExplanation":"The interpretation treats a dominant first gradient principal component as explaining most of a chosen trait direction, although the reported squared alignment with that component is small. For unit direction \\(u\\), normalized gradient energy is \\(\\sum_ke_k(v_k^\\top u)^2\\).","primaryUrl":"https://arxiv.org/html/2604.16423","noteUrl":"/review?paper=review-neurips-2026-154271","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154272","title":"Coreset-Induced Conditional Velocity Flow Matching","authors":["Xiao Wang","Zihua She","Jianxi Su"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Unconditional assignment columns do not preserve the conditional source law","plainExplanation":"Uniform-index column sums determine \\(P(B=b)\\), but after conditioning, \\(P(B=b\\mid C)=\\sum_iP(I=i\\mid C)P(B=b\\mid I=i,C)\\). The index weights are no longer uniform. In a self-consistent two-point EMS example with data \\(\\{-1,1\\}\\), equal component weights, regularization \\(1.5\\), and anchors \\(\\pm a\\), where \\(a=\\tanh(2a/1.5)\\approx0.775516\\), the printed covariance is \\((1-a^2)/2\\). At \\(t=x=0.5\\), the actual positive-component probability is approximately \\(0.795314\\), versus required Bayes weight \\(0.784704\\); at \\(t=0\\), both are \\(0.5\\). The general-time conditional preservation proof is unsupported. The initial-time identity and other conditional velocity results remain valid within their assumptions.","primaryUrl":"https://arxiv.org/html/2605.12951","noteUrl":"/review?paper=review-neurips-2026-154272","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154274","title":"(Mis)generalization of Helpful-Only Fine-Tuning","authors":["Mohammad Omar Khursheed","Baram Sosis","Fabien Roger"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"(Mis)generalization of Helpful-Only Fine-Tuning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.04413","noteUrl":"/review?paper=review-neurips-2026-154274","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154277","title":"A Unified Approach for Computing Wasserstein Barycenters of Discrete and Continuous Measures","authors":["Peng Xu","Changbo Zhu","Xiaohui Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Finite transport moments do not imply finite comparator KL","plainExplanation":"The barycenter convergence bound uses a finite KL comparator without imposing finite relative entropy, even for compactly supported input laws. For a single input law the barycenter is that law.","primaryUrl":"https://arxiv.org/html/2605.11270","noteUrl":"/review?paper=review-neurips-2026-154277","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154278","title":"LLM Judge Validation Under Sparse Overlap: From Inference to Design","authors":["Junxuan Li","Arko Mukherjee","Soumyabrata Pal"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"A diverging variance upper bound does not establish divergence of the variance","plainExplanation":"Shared-item covariance changes the variance of a model difference. A diverging variance upper bound does not establish divergence of the variance. The delta-method covariance term has the wrong sign.","primaryUrl":"https://arxiv.org/html/2609.31857","noteUrl":"/review?paper=review-neurips-2026-154278","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154279","title":"Towards Complementary Keypoint Detection via Mixture of Detectors","authors":["Xiaoyong Lu","Guobao Xiao","Dong Liang","Songlin Du"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Towards Complementary Keypoint Detection via Mixture of Detectors","plainExplanation":"No issue was found in the checked claims. Section3.2.2 gives a generic top-k positive-reward rule; Section4.1 line313 instantiates it as top-2 for both rewards. These are compatible. Class-index k and top-k cardinality are overloaded notation, but the whole-paper training rule is defined. With six detectors, chosen top2 quantity(1,1,0,0,0,0),precision(0,1,0,1,0,0),alpha.8 yields rewards(.8,1,0,.2,0,0),sum2,normalized(.4,.5,0,.1,0,0).","primaryUrl":"https://arxiv.org/html/2610.16762","noteUrl":"/review?paper=review-neurips-2026-154279","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154281","title":"Concurrent Image Understanding and Generation: Self-Correcting Coupled Markov Jump Process","authors":["Minh-Quan Le","Armand Comas","Alexandros Lattas","Stylianos Moschoglou","Pedro Velez","Amit Raj","Aaron Germuth","Thabo Beeler","Dimitris Samaras","Di Qiu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Nested floor quotas can leave a token unrevealed","plainExplanation":"The reveal schedule recomputes a quota from an idealized remaining count rather than the actual integer number left after previous floors. With \\(L=3\\), no initial reveal and schedule \\(0,1/2,1\\), the first reveal is \\(\\lfloor3/2\\rfloor=1\\).","primaryUrl":"https://arxiv.org/html/2607.13188","noteUrl":"/review?paper=review-neurips-2026-154281","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154286","title":"PermaVid: Consistent Video Generation Across Edits via Disentangled Context Memory","authors":["Shuai Yang","Bingjie Gao","Ziwei Liu","Jiaqi Wang","Dahua Lin","Tong Wu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PermaVid: Consistent Video Generation Across Edits via Disentangled Context Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.16449","noteUrl":"/review?paper=review-neurips-2026-154286","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154287","title":"Data Auctions for Retrieval Augmented Generation","authors":["Minbiao Han","Seyed A Esmaeili","Michael Albert","Haifeng Xu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The virtual-payment identity uses the wrong probability-mass index","plainExplanation":"The virtual-payment identity uses the wrong probability-mass index. The lower adjacent-IC payment choice can have zero revenue against a positive optimum.","primaryUrl":"https://arxiv.org/html/2508.16007","noteUrl":"/review?paper=review-neurips-2026-154287","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154289","title":"Test-time Scaling of Diffusions with Flow Maps","authors":["Amirmojtaba Sabour","Michael Albergo","Carles Domingo i Enrich","Nicholas Boffi","Sanja Fidler","Karsten Kreis","Eric Vanden-Eijnden"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Test-time Scaling of Diffusions with Flow Maps","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.22688","noteUrl":"/review?paper=review-neurips-2026-154289","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154290","title":"StreamPI: Streaming Multimodal Temporal Modeling for Vision-Language-Action Models","authors":["Zhe Liu","Jinghua Hou","Yuxiang Lu","Zhenya YANG","Xianzhe Fan","Junwei Luo","Junyi Li","Ruihua Han","Zhi Hou","Hengshuang Zhao"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"StreamPI: Streaming Multimodal Temporal Modeling for Vision-Language-Action Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.26067","noteUrl":"/review?paper=review-neurips-2026-154290","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154291","title":"Perception Without Engagement: Dissecting the Causal Discovery Deficit in LMMs","authors":["Jiafeng Liang","Zhihao Zhu","Zihan Zhang","Baoqi Ren","Tao Ren","Shixin Jiang","Runxuan Liu","Ming Liu","See-Kiong Ng","Bing Qin"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Correct-label and KL gradients need not be orthogonal","plainExplanation":"The gradient-geometry argument treats the correct-label log-probability direction and the KL direction as orthogonal. In logit coordinates, \\(\\nabla\\log p_0=e_0-p\\) and \\(\\nabla D_{\\rm KL}(p\\|q)=p\\odot(\\log(p/q)-D_{\\rm KL}(p\\|q))\\).","primaryUrl":"https://arxiv.org/html/2605.09422","noteUrl":"/review?paper=review-neurips-2026-154291","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154292","title":"Bidirectional Sparse Attention for Faster Video Diffusion Training","authors":["Chenlu Zhan","Wen Li","Jun Zhang","chuyu shen","Lele Cheng"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Novelty selection can discard the center required for filling","plainExplanation":"Constant raw scores equal to two give cutoff \\(p=2\\), which normalized cumulative attention cannot reach even with every key. Reading the scores as probabilities removes that range failure but not the unit mismatch: constant scores \\(0.25\\) give a one-key minimum even when the requested retained count is two. The two selection meanings are not interchangeable. A post-softmax interpretation alone does not establish the requested count or mass behavior.","primaryUrl":"https://arxiv.org/html/2509.01085","noteUrl":"/review?paper=review-neurips-2026-154292","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154293","title":"Hybrid Probabilistic Zonotopes for Identifiable and Refinable Predictive Uncertainty","authors":["Zhen Zhang","Amr Alanwar"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Shared latent offsets can confound the parameter despite distinct nominal centers","plainExplanation":"Shared latent offsets can confound the parameter despite distinct nominal centers. The first characteristic-function zero need not identify the uniform generator.","primaryUrl":"https://arxiv.org/html/2608.05454","noteUrl":"/review?paper=review-neurips-2026-154293","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154294","title":"GReFEM: Multimodal LLMs as Zero-Shot Semantic Assistants for Physics-Guided 3D Mesh Refinement","authors":["Kartik Bali","Mahish Kumar Guru","Christian J Cyron","Roland Aydin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GReFEM: Multimodal LLMs as Zero-Shot Semantic Assistants for Physics-Guided 3D Mesh Refinement","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.08798","noteUrl":"/review?paper=review-neurips-2026-154294","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154295","title":"$\\alpha$Depth: Learning Single-Pass Soft Boundary Decomposition for Stereo Conversion","authors":["Xiang Zhang","Yang Zhang","Lukas Mehl","Karlis M Briedis","Markus Gross","Christopher Schroers"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"$\\alpha$Depth: Learning Single-Pass Soft Boundary Decomposition for Stereo Conversion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.00386","noteUrl":"/review?paper=review-neurips-2026-154295","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154296","title":"Hypernetworks for Dynamic Feature Selection","authors":["Javier Fumanal Idocin","Raquel Fernandez-Peralta","Javier Andreu-Perez"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Equal coordinate variances do not determine transformed feature energy","plainExplanation":"For linear map \\(A\\), \\(\\mathbb E\\|Ae\\|^2=\\operatorname{tr}(A\\operatorname{Cov}(e)A^\\top)\\) for centered features, including off-diagonal covariance. Present features \\((R,R)\\) and absent features \\((T,-T)\\), with independent unit Rademachers, have equal coordinate variances. With \\(A=(1,1)\\), their transformed energies are four and zero, so the aggregate can scale as \\(4|S|\\). Independence across features does not impose isotropy within one feature. The stated variance assumptions are insufficient. Adding full covariance control removes this witness but does not automatically prove every energy claim.","primaryUrl":"https://arxiv.org/html/2605.12278","noteUrl":"/review?paper=review-neurips-2026-154296","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154299","title":"OptiWorld: Optimal Control for Video World Generation under Physical Constraints","authors":["Yu Yuan","Jianhao Yuan","Xijun Wang","Daiqing Li","Liu He","Lu Ling","Stanley Chan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Cartesian second differences do not discretize covariant acceleration","plainExplanation":"The continuous objective penalizes covariant acceleration under a position-dependent metric, while its proposed discretization squares ordinary coordinate second differences. For \\(q(s)=(0,0,s)\\) and \\(G=\\operatorname{diag}(1+z^2,1+z^2,1+4z^2)\\), Cartesian acceleration is zero but \\(\\nabla_{\\dot q}\\dot q=(0,0,4s/(1+4s^2))\\).","primaryUrl":"https://arxiv.org/html/2606.00499","noteUrl":"/review?paper=review-neurips-2026-154299","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154302","title":"At FullTilt: Real-Time Open-Set 3D Macromolecule Detection Directly from Tilted 2D Projections","authors":["Ming-Yang Ho","Alberto Bartesaghi"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"At FullTilt: Real-Time Open-Set 3D Macromolecule Detection Directly from Tilted 2D Projections","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.10766","noteUrl":"/review?paper=review-neurips-2026-154302","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154303","title":"Déjà View: Looping Transformers for Multi-View 3D Reconstruction","authors":["Alessandro Burzio","Tobias Fischer","Sven Elflein","Qunjie Zhou","Riccardo de Lutio","Jiawei Ren","Jiahui Huang","Shengyu Huang","Marc Pollefeys","Laura Leal-Taixé","Zan Gojcic","Haithem Turki"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Déjà View: Looping Transformers for Multi-View 3D Reconstruction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30215","noteUrl":"/review?paper=review-neurips-2026-154303","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154307","title":"Uncertainty-Aware Probabilistic Constrained Clustering from Entangled Pairwise Supervision","authors":["Shaojie Zhang","Ke Chen"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Uncertainty-Aware Probabilistic Constrained Clustering from Entangled Pairwise Supervision","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.12027","noteUrl":"/review?paper=review-neurips-2026-154307","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154309","title":"SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents","authors":["Yu Yang","Yue Liao","Jianbiao Mei","Baisen Wang","Jiangning Zhang","Xiangtai Li","Liang Lv","Hanlin Chen","Yong Liu","Shuicheng Yan","Gim Hee Lee"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.08403","noteUrl":"/review?paper=review-neurips-2026-154309","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154310","title":"Hallucination as Commitment Failure: Larger LLMs Misfire Despite Knowing the Answer","authors":["Jewon Yeom","Jaewon Sok","Heejun Kim","Seonghyeon Park","Jeongjae Park","Taesup Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Alias completeness bounds a set probability, not the emitted token’s probability","plainExplanation":"Alias completeness bounds a set probability, not the emitted token’s probability.","primaryUrl":"https://arxiv.org/html/2605.22007","noteUrl":"/review?paper=review-neurips-2026-154310","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-154311","title":"Asymptotic Anytime-Valid Inference for U-statistics","authors":["Leheng Cai","Qirui Hu","Weijia Li"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Asymptotic Anytime-Valid Inference for U-statistics","plainExplanation":"A vanishing approximation error does not preserve coverage at a deterministic support endpoint without a one-sided remainder margin.","primaryUrl":"https://arxiv.org/html/2605.14692","noteUrl":"/review?paper=review-neurips-2026-154311","collection":"neurips-2026","severity":"major","issueTypes":["implementation_mismatch","mathematical_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154312","title":"Diffusion Meta-Prompting and Steering for Generalizable Foundation Model Adaptation","authors":["Deepak Sridhar","Yi Li","Kartikeya Bhardwaj","Shuangjun Liu","TAOTAO JING","Yuan Li","Shuai Zhang","Jiancheng Lyu","Dashan Gao","Nuno Vasconcelos"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Diffusion Meta-Prompting and Steering for Generalizable Foundation Model Adaptation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.11067","noteUrl":"/review?paper=review-neurips-2026-154312","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154314","title":"One More Time: Revisiting Neural Quantum States from a Reinforcement Learning Perspective","authors":["Juan Duque","Sergio García Heredia","Vinicius Hernandes","Eliska Greplova","Thomas Spriggs","Aaron Courville","Anna Dawid"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"One More Time: Revisiting Neural Quantum States from a Reinforcement Learning Perspective","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.02292","noteUrl":"/review?paper=review-neurips-2026-154314","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154315","title":"Amortized-Precision Quantization for Early-Exit Vision Transformers","authors":["Rui Fang","Hsi-Wen Chen","Ming-syan Chen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Amortized-Precision Quantization for Early-Exit Vision Transformers","plainExplanation":"The reduction treats threshold-reachable sets as arbitrary subsets, although they form a nested chain.","primaryUrl":"https://arxiv.org/html/2605.07317","noteUrl":"/review?paper=review-neurips-2026-154315","collection":"neurips-2026","severity":"major","issueTypes":["invalid_reduction","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154317","title":"VideoMDM: Towards 3D Human Motion Generation From 2D Supervision","authors":["Amir Mann","Gal M Harari","Merav keidar","Or Litany"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The camera matrix loses an independent spatial coordinate","plainExplanation":"The appendix camera projection uses three rows that depend on the same rotated horizontal coordinate \\(q=\\cos\\theta\\,x+\\sin\\theta\\,z\\) and on \\(y\\). The resulting matrix has rank at most two.","primaryUrl":"https://arxiv.org/html/2606.13364","noteUrl":"/review?paper=review-neurips-2026-154317","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154318","title":"Agentic-imodels: Evolving agentic interpretability tools via autoresearch","authors":["Chandan Singh","Yan Shuo Tan","Weijia Xu","Zelalem Gero","Weiwei Yang","Michel Galley","Jianfeng Gao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Repeated judges do not remove analysis-level variance","plainExplanation":"The uncertainty calculation treats judge scores from regenerated analyses as independent samples when reporting the SEM of the overall mean. Under \\(Y_{ij}=\\mu+A_i+E_{ij}\\) with independent analyses and conditionally independent judge noise, the mean variance is \\(\\sigma_A^2/R+\\sigma_E^2/(RJ)\\).","primaryUrl":"https://arxiv.org/html/2605.03808","noteUrl":"/review?paper=review-neurips-2026-154318","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154319","title":"ExComm: Exploration-Stage Communication for Error-Resilient Agentic Test-Time Scaling","authors":["Woomin Song","Beomjun Kim","Daewon Choi","Sai Muralidhar Jayanthi","Saket Dingliwal","Jinwoo Shin","Aram Galstyan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"ExComm: Exploration-Stage Communication for Error-Resilient Agentic Test-Time Scaling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22102","noteUrl":"/review?paper=review-neurips-2026-154319","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154321","title":"Tool-IQA: Augmenting Image Quality Assessment with Simple Tools","authors":["Guanyi Qin","Junjie Zhang","Chunming He","Yibing Fu","Jie Liang","Tianhe Wu","Lei Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The tool-free reward baseline can average an empty cohort","plainExplanation":"The reward baseline averages candidates that did not use a tool, but the candidate-generation scheme does not guarantee that any such candidate exists. If all four candidates use tools, the tool-free index set is empty and its mean is undefined.","primaryUrl":"https://arxiv.org/html/2606.16082","noteUrl":"/review?paper=review-neurips-2026-154321","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154322","title":"OperatorSHAP: Fast and Accurate Shapley Value Estimation for Neural Operators","authors":["Joshua Stiller","Santo Thies","Felix Czaja","Eyke Hüllermeier"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OperatorSHAP: Fast and Accurate Shapley Value Estimation for Neural Operators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.28065","noteUrl":"/review?paper=review-neurips-2026-154322","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154324","title":"Extracting Search Trees from LLM Reasoning Traces Reveals Myopic Planning","authors":["Sixing Chen","Ji-An Li","Saner Cakir","Sinan Akcali","Kayla Lee","Marcelo G Mattar"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Extracting Search Trees from LLM Reasoning Traces Reveals Myopic Planning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06840","noteUrl":"/review?paper=review-neurips-2026-154324","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154326","title":"Last-Iterate Convergence of Single-Loop Stochastic Methods for Constrained Convex-Concave Minimax Problems","authors":["Taoli Zheng","Jiajin Li","Anthony Man-Cho So"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Block Lipschitz constants do not imply the stated full-operator bound","plainExplanation":"The minimax proof uses blockwise \\(\\ell_1\\)-distance Lipschitz assumptions as if they bounded the full saddle operator by the same constant in Euclidean norm. The convex–concave function \\(F(x,y)=x^2/2+xy\\) satisfies the block assumptions with \\(L=1\\), but its full operator norm is the golden ratio.","primaryUrl":"https://arxiv.org/html/2607.11056","noteUrl":"/review?paper=review-neurips-2026-154326","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154327","title":"Radial-Angular Geometry for Reliable Update Diagnosis in Noisy-Label Learning","authors":["Ningkang Peng","Jingyang Mao","Xiaoqian Peng","Qu Weiguang","Yanhui Gu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Radial-Angular Geometry for Reliable Update Diagnosis in Noisy-Label Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17429","noteUrl":"/review?paper=review-neurips-2026-154327","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154328","title":"MHWA: Multi-timescale Hierarchical World-Action Model","authors":["Pengcheng Pan","Guoqing Ma","Yuhan Zhang","Yang Chen","Yichen Liu","Ziheng Li","Shan Yu"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Hard top-k frequency balance has zero local gradient","plainExplanation":"The load-balancing loss is a KL function of discrete hard-top-k expert frequencies and is differentiated as a routing regularizer. Whenever all routing-score gaps are strict, a neighborhood of parameters preserves every selected expert.","primaryUrl":"https://arxiv.org/html/2610.14727","noteUrl":"/review?paper=review-neurips-2026-154328","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154330","title":"Differencing the Diffusion Trajectory toward Uncertain Components for Time Series Forecasting","authors":["Chen Su","Yuanhe Tian","Yan Song"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Differencing the Diffusion Trajectory toward Uncertain Components for Time Series Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.22599","noteUrl":"/review?paper=review-neurips-2026-154330","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154331","title":"UniVL: Unified Vision-Language Embedding for Spatially Grounded Contextual Image Generation","authors":["Jiayun (Peter) Wang","Yu Wang","Weijie Gan","Zhenting Wang","Wei Wei"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"UniVL: Unified Vision-Language Embedding for Spatially Grounded Contextual Image Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.21611","noteUrl":"/review?paper=review-neurips-2026-154331","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154336","title":"PEEK: Context Map as an Orientation Cache for Long-Context LLM Agents","authors":["Zhuohan Gu","Qizheng Zhang","Omar Khattab","Samuel Madden"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PEEK: Context Map as an Orientation Cache for Long-Context LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19932","noteUrl":"/review?paper=review-neurips-2026-154336","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154337","title":"Neural Reconstruction of LiDAR Point Clouds under Jamming Attacks via Full-Waveform Representation and Simultaneous Laser Sensing","authors":["Ryo Yoshida","Takami Sato","Wenlun Zhang","Yuki Hayakawa","Shota Nagai","Kentaro Yoshioka"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Neural Reconstruction of LiDAR Point Clouds under Jamming Attacks via Full-Waveform Representation and Simultaneous Laser Sensing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.00371","noteUrl":"/review?paper=review-neurips-2026-154337","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154338","title":"CausalCine: Real-Time Autoregressive Generation for Multi-Shot Video Narratives","authors":["Yihao Meng","Zichen Liu","Hao Ouyang","Qiuyu Wang","Ka Leong Cheng","Yue Yu","Hanlin Wang","Xing Zhu","Yujun Shen","Qifeng Chen","Huamin Qu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"CausalCine: Real-Time Autoregressive Generation for Multi-Shot Video Narratives","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12496","noteUrl":"/review?paper=review-neurips-2026-154338","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154339","title":"TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining","authors":["Shih-Ying Yeh","Daniel Z Kaplan","Xuehai Wang","Fu-En Yang","Min-Hung Chen","Shang-Hong Lai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.33419","noteUrl":"/review?paper=review-neurips-2026-154339","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154340","title":"Approaching I/O-optimality for Approximate Attention","authors":["Pál András Papp","Alek Sobczyk","Anastasios Zouzias"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The attention scaling and one factorial estimate need repair","plainExplanation":"Row normalization does not cancel a change of temperature inside the exponential. Entry bounds also do not bound dot products by the same number.","primaryUrl":"https://arxiv.org/html/2605.23751","noteUrl":"/review?paper=review-neurips-2026-154340","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154343","title":"Orthogonal Uplift Learning with Permutation-Invariant Representations for Combinatorial Treatments","authors":["Jiacan Gao","Xinyan Su","Mingyuan Ma","Xiao Xu","Xinrui Wan","Tianqi Gu","Enyun Yu","Jiecheng Guo","zhiheng zhang"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"A second-moment score bound does not control the claimed squared nuisance error","plainExplanation":"A second-moment score bound does not control the claimed squared nuisance error. The plug-in product expansion omits a nuisance cross term. The stated orthogonal-score bound misses a propensity-error square.","primaryUrl":"https://arxiv.org/html/2602.19851","noteUrl":"/review?paper=review-neurips-2026-154343","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"lomo-supervised-loss-kl-alignment","title":"LoMo: Local Modality Substitution for Deeper Vision-Language Fusion","authors":["Feng Han","Zhixiong Zhang","Zheming Liang","Yibin Wang","Jiaqi Wang"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Fixed-label training does not add the claimed alignment KL","plainExplanation":"The KL identity averages over model-sampled answers. Fixed corpus labels follow a different distribution, and improving their supervised loss can increase cross-carrier KL.","primaryUrl":"https://arxiv.org/abs/2605.30265v1","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154346","title":"Probing Visual Planning in Image Editing Models","authors":["Zhimu Zhou","Yanpeng Zhao","Qiuyu Liao","Bo Zhao","Xiaojian (Shawn) Ma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Probing Visual Planning in Image Editing Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.22868","noteUrl":"/review?paper=review-neurips-2026-154346","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154347","title":"The Causal Description Gap: Information-Theoretic Separations Across Pearl's Hierarchy","authors":["Seyedmorteza Emadi"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"The Causal Description Gap: Information-Theoretic Separations Across Pearl's Hierarchy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.02177","noteUrl":"/review?paper=review-neurips-2026-154347","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154348","title":"USAD: Uncertainty-aware Statistical Adversarial Detection","authors":["Zhijian Zhou","Xunye Tian","Jiacheng Zhang","Zesheng Ye","Yiyi Guo","Donghao Zhang","Liuhua Peng","Feng Liu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"USAD: Uncertainty-aware Statistical Adversarial Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.27832","noteUrl":"/review?paper=review-neurips-2026-154348","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154349","title":"ToolSearcher: Optimizing Tool Selection at Scale via Reinforcement Learning","authors":["Zhenlong Dai","Xujie Song","Zitong Wang","Tong Niu","Jian liu","Weiqiang Wang","Xiu Tang","Sai Wu","Chang Yao","Jingyuan Chen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The selected-event product is not the expected discovery reward","plainExplanation":"Constant rewards make the printed group advantage undefined. The selected-event product is not the expected discovery reward.","primaryUrl":"https://arxiv.org/html/2609.30906","noteUrl":"/review?paper=review-neurips-2026-154349","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154356","title":"PRPO: Perception-Reinforced Policy Optimization via Token-Level Dynamic Advantage Reshaping","authors":["Qiming Li","Tianlun Li","Xiaolong Cheng","Hangyu Li","Ruiyan Gong","Kangning Niu","kaitao jiang","Mu Xu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"PRPO: Perception-Reinforced Policy Optimization via Token-Level Dynamic Advantage Reshaping","plainExplanation":"Multiplying a fixed advantage by a positive number cannot make a locally constant clipped objective have a nonzero derivative.","primaryUrl":"https://arxiv.org/html/2606.08708","noteUrl":"/review?paper=review-neurips-2026-154356","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154358","title":"Wavefunction Flows: Efficient Quantum Simulation of Continuous Flow Models","authors":["David Layden","Ryan Sweke","Vojtech Havlicek","Anirban Chowdhury","Kirill Neklyudov"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"Wavefunction Flows: Efficient Quantum Simulation of Continuous Flow Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.08462","noteUrl":"/review?paper=review-neurips-2026-154358","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154359","title":"Scale Where It Matters: Training-Free Localized Scaling for Diffusion Models","authors":["Qin Ren","Yufei Wang","Lanqing Guo","Wen Zhang","Zhiwen Fan","Chenyu You"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Expected precision cannot be replaced by a ratio of expected counts","plainExplanation":"Expected precision cannot be replaced by a ratio of expected counts. Recall alone does not determine the benefit of a localized repair. An oracle choosing one candidate cannot combine repairs from different candidates.","primaryUrl":"https://arxiv.org/html/2511.19917","noteUrl":"/review?paper=review-neurips-2026-154359","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154361","title":"Where to Look Is Not How to Fix: Pre-Denoising Diagnostics and Modality-Dependent Control in Diffusion Composition","authors":["Fangzheng Wu","Brian Summa"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Where to Look Is Not How to Fix: Pre-Denoising Diagnostics and Modality-Dependent Control in Diffusion Composition","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03068","noteUrl":"/review?paper=review-neurips-2026-154361","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154362","title":"Discrete Langevin-Inspired Posterior Sampling","authors":["Sattwik Basu","Chaitanya Amballa","Jorge V Sampedro","Romit Roy Choudhury"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Discrete Langevin-Inspired Posterior Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09302","noteUrl":"/review?paper=review-neurips-2026-154362","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154364","title":"BrickAnything: Geometry-Conditioned Buildable Brick Generation with Structure-Aware Tokenization","authors":["zhengyang Ni","Feng yan","Yu Guo","Fei Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"BrickAnything: Geometry-Conditioned Buildable Brick Generation with Structure-Aware Tokenization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26182","noteUrl":"/review?paper=review-neurips-2026-154364","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154368","title":"Every Bit, Everywhere, All At Once: A Binomial Multibit LLM Watermark","authors":["Thibaud Gloaguen","Robin Staab","Mark Vero","Martin Vechev"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The recovery score uses total rather than remaining future variance","plainExplanation":"Write \\(S=\\sum_uq_uG_u\\). From \\(q_u=p_u(1+G_u-S)\\), one obtains \\(S=(\\mathbb E_pG+\\mathbb E_pG^2)/(1+\\mathbb E_pG)\\) and \\(\\sum_uq_u=1+\\mathbb E_pG-S\\), not generally one. For binary scores and \\(p=(1/2,1/2)\\), \\(S=2/3\\) and total mass is \\(5/6\\). This is a local index/normalization typo. The later correct form and actual empirical implementation are not refuted.","primaryUrl":"https://arxiv.org/html/2605.11653","noteUrl":"/review?paper=review-neurips-2026-154368","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154369","title":"Curriculum Learning for Safety Alignment","authors":["Sandeep Kumar","Virginia Smith","Chhavi Yadav"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Curriculum Learning for Safety Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26315","noteUrl":"/review?paper=review-neurips-2026-154369","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154370","title":"POP: Online Structural Pruning Enables Efficient Inference of Large Foundation Models","authors":["Yi Chen","Wonjin Shin","Shuhong Liu","Tho Mai","Jeongmo Lee","Jun Liu","Chuanbo Hua","Kun Wang","Joo-Young Kim"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"POP: Online Structural Pruning Enables Efficient Inference of Large Foundation Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.06822","noteUrl":"/review?paper=review-neurips-2026-154370","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154372","title":"GeneZip: Region-Aware Compression for Long Context DNA Modeling","authors":["Jianan Zhao","Xixian Liu","Zhihao Zhan","XINYU YUAN","Hongyu Guo","Jian Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Whole-window routing is not guaranteed prefix-causal","plainExplanation":"Two windows can have identical prefix boundary probabilities but different suffix confidence. The global confidence projection can then retain different prefix boundaries, even with the stated minimum retained count and first-position rule. Earlier compressed representations can therefore depend on suffix information before entering a causal mixer; causal mixing and dechunking do not remove that dependency. The printed autoregressive path lacks a causality argument. This does not establish that learned outputs exploit the dependency or that an unrelated diagnostic leaked into the reported PPL.","primaryUrl":"https://arxiv.org/html/2602.17739","noteUrl":"/review?paper=review-neurips-2026-154372","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154373","title":"Causal Evaluation of Membership Inference Attacks","authors":["Mathieu Even","Clément Berenfeld","Linus Bleistein","Tudor Cebere","Julie Josse","Aurélien Bellet"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The transported control term needs the treated-population normalizer","plainExplanation":"With propensity \\(\\pi(X)\\), \\(\\mathbb E[\\pi(X)(1-A)Y/(1-\\pi(X))]=\\mathbb E[\\pi(X)g(X)]\\), where \\(g\\) is the control regression. The treated-population density ratio is \\(\\pi(X)/P(A=1)\\), so this numerator equals \\(P(A=1)\\mathbb E_TY(0)\\). Dividing by \\(P(A=0)\\) scales the target by the treated/control prevalence ratio. The appendix normalization is wrong for unequal pools. Balanced experiments hide it; the main-text proposition and estimator are correct on this point.","primaryUrl":"https://arxiv.org/html/2602.02819","noteUrl":"/review?paper=review-neurips-2026-154373","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154375","title":"Density-Ratio Losses for Post-Hoc Learning to Defer","authors":["Alexander Soen","Ragnar Thobaben","Joakim Jaldén","Richard Nock"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Density-Ratio Losses for Post-Hoc Learning to Defer","plainExplanation":"An added cost makes deferral harder, so the loss difference must beat the negative cost threshold.","primaryUrl":"https://arxiv.org/html/2605.19557","noteUrl":"/review?paper=review-neurips-2026-154375","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154376","title":"Articulation in Prime: Primitive-Based Articulated Object Understanding from a Single Casual Video","authors":["Arslan Artykov","Tom Ravaud","Nicolás Violante","Vincent Lepetit"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Articulation in Prime: Primitive-Based Articulated Object Understanding from a Single Casual Video","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18645","noteUrl":"/review?paper=review-neurips-2026-154376","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154377","title":"Fully Distributed Tâtonnement for Chores Markets","authors":["Bhaskar Ray Chaudhury","Christian Kroer","Ruta Mehta","Tianlong Nan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The iteration bound omits the time needed for tiny initial prices to grow","plainExplanation":"The price escape estimate is multiplicative rather than additive. The iteration bound omits the time needed for tiny initial prices to grow.","primaryUrl":"https://arxiv.org/html/2607.00300","noteUrl":"/review?paper=review-neurips-2026-154377","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154378","title":"Learning Deployable Causal Action Geometry under Temporal Non-Stationarity","authors":["Changjian Liu","Tianyu Wang","Yuwei Xu","Xiaoxuan Deng","Zhilin Zhang","Yong Gao","Chuan Yu","Jian Xu","Bo Zheng"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Distance to an oracle set does not control a fixed oracle representative","plainExplanation":"The optimization criterion bounds the infimum of distance to the oracle minimizer set, but the later policy bridge compares with a fixed selected oracle minimizer. Zero distance to the set does not imply zero distance to every representative.","primaryUrl":"https://arxiv.org/html/2610.16644","noteUrl":"/review?paper=review-neurips-2026-154378","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154379","title":"Instructing LLMs to Negotiate using Reinforcement Learning with Verifiable Rewards","authors":["Shuze D Liu","Claire Chen","Jiabao S Xiao","Lei Lei","Yuheng Zhang","Yisong Yue","David Simchi-Levi"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Instructing LLMs to Negotiate using Reinforcement Learning with Verifiable Rewards","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.09855","noteUrl":"/review?paper=review-neurips-2026-154379","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154380","title":"A Unified Framework for Uniform-Price Resource Allocation Mechanisms","authors":["Ioannis Caragiannis","Dimitris Fotakis","Stratis Skoulakis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Equilibrium existence needs stronger valuation assumptions","plainExplanation":"If only one bidder values the resource, any positive bid by that bidder wins the whole unit. Reducing that bid lowers payment without changing allocation. Bidding zero is no solution either: the paper assigns zero allocation when every bid is zero, allowing a profitable small positive bid. These boundary cases prevent equilibrium under the stated valuation assumptions. The example refutes Theorem 2's universal existence claim and its step asserting $\\mu>0$. Restricting to $n\\ge2$ continuously differentiable, strictly increasing, concave valuations repairs this existence argument. Welfare bounds conditional on an existing equilibrium, separate nonsmooth claims and numerical learning results require their own arguments.","primaryUrl":"https://arxiv.org/html/2606.06151","noteUrl":"/review?paper=review-neurips-2026-154380","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154381","title":"Theory of Optimal Learning Rate Schedules and Scaling Laws for a Random Feature Model","authors":["Blake Bordelon","Francesco Mori"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The momentum correlation recurrence uses the previous schedule index","plainExplanation":"Under the displayed definition, \\(v_{t+1}\\) contains \\(\\beta_{t+1}\\). Its correlations with \\(\\Delta_{t+1}\\) must therefore carry \\(\\beta_{t+1}\\), not \\(\\beta_t\\), whenever the schedule varies. Changing gradient evaluation timing does not alter that coefficient index. This is a local recurrence inconsistency. Consistent reindexing may preserve the optimized schedule family; implementation behavior is unknown.","primaryUrl":"https://arxiv.org/html/2602.04774","noteUrl":"/review?paper=review-neurips-2026-154381","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154385","title":"TrajLoc: Trajectory-Attention Localization for Multi-Object Motion Control","authors":["Omer Sela","Inbar Huberman-Spiegelglas","Michael Rotman","Sagie Benaim","Avi Ben-Cohen"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"TrajLoc: Trajectory-Attention Localization for Multi-Object Motion Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.00861","noteUrl":"/review?paper=review-neurips-2026-154385","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154387","title":"On the Relaxation of Conditional Independence Assumption for Image Segmentation","authors":["Zixun Wang","Ben Dai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"On the Relaxation of Conditional Independence Assumption for Image Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38930","noteUrl":"/review?paper=review-neurips-2026-154387","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154389","title":"When Should LLMs Be Less Specific? Selective Abstraction for Reliable Long-Form Text Generation","authors":["Shani Goren","Ido Galil","Ran El-Yaniv"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Exchangeability alone does not give the conditional Beta coverage law","plainExplanation":"Exchangeability alone does not give the conditional Beta coverage law.","primaryUrl":"https://arxiv.org/html/2602.11908","noteUrl":"/review?paper=review-neurips-2026-154389","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-154392","title":"What Time Is It? How Data Geometry Makes Time Conditioning Optional for Flow Matching","authors":["Alec Helbling","Sebastian Gutierrez Hernandez","Benjamin Hoover","Duen Horng Chau","Parikshit Ram"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Conditional-mean error is not unconditional target variance","plainExplanation":"Term I is \\(\\mathbb E\\|f(Z)-\\mathbb E[U\\mid Z]\\|^2\\), but its claimed characterization uses unconditional \\(\\operatorname{Var}(U)\\) independently of \\(f\\). Taking \\(f(Z)=\\mathbb E[U\\mid Z]\\) makes Term I zero even when \\(\\operatorname{Var}(U)>0\\).","primaryUrl":"https://arxiv.org/html/2605.08344","noteUrl":"/review?paper=review-neurips-2026-154392","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154394","title":"TopoPrune: Robust Data Pruning via Unified Latent Space Topology","authors":["Arjun Roy","Prajna Malettira","Manish Nagaraj","Kaushik Roy"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Positive scalarization does not imply strict superiority to both endpoints","plainExplanation":"An exact maximizer of a positive weighted sum is nondominated, but cannot strictly dominate an endpoint that maximizes one of those component objectives unless that endpoint has ties. Low correlation does not change this optimality fact. Stratified sampling also need not maximize the weighted sum at all, and downstream accuracy is a third objective not covered by the scalarization argument. The strict-superiority inference is unsupported. Useful empirical mixtures and the valid scalarization nondomination result remain possible.","primaryUrl":"https://arxiv.org/html/2602.02739","noteUrl":"/review?paper=review-neurips-2026-154394","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154398","title":"Flow Map Denoisers: Traversing the Distortion-Perception Plane for Inverse Problems","authors":["Nicolas Zilberstein","Morteza Mardani","Santiago Segarra"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Flow Map Denoisers: Traversing the Distortion-Perception Plane for Inverse Problems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19802","noteUrl":"/review?paper=review-neurips-2026-154398","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154399","title":"Expected Batch Optimal Transport Plans and Consequences for Flow Matching","authors":["Samuel Boïté","Julie Delon","Kimia Nadjahi"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Expected Batch Optimal Transport Plans and Consequences for Flow Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12174","noteUrl":"/review?paper=review-neurips-2026-154399","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154405","title":"Decoding the Critique Mechanism in Large Reasoning Models","authors":["Hoang Phan","Nguyen Hung-Quang","Thanh Quoc Hung Le","Xiusi Chen","Heng Ji","Khoa D Doan"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Decoding the Critique Mechanism in Large Reasoning Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.16331","noteUrl":"/review?paper=review-neurips-2026-154405","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154406","title":"Autoregressive Learning in Joint KL: Sharp Oracle Bounds and Lower Bounds","authors":["Yunbei Xu","Yuzhe Yuan","Ruohan Zhan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Fano calculation does not yield the claimed horizon-scaled lower bound","plainExplanation":"A fitted batch posterior can violate the claimed deterministic population-KL oracle bound on a positive-probability training event. The dependent-class lower-bound proof has an invalid Fano step, and its uniform horizon scaling exceeds the declared model’s KL range. The decomposable proof also confuses one-step and full policy-class cardinalities. Separately, signed reward control needs twice the printed total-variation coefficient. The reports state weaker valid guarantees with their learner, cardinality, and parameter restrictions.","primaryUrl":"https://arxiv.org/html/2605.12316","noteUrl":"/review?paper=review-neurips-2026-154406","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154408","title":"Discrete Flow Matching: Convergence Guarantees Under Minimal Assumptions","authors":["Le-Tuyet-Nhi PHAM","Giovanni Conforti","Zhenjie Ren","Alain Durmus"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The binary-torus generator loses neighbor multiplicity","plainExplanation":"The torus generator assigns rate \\(1/2\\) for each of the two directions and diagonal rate \\(-1\\), using a piecewise off-diagonal definition. On \\(\\mathbb Z_2\\), the \\(+1\\) and \\(-1\\) neighbors are the same state.","primaryUrl":"https://arxiv.org/html/2605.08882","noteUrl":"/review?paper=review-neurips-2026-154408","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154412","title":"TTCS: Test-Time Curriculum Synthesis for Self-Evolving","authors":["Chengyi Yang","Zhishang Xiang","Yunbo Tang","Zongpei Teng","Chengsong Huang","Fei Long","Yuhan Liu","Jinsong Su"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Modal agreement and reward variance do not justify the stated gradient claims","plainExplanation":"The reference answer is selected from the same observations. Its agreement indicators are therefore not independent correctness trials, even when the answer distribution is symmetric.","primaryUrl":"https://arxiv.org/html/2601.22628","noteUrl":"/review?paper=review-neurips-2026-154412","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154413","title":"Retrieval-Augmented Diffusion Modeling for Stochastic Discount Factor Portfolios","authors":["Kelvin J.L. Koa","XinYang Li","Ke-Wei Huang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Retrieval-Augmented Diffusion Modeling for Stochastic Discount Factor Portfolios","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.35086","noteUrl":"/review?paper=review-neurips-2026-154413","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154414","title":"VISD: Enhancing Video Reasoning via Structured Self-Distillation","authors":["Hao Lin","Kunyang Lv","Xu Jiang","Jingqi Tian","Zhongjing Du","Jiayu Ding","Qiaoman Zhang","Hongbo Jin"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"VISD: Enhancing Video Reasoning via Structured Self-Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06094","noteUrl":"/review?paper=review-neurips-2026-154414","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154417","title":"ActWorld: From Explorable to Interactive World Model via Action-Aware Memory","authors":["Zhexiao Xiong","Yizhi Song","Hao Kang","Qing Yan","Liming Jiang","Jenson Yang","ZHOUJIE FU","Stathi Fotiadis","Zichuan Liu","Bo Liu","Yiding Yang","Xin Lu","Nathan Jacobs"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Zero-valued tokens still change attention normalization","plainExplanation":"The memory-bank initialization claims to preserve a pretrained forward pass exactly because every prepended bank token initially has zero value. A zero key and value contribute no numerator but add \\(e^0=1\\) to the softmax denominator.","primaryUrl":"https://arxiv.org/html/2606.17730","noteUrl":"/review?paper=review-neurips-2026-154417","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"discriminative-factorization-label-information","title":"Black-box model classification under the discriminative factorization","authors":["Hayden Helm","Merrick Ohata","Carey Priebe"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Cross-class pair separation does not preserve label information","plainExplanation":"Existential cross-class separation can hold while an embedding discards every label bit. Discriminative factorization rank also does not bound the geometric MDS dimension by its square, and a finite sample can miss directions. SVD recovers a column space without necessarily recovering individual factor loadings or zero sets.","primaryUrl":"https://arxiv.org/html/2605.07878v1","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154424","title":"Dirichlet-Guided Group Forecasting for Alleviating Over-smoothing in Time Series Forecasting","authors":["Xingyu Zhang","Jingyao Wang","Xin Yu","Zeen Song","Jianqi Zhang","Changwen Zheng","Wenwen Qiang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Dirichlet-Guided Group Forecasting for Alleviating Over-smoothing in Time Series Forecasting","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.10592","noteUrl":"/review?paper=review-neurips-2026-154424","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154425","title":"Policy-DRIFT: Dynamic Reward-Informed Flow Trajectory Steering","authors":["Atharva Mahajan","Abhijeet Yashwant Vishwasrao","Yuning Wang","Ricardo Vinuesa"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Pre- and post-guidance share the same first-order dynamics","plainExplanation":"The explanation attributes a general physical-feasibility advantage to applying guidance before the learned velocity step rather than after it. For smooth fields \\(v,g\\), the pre-step is \\(u+h g(u)+h v(u+h g(u))=u+h(v+g)+h^2Dv\\,g+O(h^3)\\).","primaryUrl":"https://arxiv.org/html/2605.14022","noteUrl":"/review?paper=review-neurips-2026-154425","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154426","title":"CORAL: Learning Amyloid Fibril Ligand Docking with Cooperative Binding Rewards","authors":["Yasheng SUN","Bohan Li","Youqi Tao","Jürgen Schmidhuber"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"CORAL: Learning Amyloid Fibril Ligand Docking with Cooperative Binding Rewards","plainExplanation":"No issue was found in the checked claims. The minus sign aligns reward maximization with the stated preference for lower binding scores.","primaryUrl":"https://arxiv.org/html/2607.17412","noteUrl":"/review?paper=review-neurips-2026-154426","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154428","title":"Decoupling Exploration and Policy Optimization: Uncertainty Guided Tree Search for Hard Exploration","authors":["Zakaria Mhammedi","James Cohan"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The finite-population guarantee needs a conditional concentration argument","plainExplanation":"The recurrence describes a ratio of expected counts, rather than the expectation of the random resampling ratio. Its fixed point does not establish the claimed finite-population expectation bound. Concentration must be applied conditionally to the actual population and its resampling step.","primaryUrl":"https://arxiv.org/html/2603.22273","noteUrl":"/review?paper=review-neurips-2026-154428","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154429","title":"RVR: Retrieve-Verify-Retrieve for Comprehensive Question Answering","authors":["Deniz Qian","Hung-Ting Chen","Eunsol Choi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Training can exhaust the set of remaining positive documents","plainExplanation":"The training procedure samples \\(m\\) uniformly through \\(\\min(M,|D^*|)\\), places those gold documents in context, and then samples a positive from \\(D^*\\setminus D_{\\rm ctx}\\). For any \\(1\\leq|D^*|\\leq M\\), the allowed event \\(m=|D^*|\\) has positive probability and leaves an empty positive set.","primaryUrl":"https://arxiv.org/html/2602.18425","noteUrl":"/review?paper=review-neurips-2026-154429","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154430","title":"Learning to Recommend in Unknown Games","authors":["Arwa Alanqary","Zakaria Baba","Manxi Wu","Alexandre Bayen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The ratio-search bracket must depend on the lower weight bound","plainExplanation":"Definition 7 excludes domination by another pure action. Theorem 6 represents the full labeled best-response indistinguishability set using extreme points of polyhedra normally equivalent to $\\widetilde{\\mathcal P}_i=\\mathcal P_i+C_i^\\circ$, where $C_i=\\mathbb R_+^{d_i}$. Here $\\sigma$ assigns action labels to vertices and $\\mathcal N$ denotes the normal fan. The proof assumes every action remains a vertex after adding the nonpositive orthant. An action can be dominated by a mixture without being dominated by another pure action. Such an action can be a vertex of the original utility polytope $\\mathcal P_i$ but cease to be a vertex after adding the nonpositive orthant $C_i^\\circ$. Extreme points of the polarized polyhedron then lose an action label required by the theorem's set equality.","primaryUrl":"https://arxiv.org/html/2602.16998","noteUrl":"/review?paper=review-neurips-2026-154430","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154432","title":"Drift-React: One-step Generation of Reaction Pathways via SE(3) Drifting Fields","authors":["Rémi Schlama","Philippe Schwaller"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A negative saddle eigenvalue gives quadratic energy decrease","plainExplanation":"The geometric-mean normalization defines a different empirical field rather than an equivalent product-kernel estimator. The transition-state explanation also confuses negative curvature with a first-order energy term, which vanishes at a stationary point.","primaryUrl":"https://arxiv.org/html/2605.22990","noteUrl":"/review?paper=review-neurips-2026-154432","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154433","title":"Midpoint Generative Models","authors":["Daniil Shlenskii","Nikita Gushchin","Lev Novitskiy","Dmitry V. Dylov","Aleksandr Korotin"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Midpoint Generative Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29920","noteUrl":"/review?paper=review-neurips-2026-154433","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154434","title":"Active Learning From Positive and Unlabeled Examples","authors":["Farnam Mansouri","Sandra Zilles","Shai Ben-David"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The printed low-response implication fails at initialization","plainExplanation":"The displayed learners have undefined states, invalid small-sample radii and a missing pre-query stopping check. Several supporting lemmas also reverse response/tail directions, overextend a Chernoff bound or use a global class condition for current disagreement. EstRate’s dyadic stopping is not the asserted negative-binomial law; local sign repairs do not complete its stochastic proof.","primaryUrl":"https://arxiv.org/html/2602.02081","noteUrl":"/review?paper=review-neurips-2026-154434","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154435","title":"Learning to Foresee: Unveiling the Unlocking Efficiency of On-Policy Distillation","authors":["Yuchen Cai","Ding Cao","Liang Lin","Chunxi Luo","Xin Xu","Kai Yang","Weijie Liu","Saiyong Yang","Tianxiang Zhao","Guangzhong Sun","Guiquan Liu","Junfeng Fang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The away-from-teacher KL Hessian has an additional residual term","plainExplanation":"For binary logits \\((\\theta,0)\\) and teacher logit \\(t^*\\), reverse-KL derivative is \\(p(1-p)(\\theta-t^*)\\). Its Hessian is \\(p(1-p)+(\\theta-t^*)p(1-p)(1-2p)\\). At base \\(\\theta=1\\) and teacher \\(t^*=5\\), this is approximately \\(0.560043\\), while the base Fisher is \\(0.196612\\). Linear logits remove network-linearization error, and adding a constant cannot change the Hessian. The near-teacher case where residual vanishes remains valid. The away-from-teacher extension and its downstream interpretation need additional error control.","primaryUrl":"https://arxiv.org/html/2605.11739","noteUrl":"/review?paper=review-neurips-2026-154435","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154437","title":"EvoDiagram: Agentic Editable Diagram Creation via Design Expertise Evolution","authors":["Tianfu Wang","Leilei Ding","Ziyang Tao","Yi Zhan","zhiyuan","Wei Wu","Yuxuan Lei","Junyang Wang","Yizhao Xu","Yin WU","Zhengyu Hu","Hongyuan Zhu","Nicholas Jing Yuan","Yanyong Zhang","Hui Xiong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EvoDiagram: Agentic Editable Diagram Creation via Design Expertise Evolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.09568","noteUrl":"/review?paper=review-neurips-2026-154437","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154438","title":"OrthoReg: Orthogonal Regularization for Hybrid Symbolic-Neural Dynamical Systems","authors":["Till Richter","Niki Kilbertus"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OrthoReg: Orthogonal Regularization for Hybrid Symbolic-Neural Dynamical Systems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19145","noteUrl":"/review?paper=review-neurips-2026-154438","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154440","title":"DORA: A Scalable Asynchronous Reinforcement Learning System for Language Model Training","authors":["Tianhao Hu","Xiangcheng Liu","Yuchun Miao","Youshao Xiao","Hong Yu Zang","Yang Zheng","huang xuan","Jinrui Ding","Yufei Zhang","Yu Yang","Yi-Kai Zhang","Yueqing Sun","Chengcheng Han","Xiandi Ma","Wei Wang","Qi GU","Yerui Sun","Yuchen Xie","Xunliang Cai"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A signed zero-mean advantage gap cannot bound the weighted surrogate error","plainExplanation":"The substitution-bias bound uses an identically zero signed advantage gap even when the importance-weighted gap is nonzero. The baseline characterization also mistakes sufficient conditions for necessary ones, excluding valid compensating coupling terms.","primaryUrl":"https://arxiv.org/html/2604.26256","noteUrl":"/review?paper=review-neurips-2026-154440","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154441","title":"H-Flow: Self-supervised Human Scene Flow via Physics-inspired Joint Multi-modal Learning","authors":["Zhanbo Huang","Xiaoming Liu","Yu Kong"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The physical-prior losses are not scale-invariant","plainExplanation":"The physical-prior losses do not remain unchanged under metric rescaling. The silhouette term scales in magnitude, and fixed skeletal/effort thresholds can change even their zero sets.","primaryUrl":"https://arxiv.org/html/2605.22629","noteUrl":"/review?paper=review-neurips-2026-154441","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154442","title":"Latent-space Attacks for Refusal Evasion in Language Models","authors":["Giorgio Piras","Raffaele Mura","Fabio Brau","Maura Pintor","Luca Oneto","Fabio Roli","Battista Biggio"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The SVM objective does not guarantee minimum zero-one error","plainExplanation":"LinearSVC with \\(C=0.1\\) minimizes regularized squared hinge loss, not zero-one classification error. Use \\(128\\) negative samples at \\(-0.01\\), \\(64\\) positives at \\(10^{-5}\\), and \\(64\\) positives at one, with the stated regularized intercept convention. The all-active optimum has weight about \\(1.210235\\) and intercept \\(-0.290830\\), misclassifying the \\(64\\) small positives. The DiM mean-difference direction is positive and separates all points at zero threshold. Normalizing weight and bias together does not change these signs. The optimization rationale is wrong. The measured SVM-versus-DiM comparison need not be false.","primaryUrl":"https://arxiv.org/html/2605.21706","noteUrl":"/review?paper=review-neurips-2026-154442","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154445","title":"Robust Inference-Time Steering of Protein Diffusion Models via Embedding Optimization","authors":["Minhuan Li","Jiequn Han","Pilar Cossio","Luhuan Wu"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Robust Inference-Time Steering of Protein Diffusion Models via Embedding Optimization","plainExplanation":"The denominator in Algorithm 1 does not follow from the stated Euler discretization and becomes zero at a schedule endpoint with \\(\\sigma_{t-1}=0\\).","primaryUrl":"https://arxiv.org/html/2602.05285","noteUrl":"/review?paper=review-neurips-2026-154445","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154448","title":"Attention-based PCA","authors":["Rodrigo Maulen Soto","Claire Boyer"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The orthogonal stationary set is a sphere rather than finitely many points","plainExplanation":"The gradient has form \\(2(A\\mu+B\\alpha v)\\), with \\(\\alpha=\\langle\\mu,v\\rangle\\) and \\(A=a_1\\|\\mu\\|^2+a_2\\alpha^2-a_3\\). Every \\(\\mu\\perp v\\) with radius \\(\\sqrt{a_3/a_1}\\) is stationary. For \\(d\\geq3\\), this is a \\((d-2)\\)-sphere, not \\(2(d-1)\\) isolated points. The zero tangential Hessian directions agree with this symmetry; negative spike and positive radial curvature make them saddles. The finite isolated-saddle enumeration is wrong. Global minimizers and a repaired manifold-based convergence argument remain possible but require proof.","primaryUrl":"https://arxiv.org/html/2605.18315","noteUrl":"/review?paper=review-neurips-2026-154448","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154449","title":"Interpretable Relational Inference with LLM-Guided Symbolic Dynamics Modeling","authors":["Xiaoxiao Liang","Juyuan Zhang","Liming Pan","Linyuan Lü"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Interpretable Relational Inference with LLM-Guided Symbolic Dynamics Modeling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.12806","noteUrl":"/review?paper=review-neurips-2026-154449","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154450","title":"Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces","authors":["William Gitta Lugoloobi","Samuele Marro","Jabez Magomere","Joss Wright","Chris Russell"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.14786","noteUrl":"/review?paper=review-neurips-2026-154450","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154453","title":"Multi-Head Recurrent Memory Agents","authors":["Jiatong Li","Samuel (Min-Hsuan) Yeh","Sharon Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Retention numerator and denominator use different capture windows","plainExplanation":"The retention denominator counts answers captured at any time through \\(T\\), whereas its numerator requires capture before \\(T\\) and presence at \\(T\\). An answer first captured at the final step is present in final memory, enters the denominator, but cannot enter the numerator.","primaryUrl":"https://arxiv.org/html/2607.01523","noteUrl":"/review?paper=review-neurips-2026-154453","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154455","title":"Neuron Populations Exhibit Divergent Selectivity with Scale","authors":["Amil Dravid","Yasaman Bahri","Alexei Efros","Yossi Gandelsman"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neuron Populations Exhibit Divergent Selectivity with Scale","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03990","noteUrl":"/review?paper=review-neurips-2026-154455","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154456","title":"Is $\\sqrt{d}$ Separation Necessary for Gradient EM to Learn Gaussian Mixtures in High Dimensions?","authors":["Yiran Zhang","Mo Zhou","Weihang Xu","Maryam Fazel","Simon Du"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Is $\\sqrt{d}$ Separation Necessary for Gradient EM to Learn Gaussian Mixtures in High Dimensions?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07551","noteUrl":"/review?paper=review-neurips-2026-154456","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154457","title":"SteerCast: Retrieval-Based Latent Steering for Decoder-Only Time Series Forecasting","authors":["Van Dai Do","Huu H Nguyen","Minh Hoang Nguyen","Hung Le"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SteerCast: Retrieval-Based Latent Steering for Decoder-Only Time Series Forecasting","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.11229","noteUrl":"/review?paper=review-neurips-2026-154457","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154458","title":"LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails","authors":["Nanxi Li","Zhengyue Zhao","Chaowei Xiao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.17329","noteUrl":"/review?paper=review-neurips-2026-154458","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154460","title":"Can Pixels Alone Reveal Image Origin? Minimax Limits and Learnable Interfaces for Passive Provenance","authors":["Kai Yao"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Can Pixels Alone Reveal Image Origin? Minimax Limits and Learnable Interfaces for Passive Provenance","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30997","noteUrl":"/review?paper=review-neurips-2026-154460","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154463","title":"Neural‑Visual Decoding via Cognitive‑guided Adaptive Blurring and Information‑Constrained Alignment","authors":["Fan Yin","Chuhang Zheng","Peiliang Gong","Donghai Guan","Qi Zhu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Neural‑Visual Decoding via Cognitive‑guided Adaptive Blurring and Information‑Constrained Alignment","plainExplanation":"No issue was found in the checked claims. The earlier identification of two different entropies has been removed. The current statement correctly describes what the regularizer directly controls: concentration of the spectral-selection weights.","primaryUrl":"https://arxiv.org/html/2605.16418","noteUrl":"/review?paper=review-neurips-2026-154463","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154464","title":"Language-Conditioned World Modeling for Visual Navigation","authors":["Yifei Dong","Fengyi Wu","Yilong Dai","Lingdong Kong","Guangyu Chen","Qiyu Hu","Yetong Sha","Feng Liu","Siyu Huang","Qi Dai","Zhi-Qi Cheng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The training actor reads expert states rather than its imagined rollout","plainExplanation":"The actor reads expert \\(s_k^E\\) after generated states have been appended to the world-model context. With \\(\\pi(s)=s\\), \\(F(s,a)=s+2a\\), and initial state one, the generated next state is three; the next on-policy action is therefore three, while the expert-state call can produce two. The resulting reward and return differ even with language and plan fixed. The printed rollout law conflicts with the main return definition. Actual code and empirical behavior are unverified.","primaryUrl":"https://arxiv.org/html/2603.26741","noteUrl":"/review?paper=review-neurips-2026-154464","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154466","title":"GRINQH: Graded Input-based Quantization Hierarchy for Efficient LLM Generation","authors":["Jette Oberländer","Jan Finkbeiner","Catherine Schoefmann","Emre Neftci"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"GRINQH: Graded Input-based Quantization Hierarchy for Efficient LLM Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.23419","noteUrl":"/review?paper=review-neurips-2026-154466","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154467","title":"Efficient Adaptive Data Analysis over Dense Data Distributions","authors":["Joon Suk Huh"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The stated uniform threshold fails for one fixed query","plainExplanation":"The estimator’s logarithm gives a nonpositive sample count. A separate concentration threshold also fails on a single fixed two-point query, so its constants need rederivation.","primaryUrl":"https://arxiv.org/html/2602.07732v1","noteUrl":"/review?paper=review-neurips-2026-154467","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154470","title":"BLEND: Balancing Personalization vs. Generalization in Federated Vision–Language Models","authors":["Payam Abdisarabshali","Fardis Nadimi","Kasra Borazjani","Naji Khosravan","Richeng Jin","Seyyedali Hosseinalipour"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"BLEND: Balancing Personalization vs. Generalization in Federated Vision–Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.12907","noteUrl":"/review?paper=review-neurips-2026-154470","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154473","title":"CSFlow: Aligning Flow Matching with Human Contrast Sensitivity","authors":["Malgorzata Galinska","Bart Pogodzinski","Jan Eric Lenssen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"CSFlow: Aligning Flow Matching with Human Contrast Sensitivity","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.08833","noteUrl":"/review?paper=review-neurips-2026-154473","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154476","title":"A World Model of Radiologist Reading for Medical Image Representation Learning","authors":["Yiwei Li","Zihao Wu","huaqin zhao","Yifan Zhou","Chao Cao","Dajiang Zhu","Tianming Liu","Lin Zhao"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"A prepended causal readout token cannot see the image sequence","plainExplanation":"The linear probe concatenates mean image features with a learned token prepended to a sequence processed by the frozen causal predictor. The position-zero token attends only to itself under the stated causal mask.","primaryUrl":"https://arxiv.org/html/2605.23992","noteUrl":"/review?paper=review-neurips-2026-154476","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154479","title":"NDPP-Grasp: Non-Differentiable Physical Plausibility Constraint-Guided Task-Oriented Dexterous Grasp Generation","authors":["Qiuchi Xiang","Haoxuan Qu","Hossein Rahmani","Jun Liu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"NDPP-Grasp: Non-Differentiable Physical Plausibility Constraint-Guided Task-Oriented Dexterous Grasp Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02432","noteUrl":"/review?paper=review-neurips-2026-154479","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154480","title":"Neural Statistical Functions","authors":["Daniel Xu","Yuxin Xie","Minghao Guo","Haixu Wu","Wojciech Matusik"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Differentiability at one point does not control moving secants","plainExplanation":"Proposition 4 infers convergence of every two-endpoint secant as both endpoints approach \\(s^*\\), assuming only differentiability at \\(s^*\\). Let \\(g(s)=(s-1/2)^2\\sin(1/(s-1/2))\\), extended by zero at \\(1/2\\).","primaryUrl":"https://arxiv.org/html/2605.11327","noteUrl":"/review?paper=review-neurips-2026-154480","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154482","title":"HAPS: Hierarchical LLM Routing with Joint Architecture and Parameter Search","authors":["Zihang Tian","Rui Li","Jingsen Zhang","Xiaohe Bo","Wei Huo","Xu Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Joint optimization does not justify the intermediate coordinate comparison","plainExplanation":"The loss chain replaces one coordinate by its jointly optimal value while holding the other coordinate at a separately optimized value. The convex loss \\(L(\\theta,\\Gamma)=(\\theta-\\Gamma)^2+\\Gamma^2/4\\) has joint optimum \\((0,0)\\).","primaryUrl":"https://arxiv.org/html/2601.05903","noteUrl":"/review?paper=review-neurips-2026-154482","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154483","title":"Sketching the Readout of Large Language Models for Scalable Data Attribution and Valuation","authors":["yide ran","Jianwen Xie","Minghui Wang","W. Jim Zheng","Denghui Zhang","Chuan Li","Zhaozhuo Xu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Matched-memory total variance need not have a dimension-free advantage","plainExplanation":"The CountSketch variance omits a swapped-index contribution. The matched-memory comparison also drops full product-variance terms, and the internal-gradient energy argument substitutes layer outputs for the inputs/Jacobians that actually determine parameter derivatives.","primaryUrl":"https://arxiv.org/html/2604.16197","noteUrl":"/review?paper=review-neurips-2026-154483","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154484","title":"Hybrid-LoRA: Bridging Full Fine-Tuning and Low-Rank Adaptation for Post-Training","authors":["Chengqian Zhang","Wei Zhu","Kyumin Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mean times standard deviation does not measure consistent high sensitivity","plainExplanation":"The module score is \\(H=\\mu\\sigma\\) for nonnegative sensitivity samples, and is interpreted as rewarding high, consistent sensitivity. At fixed positive mean, \\(H\\) increases with variability.","primaryUrl":"https://arxiv.org/html/2605.18822","noteUrl":"/review?paper=review-neurips-2026-154484","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154485","title":"Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role-Aware Policy Optimization","authors":["Zhenhua Xu","Dongsheng Chen","Jian Li","Yitong Lin","Zhebo Wang","Jiafu Wu","Yizhang Jin","Chengjie Wang","Meng Han","Yabiao Wang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Prompt-deletion entropy change is not conditional mutual information","plainExplanation":"The fixed-profile versus deleted-prompt entropy difference is not the conditional mutual information of an identified joint law. It can be negative or vanish despite a large change in token probabilities.","primaryUrl":"https://arxiv.org/html/2606.27025","noteUrl":"/review?paper=review-neurips-2026-154485","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154488","title":"How Accurately Can a Gaussian Approximate Stochastic Approximation Iterates?","authors":["Shaan U Haque","Zedong Wang","Zixuan Zhang","Siva Theja Maguluri"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"How Accurately Can a Gaussian Approximate Stochastic Approximation Iterates?","plainExplanation":"The scalar variance recurrence contradicts these target normalizations.","primaryUrl":"https://arxiv.org/html/2602.13906","noteUrl":"/review?paper=review-neurips-2026-154488","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154490","title":"A meshfree exterior calculus for generalizable and data-efficient learning of physics from point clouds","authors":["Benjamin Shaffer","Brooks Kinch","M. Ani Hsieh","Nathaniel Trask"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Signed stencil remainders require absolute weights","plainExplanation":"The meshfree Taylor-error bound sums stencil weights without absolute values although the stated quadratic program permits negative weights. The triangle inequality controls the remainder by \\(\\sum_j|a_j|\\,|r_j|\\), not \\(\\sum_ja_j|r_j|\\).","primaryUrl":"https://arxiv.org/html/2605.08436","noteUrl":"/review?paper=review-neurips-2026-154490","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154491","title":"Reflection with Action-Induced Visual Differences for Desktop GUI Agents","authors":["Yijie Ma","Chaoyue Niu","Fan Wu","Guihai Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Reflection with Action-Induced Visual Differences for Desktop GUI Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.24015","noteUrl":"/review?paper=review-neurips-2026-154491","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154492","title":"Efficient Dynamic Algorithms for Graph Neural Networks with Non-Linearity","authors":["Kiarash Banihashem","MohammadTaghi Hajiaghayi","Mahdi JafariRaviz","Silvio Lattanzi","Danny Mittal"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Efficient Dynamic Algorithms for Graph Neural Networks with Non-Linearity","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.32929","noteUrl":"/review?paper=review-neurips-2026-154492","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154497","title":"The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching","authors":["Sankaran Vaidyanathan","David Arbour","Aaron Mueller","Scott Niekum","David Jensen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.27510","noteUrl":"/review?paper=review-neurips-2026-154497","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154499","title":"Data-Free Reservoir Features for Efficient Long-Horizon Cold-Start Continual Learning","authors":["Augustinas Jučas","Yangchen Pan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Data-Free Reservoir Features for Efficient Long-Horizon Cold-Start Continual Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.27095","noteUrl":"/review?paper=review-neurips-2026-154499","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154500","title":"State-of-art minibatches via novel DPP kernels: discretization, wavelets, and rough objectives","authors":["Hoang Son Tran","Pranav Gupta","Rémi Bardenet","Subhroshekhar Ghosh"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"State-of-art minibatches via novel DPP kernels: discretization, wavelets, and rough objectives","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13127","noteUrl":"/review?paper=review-neurips-2026-154500","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154501","title":"Flexformer: Flexible Linear Transformer with Learnable Attention Kernel","authors":["Haoran Zhang","Zisu Dong","Feng Zhou"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The Fourier feature map is smaller than the claimed kernel by a factor of four","plainExplanation":"For \\(d=n=\\sigma=1\\), \\(\\tau=\\log2\\), and both frequencies one, the printed map is \\(e^{x^2/2}(\\cos x,\\sin x)/2\\). Query zero against keys zero and \\(\\pi\\) gives opposite-sign similarities. Normalizing their sum therefore leaves a negative weight, whose real logarithm is undefined. Positive semidefiniteness does not require nonnegative off-diagonal entries. The stated positive-kernel prerequisite is not ensured by the learned real frequencies. No implementation clipping or actual failure is inferred.","primaryUrl":"https://arxiv.org/html/2606.27748","noteUrl":"/review?paper=review-neurips-2026-154501","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154502","title":"Two is better than one: A Collapse-free Multi-Reward RLIF Training Framework","authors":["Shourov Joarder","Diganta Sikdar","Ahsan H Akash","Binod Bhattarai","Prashnna K Gyawali"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Two is better than one: A Collapse-free Multi-Reward RLIF Training Framework","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.22620","noteUrl":"/review?paper=review-neurips-2026-154502","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154503","title":"UniCon3R: Unified Contact-aware 4D Human-Scene Reconstruction from Monocular Video","authors":["TANUJ SUR","Shashank Tripathi","Nikos Athanasiou","Ha Linh Nguyen","Kai Xu","Michael Black","Angela Yao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"UniCon3R: Unified Contact-aware 4D Human-Scene Reconstruction from Monocular Video","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.19923","noteUrl":"/review?paper=review-neurips-2026-154503","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154504","title":"idSCD: Identifying Training Datasets through Semantic Correlation Descriptors","authors":["Ionuț Hodoroagă","Andrada Gobeajă","Marius Leordeanu","Elena Burceanu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"idSCD: Identifying Training Datasets through Semantic Correlation Descriptors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30462","noteUrl":"/review?paper=review-neurips-2026-154504","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154505","title":"Parallel Rollout Approximation for Pixel-Space Autoregressive Image Generation","authors":["Jiayi Xu","Di He","Guolin Ke"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Parallel Rollout Approximation for Pixel-Space Autoregressive Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.27978","noteUrl":"/review?paper=review-neurips-2026-154505","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154506","title":"MemReg: Streaming Outdoor LiDAR Point Cloud Registration with Hybrid Memory Buffers","authors":["Jiuming Liu","Mengmeng Liu","Guangming Wang","Lihao Liu","Michael Yang","Yunpeng Zhang","Francesco Nex","Hao Cheng","Hesheng Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The rotation angle subtracts the identity inside the trace","plainExplanation":"The rotation-error formula places the subtraction in \\(\\operatorname{Tr}(Q-I)/2\\) rather than using the standard \\((\\operatorname{Tr}Q-1)/2\\), where \\(Q\\) is a relative three-dimensional rotation. Subtracting the identity inside the trace gives \\((\\operatorname{Tr}Q-3)/2=\\cos\\theta-1\\), not \\(\\cos\\theta\\).","primaryUrl":"https://arxiv.org/html/2610.12624","noteUrl":"/review?paper=review-neurips-2026-154506","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154507","title":"Audible World Models: Spatially Aware Sound Generation for 3D Worlds","authors":["Duowen Chen","Jinjin He","Gouthaman KV","Sandeep Bangalore Venkatesh","Bo Zhu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Audible World Models: Spatially Aware Sound Generation for 3D Worlds","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38444","noteUrl":"/review?paper=review-neurips-2026-154507","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154508","title":"Neural Bayesian Filtering","authors":["Christopher Solinas","Radovan Haluška","David Sychrovský","Finbarr Timbers","Nolan Bard","Michael Buro","Martin Schmid","Nathan Sturtevant","Michael Bowling"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Clipping the residual features destroys observation recoverability","plainExplanation":"The inversion needs uncensored residual differences. With fixed distinct particles zero and one, observations \\(30\\) and \\(40\\) both saturate clipping to \\([-20,20]\\), giving identical encoder features but different observation likelihood ratios. Standardizing the unchanged particle states cannot recover the lost values. Open saturation regions also avoid a probability-zero degeneracy. The recoverability application fails for the clipped feature set. The conditional lemma and theorem under their stated sufficiency assumptions remain valid.","primaryUrl":"https://arxiv.org/html/2510.03614","noteUrl":"/review?paper=review-neurips-2026-154508","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154509","title":"Learn from Your Mistakes: Self-Correcting Masked Diffusion Models","authors":["Yair Schiff","Omer Belhasin","Roy Uziel","Guanghan Wang","Marianne Arriola","Gilad Turok","Ran Zilberstein","Miki Elad","Volodymyr Kuleshov"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Optimal tokenwise correction need not preserve the joint latent distribution","plainExplanation":"A factorized likelihood optimum recovers the individual conditional token marginals. It need not recover their joint conditional dependence, which the stationary latent distribution can require.","primaryUrl":"https://arxiv.org/html/2602.11590","noteUrl":"/review?paper=review-neurips-2026-154509","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154514","title":"MindVLM: Neural-Grounded Visual Captioning via Subject-Aware Semantic Evidence Selection","authors":["Zixiang Yin","Yu-Ping Wang","Zhengming Ding"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"MindVLM: Neural-Grounded Visual Captioning via Subject-Aware Semantic Evidence Selection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.12673","noteUrl":"/review?paper=review-neurips-2026-154514","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154518","title":"AVI-HT: Adaptive Vision-IMU Fusion for 3D Hand Tracking","authors":["Ziyi Kou","Ankit Kumar","Mia Huang","Taylor Niehues","Vatsal Mehta","Ergys Ristani","Li Guan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Sensor importance needs the sensor-column slice","plainExplanation":"The full attention row contains the visual self column as well as twelve sensor columns. Calling it a twelve-entry sensor vector omits a necessary column slice. This is an omitted slice in the importance definition. The underlying attention computation remains well-shaped; the finding does not diagnose implementation behavior.","primaryUrl":"https://arxiv.org/html/2605.21714","noteUrl":"/review?paper=review-neurips-2026-154518","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-154520","title":"AMGenC: Generating Charge Balanced Amorphous Materials","authors":["Yan Lin","Jilin Hu","N M Anoop Krishnan","Morten Smedskjaer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AMGenC: Generating Charge Balanced Amorphous Materials","plainExplanation":"The soft projection is approximate; the final discrete correction provides exact charge balance. No proof, formula, or statistical error was identified in the material examined. Equivalence with the accepted conference manuscript has not been established.","primaryUrl":"https://arxiv.org/html/2604.27613","noteUrl":"/review?paper=review-neurips-2026-154520","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154521","title":"Gaussian Splatting-based Volumetric Video Compression with Sparse 4D Anchors","authors":["Ge Gao","Siyue Teng","Changqi Wang","Fan Zhang","Nantheera Anantrasirichai","Jui-Chiu Chiang","Wen-Hsiao Peng","David Bull"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Gaussian Splatting-based Volumetric Video Compression with Sparse 4D Anchors","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.33969","noteUrl":"/review?paper=review-neurips-2026-154521","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154523","title":"LINE: LLM-based Iterative Neuron Explanations for Vision Models","authors":["Vladimir Zaigrajew","Michał Piechota","Gaspar Sekula","Paweł Gelar","Przemyslaw Prem Biecek"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strict superiority omits the half-credit contribution of AUC ties","plainExplanation":"Trapezoidal ROC AUC is \\(P(\\mathrm{control}<\\mathrm{concept})+\\tfrac12P(\\mathrm{control}=\\mathrm{concept})\\). The printed statistic includes only the first event. Nonnegative activation samples can contain ties while controls still have positive variance, so the two differ without any degenerate-variance pathology. This is a local probability identity mismatch. Actual tie prevalence and resulting benchmark ranking effects are unknown.","primaryUrl":"https://arxiv.org/html/2604.08039","noteUrl":"/review?paper=review-neurips-2026-154523","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154524","title":"What Cohort INRs Encode, and Where to Freeze Them","authors":["Vasiliki Sideri-Lampretsa","Sophie Starck","Robbie Holland","Julian McGinnis","Daniel Rueckert"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"What Cohort INRs Encode, and Where to Freeze Them","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08298","noteUrl":"/review?paper=review-neurips-2026-154524","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154526","title":"Autoregressive Appearance Prediction for 3D Gaussian Avatars","authors":["Michael Steiner","Zhang Chen","Alexander Richard","Vasu Agrawal","Markus Steinberger","Michael Zollhöfer"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Autoregressive Appearance Prediction for 3D Gaussian Avatars","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.00928","noteUrl":"/review?paper=review-neurips-2026-154526","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154528","title":"The Spectral Amplitude Principle for Dynamics of Quantum Neural Networks","authors":["Yihang Xu","Dan-Bo Zhang","Junchi Yan"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The Spectral Amplitude Principle for Dynamics of Quantum Neural Networks","plainExplanation":"The learning order depends on the Fourier Jacobian or kernel. A conditional equal-kernel result and the exact discrete residual solution give precise replacements.","primaryUrl":"https://arxiv.org/html/2409.06682","noteUrl":"/review?paper=review-neurips-2026-154528","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","incorrect_bound","invalid_inference","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-154529","title":"Intersectional Fairness via Mixed-Integer Optimization","authors":["Jiří Němeček","Mark Kozdoba","Illia Kryvoviaz","Tomáš Pevný","Jakub Marecek"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Intersectional Fairness via Mixed-Integer Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.19595","noteUrl":"/review?paper=review-neurips-2026-154529","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154531","title":"Breaking the Bias Barrier in Concave Multi-Objective Reinforcement Learning","authors":["Swetha Ganesh","Jason Chia","Vaneet Aggarwal"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Breaking the Bias Barrier in Concave Multi-Objective Reinforcement Learning","plainExplanation":"Making the scalarization estimate exact does not remove the independent trajectory variance.","primaryUrl":"https://arxiv.org/html/2603.08518","noteUrl":"/review?paper=review-neurips-2026-154531","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154532","title":"Enhancing Speech Large Language Models through Reinforced Behavior Alignment","authors":["Yansong Liu","Yuan Liu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Enhancing Speech Large Language Models through Reinforced Behavior Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2509.03526","noteUrl":"/review?paper=review-neurips-2026-154532","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154536","title":"Diversity Curves for Graph Representation Learning","authors":["Katharina Limbeck","Nadja Häusermann","Martin Carrasco","Guy Wolf","Bastian Rieck"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The normalized-Laplacian embedding can collapse distinct nodes","plainExplanation":"The heat expression is positive on the diagonal, while the proposed normalized-Laplacian diffusion embedding can map distinct nodes to the same point. Both constructions need correction before being used as metrics.","primaryUrl":"https://arxiv.org/html/2605.06466","noteUrl":"/review?paper=review-neurips-2026-154536","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154538","title":"KV Packet: Recomputation-Free Context-Independent KV Caching for LLMs","authors":["Chuangtao Chen","Grace Li Zhang","Xunzhao Yin","Cheng Zhuo","Bing Li","Ulf Schlichtmann"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"KV Packet: Recomputation-Free Context-Independent KV Caching for LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.13226","noteUrl":"/review?paper=review-neurips-2026-154538","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154539","title":"Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care","authors":["Shreyas Rajesh","Kartik Sharma","Tonmoy Monsoor","Mehmet Yigit Turali","Richard Idro","Juliana Kayaga","Robert Sebunya","Tracy T Namata","Jessica N Pasqua","vwani Roychowdhury","Rajarshi Mazumder"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.31036","noteUrl":"/review?paper=review-neurips-2026-154539","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154540","title":"FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies","authors":["Xintong Hu","Xuhong Huang","JINYU ZHANG","Yutong Yao","Yuchong Sun","Qiuyue Wang","Mingsheng Li","Yitao Liu","Yixuan Chen","Yingming Zheng","Shuai Bai","Tao Yu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27284","noteUrl":"/review?paper=review-neurips-2026-154540","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154541","title":"Recursive Language Models","authors":["Alex Zhang","Tim Kraska","Omar Khattab"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Recursive Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.24601","noteUrl":"/review?paper=review-neurips-2026-154541","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154543","title":"Can Model Merging Improve Aggregation in DiLoCo?","authors":["Stefan Horoi","Benjamin Thérien","Guy Wolf","Eugene Belilovsky"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Subtracting a descent displacement reverses the outer step","plainExplanation":"The pseudo-gradient is defined as local weights minus global weights, but the outer rule subtracts that displacement. For loss \\(\\|W\\|_F^2/2\\), initial \\(W=I_2\\), and one local step of size \\(0.1\\), the displacement is \\(-0.1I_2\\).","primaryUrl":"https://arxiv.org/html/2607.03011","noteUrl":"/review?paper=review-neurips-2026-154543","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154545","title":"Taming the Curses of Multiagency in Robust Markov Games with Large State Space through Linear Function Approximation","authors":["Jingchu Gai","Laixi Shi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Total signed mass does not bound integration against a varying function","plainExplanation":"The robust dual uses a ridge center that may not be a probability measure. The analysis also bounds signed integrals by cancelling total mass and interchanges policy-specific best responses with a mixture maximum; neither step follows from the stated assumptions.","primaryUrl":"https://arxiv.org/html/2605.03125","noteUrl":"/review?paper=review-neurips-2026-154545","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154546","title":"interwhen: A Generalizable Framework for Steering Reasoning Models with Test-time Verification","authors":["Vishak K Bhat","Prateek Chanda","Vijval Ekbote","Ashmit Khandelwal","Maitreyi Swaroop","Subbarao Kambhampati","Vineeth N Balasubramanian","Nagarajan Natarajan","Amit Sharma"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The stationary theorem needs separate degenerate-chain cases","plainExplanation":"Allowed \\(a_1=1,b_1=0\\) can make the stationary denominator zero, so the displayed ratio is undefined. With \\(\\gamma=\\delta=0\\), \\(\\beta=1,\\beta^{\\prime}=0\\), zero safe/fix rates and initial compliance one, baseline compliance remains one while the steered chain’s compliance becomes zero. Thus the zero-\\(\\gamma\\) clause is not universal. The interior calculation and separate finite-horizon results with stronger false-alarm hypotheses are not refuted.","primaryUrl":"https://arxiv.org/html/2602.11202","noteUrl":"/review?paper=review-neurips-2026-154546","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154548","title":"On the Sparsity-Storage-Accuracy Tradeoff in Parsimoniously Activated Dictionary Learning","authors":["Zihui Zhao","Yang Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"On the Sparsity-Storage-Accuracy Tradeoff in Parsimoniously Activated Dictionary Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.22352","noteUrl":"/review?paper=review-neurips-2026-154548","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154549","title":"Magnifying What Matters: Attention-Guided Adaptive Rendering for Visual Text Comprehension","authors":["Shenglai Zeng","Qirui Wang","Kai Guo","Xinnan Dai","Xianxuan Long","Hui Liu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Magnifying What Matters: Attention-Guided Adaptive Rendering for Visual Text Comprehension","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.12898","noteUrl":"/review?paper=review-neurips-2026-154549","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154550","title":"Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics","authors":["Maciej Chrabaszcz","Aleksander Szymczyk","Marcin Sendera","Tomasz Trzcinski","Sebastian Cygert"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18549","noteUrl":"/review?paper=review-neurips-2026-154550","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154551","title":"Even Sharper Bounds for Transductive Learning and Its Applications","authors":["Yingzhen Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Even Sharper Bounds for Transductive Learning and Its Applications","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.28459","noteUrl":"/review?paper=review-neurips-2026-154551","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154552","title":"Retrieve-then-Steer: Online Success Memory for Test-Time Adaptation of Generative VLAs","authors":["Jianchao Zhao","Huoren Yang","Yusong Hu","Yuyang Gao","Qiguan Ou","Cong Wan","SongLin Dong","Zhiheng Ma","Yihong Gong"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Fractional critic labels are accumulated with percentage units","plainExplanation":"On a successful trajectory ending at time \\(100\\), intervals \\(0\\to50\\) and \\(50\\to75\\) each have label \\(c=1/2\\). Fractional accumulation gives \\(1-v=\\prod(1-c)\\), hence \\(v=0.75\\). The printed recurrence gives \\(1-v/100=\\prod(1-c/100)\\), hence \\(v=0.9975\\), above threshold \\(0.95\\). A single interval \\(0\\to75\\) instead gives \\(0.75\\), below threshold, despite identical final completion. The manuscript units are inconsistent even for ideal labels. Actual robot wrapper conversions and runs are not verified.","primaryUrl":"https://arxiv.org/html/2605.10094","noteUrl":"/review?paper=review-neurips-2026-154552","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154553","title":"When Can Digital Personas Reliably Approximate Human Survey Findings?","authors":["Mumin Jia","Yilin Chen","Divya Sharma","Jairo Diaz-Rodriguez"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Global MMD compares profile distributions without preserving respondent identity","plainExplanation":"The displayed statistic discards the pairing between a real person and that person’s digital persona.","primaryUrl":"https://arxiv.org/html/2605.10659","noteUrl":"/review?paper=review-neurips-2026-154553","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154554","title":"Robust Nash Alignment under Preference Uncertainty","authors":["Shihab Ahmed","Debamita Ghosh","David Tang","Yudan Wang","Alvaro Velasquez","Yue Wang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The optimistic regret inequality fails at the first step","plainExplanation":"Lemma 2 includes a negative consecutive-iterate KL term in its optimistic regret bound without sufficient compensating initial terms. Take one round with comparator equal to the initial uniform policy.","primaryUrl":"https://arxiv.org/html/2610.00715","noteUrl":"/review?paper=review-neurips-2026-154554","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154556","title":"Monroe: A Molecular Foundation Model for In-context Probabilistic Inference","authors":["Błażej Banaszewski","Andrew Fitzgibbon"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Coordinatewise Huber loss is not globally rotation-invariant after alignment","plainExplanation":"Coordinatewise Huber penalties lose global-frame invariance outside their quadratic regime even after Kabsch alignment. Adding constants inside jointly learned uncertainty weights also changes task weighting and training dynamics.","primaryUrl":"https://arxiv.org/html/2608.18982","noteUrl":"/review?paper=review-neurips-2026-154556","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154557","title":"CORTEG: Foundation Models Enable Cross-Modality Representation Transfer from Scalp to Intracranial Brain Recordings","authors":["Liuyin Yang","Qiang Sun","Bob Van Dyck","Eva C Merino","Marc M Van Hulle"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"CORTEG: Foundation Models Enable Cross-Modality Representation Transfer from Scalp to Intracranial Brain Recordings","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.10337","noteUrl":"/review?paper=review-neurips-2026-154557","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154560","title":"ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence","authors":["Yiwei Chen","Amila Weerasinghe","Siddharth Biswal","Utkrisht Rajkumar","Olcay Boz","Taylor Berg-Kirkpatrick","Rose Yu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.14381","noteUrl":"/review?paper=review-neurips-2026-154560","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154563","title":"ElasticFit: Fit-Aware 3D Object Insertion via VLM Reasoning and Generative Adaptation","authors":["Tzu-Hsin Hsieh","Ricardo Marroquim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"ElasticFit: Fit-Aware 3D Object Insertion via VLM Reasoning and Generative Adaptation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07460","noteUrl":"/review?paper=review-neurips-2026-154563","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154564","title":"A Theory of Online Learning with Autoregressive Chain-of-Thought Reasoning","authors":["Idan Mehalel","Ilan Doron-Arad","Elchanan Mossel"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A support-cardinality bound does not separate shifted lengths","plainExplanation":"The gluing proof uses a bound on the number of supported strings as if it bounded their maximum length, then chooses fixed shifts to separate the supports. A finite support of small cardinality may contain strings of arbitrarily large length.","primaryUrl":"https://arxiv.org/html/2605.06819","noteUrl":"/review?paper=review-neurips-2026-154564","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154568","title":"Half-Truths Break Similarity-Based Retrieval","authors":["Bora Kargi","Arnas Uselis","Seong Joon Oh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Half-Truths Break Similarity-Based Retrieval","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.23906","noteUrl":"/review?paper=review-neurips-2026-154568","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154570","title":"METRO: Metric-Enhanced Token Routing Operator","authors":["Nodens Koren","Thomas Hofmann","Georgios Kissas"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"METRO: Metric-Enhanced Token Routing Operator","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.05100","noteUrl":"/review?paper=review-neurips-2026-154570","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154575","title":"Heavy-Tailed Flow Matching via Random Clocks","authors":["Zhouhao Yang","Yezhen Wang","Kenji Kawaguchi","Vladimir Braverman","Haoyang Cao"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Heavy-Tailed Flow Matching via Random Clocks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.13841","noteUrl":"/review?paper=review-neurips-2026-154575","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154576","title":"Multi-Armed Bandits With Best-Action Queries","authors":["Francesco Bacchiocchi","Matteo Castiglioni","Alberto Marchesi","Francesco Emanuele Stradi"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The unrestricted bandit lower bound can exceed the maximum possible regret","plainExplanation":"The stated lower bound can exceed the horizon-wide maximum regret when the arm count is unrestricted. It needs a horizon cap or an explicit parameter range.","primaryUrl":"https://arxiv.org/html/2605.08287","noteUrl":"/review?paper=review-neurips-2026-154576","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154577","title":"Beckmann Transport Models: From Autonomous Flows to One-Step Maps","authors":["Lee Kit","Florentin Coeurdoux","Peter Potaptchik","Yilun Du","Michael Albergo","Eric Vanden-Eijnden"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Beckmann Transport Models: From Autonomous Flows to One-Step Maps","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.01692","noteUrl":"/review?paper=review-neurips-2026-154577","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154580","title":"Learning What Not to Impute: An Uncertainty-Aware Diffusion Framework for Meaningful Missingness","authors":["Lixing Zhang","Yidong Ouyang","Weifu Li","Guang Cheng","Shixiang Zhu","Liyan Xie"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Observed-support identification does not identify a mechanism everywhere","plainExplanation":"The missing-data identifiability statement infers equality of global mechanism functions from equality of the observed complete-data law, without a covariate-support rank condition. Let one covariate be identically zero and consider logistic mechanisms \\(\\phi_a(z;1)=\\sigma(az)\\).","primaryUrl":"https://arxiv.org/html/2606.05073","noteUrl":"/review?paper=review-neurips-2026-154580","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154584","title":"INEUS: Iterative Neural Solver for High-Dimensional PIDEs","authors":["Jean-Loup Dupret","Davide Gallon","Patrick Cheridito"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"INEUS: Iterative Neural Solver for High-Dimensional PIDEs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06281","noteUrl":"/review?paper=review-neurips-2026-154584","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154586","title":"High-Dimensional Learning Dynamics of Attention-Indexed Models","authors":["Yizhou Xu","Margarita Sagitova","Florent Krzakala","Lenka Zdeborová"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"High-Dimensional Learning Dynamics of Attention-Indexed Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.03858","noteUrl":"/review?paper=review-neurips-2026-154586","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154588","title":"Signed-Permutation Coordinate Transport for RMSNorm Transformers","authors":["John Sweeney"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Signed-Permutation Coordinate Transport for RMSNorm Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.31963","noteUrl":"/review?paper=review-neurips-2026-154588","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154590","title":"Transformers Provably Implement In-Context Reinforcement Learning with Policy Improvement","authors":["Haodong Liang","Lifeng LAI"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Transformers Provably Implement In-Context Reinforcement Learning with Policy Improvement","plainExplanation":"The convergence theorem’s premises cannot all hold as written. This is an empty-premise problem, while the separate explicit transformer constructions remain outside the contradiction.","primaryUrl":"https://arxiv.org/html/2605.05755","noteUrl":"/review?paper=review-neurips-2026-154590","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154592","title":"An Equivariance Principle for Optimizer Design: Symmetry-Compatible Updates for Embeddings, LM Heads, and MoE Routers","authors":["Tim Tsz-Kit Lau","Weijie Su"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The equivariance proposition uses one untransformed objective twice","plainExplanation":"For orthogonal \\(P,Q\\), the chain rule concerns \\(\\widetilde f(V)=f(P^\\top VQ)\\), giving \\(\\nabla\\widetilde f(PWQ^\\top)=P\\nabla f(W)Q^\\top\\). With scalar \\(f(w)=(w-1)^2/2,w=0.3,P=-1,Q=1\\), the same untransformed objective instead has gradient \\(-1.3\\), not transformed gradient \\(0.7\\). Zero initial momentum carries the mismatch into the first Nesterov update. The local statement and proof conventions disagree. Momentum equivariance as a map of an equivariantly transformed gradient sequence and the broader design principle remain intact.","primaryUrl":"https://arxiv.org/html/2605.18106v5","noteUrl":"/review?paper=review-neurips-2026-154592","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154598","title":"How Neural Reward Models Learn Features for Policy Optimization: A Single-Index Analysis","authors":["Rei Higuchi","Ryotaro Kawata","Akifumi Wachi","Shokichi Takakura","Kohei Miyaguchi","Taiji Suzuki"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"How Neural Reward Models Learn Features for Policy Optimization: A Single-Index Analysis","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24749","noteUrl":"/review?paper=review-neurips-2026-154598","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154602","title":"Inverted Detection and Control in Steering Vectors","authors":["Max Torop","Aria Masoomi","Jennifer Dy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Inverted Detection and Control in Steering Vectors","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.02957","noteUrl":"/review?paper=review-neurips-2026-154602","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154603","title":"CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models","authors":["Chengqian Zhang","Yucheng Jin","Duo Zhang","Tiejun Li","Han Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08960","noteUrl":"/review?paper=review-neurips-2026-154603","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154605","title":"Personal Visual Memory from Explicit and Implicit Evidence","authors":["Viet Nguyen","Thao Nguyen","Vishal Patel","Yuheng Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Personal Visual Memory from Explicit and Implicit Evidence","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.28806","noteUrl":"/review?paper=review-neurips-2026-154605","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154608","title":"Pose-Free Feed-Forward 3D Inpainting via Learnable Mask Attention and Support Token Refinement","authors":["Jingyi Pan","Dan Xu","Qiong Luo"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Pose-Free Feed-Forward 3D Inpainting via Learnable Mask Attention and Support Token Refinement","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.11857","noteUrl":"/review?paper=review-neurips-2026-154608","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154610","title":"Black-Box Inference of LLM Architectural Properties with Restrictive API Access","authors":["Christopher Ellis","Mei-Yu Wang","Shreyas Chaudhari","Leighton P Barnes","Giulia Fanti","José M F Moura"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Black-Box Inference of LLM Architectural Properties with Restrictive API Access","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.01313","noteUrl":"/review?paper=review-neurips-2026-154610","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154611","title":"Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge","authors":["Wonjun Lee","Kyungsik Yang","Gaeun Ji","Vaidehi Patil","Haon Park","Bumsub Ham","Mohit Bansal","Suhyun Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Top-one normalization destroys the scalar probability signal","plainExplanation":"The top-k ablation selects a fixed set of probability coordinates and normalizes each selected vector by its own sum before computing distances. At \\(k=1\\), every positive selected probability becomes the scalar \\(1\\).","primaryUrl":"https://arxiv.org/html/2610.07532","noteUrl":"/review?paper=review-neurips-2026-154611","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154613","title":"Enhancing LLMs with Cognitive-Affective Personality Inference for Simulating Human Social-Psychological Behavior","authors":["Zhibo Deng","Dongyuan Li","Shuwen Ge","Ziqing Zhang","Ying Zhang","Renhe Jiang"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Enhancing LLMs with Cognitive-Affective Personality Inference for Simulating Human Social-Psychological Behavior","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.08328","noteUrl":"/review?paper=review-neurips-2026-154613","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154615","title":"Trainable Sparse Attention via Hybrid Top-k+Top-p Masking and Distillation Fine-Tuning","authors":["Jintao Zhang","Kai Jiang","Chendong Xiang","Weiqi Feng","Yuezhou Hu","Haocheng Xi","Shuo Yang","Mengfei Xia","Jianfei Chen","Jun Zhu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A more concentrated attention distribution can have a larger output error","plainExplanation":"The decomposition is correct, but the claimed monotonicity does not follow. Value vectors can have opposite signs, so reducing the dropped probability mass can remove cancellation and increase the value-weighted error.","primaryUrl":"https://arxiv.org/html/2602.13515","noteUrl":"/review?paper=review-neurips-2026-154615","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154616","title":"Marrying Optimal Transport and ODEs for Unified Continuous-Time 4D Reconstruction and Tracking","authors":["Liying Yang","Hao Mo","Jialun Liu","Chen Liu","Xinxing Yu","Chenhao Guan","Hui Ma","Xiao Cao","Ajian Liu","Yanyan Liang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Straight paths minimize pairwise action only after a coupling is chosen","plainExplanation":"For equally weighted sources \\((-1,0),(1,0)\\) and ordered targets \\((1,1),(-1,-1)\\), straight index-paired paths cost five, while swapping targets costs one. The paired paths are \\(\\pm m(t)\\), \\(m(t)=(-1+2t,t)\\), never zero. Field \\(v(x,t)=(2,1)(m(t)^\\top x)/\\|m(t)\\|^2\\) realizes both paired velocities exactly, so zero conditional flow-matching loss does not force an optimal pairing. Conditional straight-path minimality remains correct. The missing coupling optimum, rather than interpolation or empirical quality, is the issue.","primaryUrl":"https://arxiv.org/html/2608.09613","noteUrl":"/review?paper=review-neurips-2026-154616","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154617","title":"DiM$^3$: Bridging Multilingual and Multimodal Models via Direction- and Magnitude-Aware Merging","authors":["Zijing Wang","Mingyang Wang","Ercong Nie","Yongkang Liu","Shi Feng","Mengjie Zhao","Daling Wang","Xiaocui Yang","Hinrich Schuetze"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DiM$^3$: Bridging Multilingual and Multimodal Models via Direction- and Magnitude-Aware Merging","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12960","noteUrl":"/review?paper=review-neurips-2026-154617","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154623","title":"Augmented Lagrangian Predictive Coding","authors":["Jeffrey Seely","Julian J Gould"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Augmented Lagrangian Predictive Coding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31022","noteUrl":"/review?paper=review-neurips-2026-154623","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154625","title":"Action Chunking Proximal Policy Optimization with Feedback Correction","authors":["Sanghyun Hahn","Jonghyun Choi"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Action Chunking Proximal Policy Optimization with Feedback Correction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36250","noteUrl":"/review?paper=review-neurips-2026-154625","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154626","title":"RePercENT: Scaling Disentangled Representation Learning Beyond Two Modalities","authors":["Vasiliki Rizou","Pascal Frossard","Dorina Thanou"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The equal-concentration vMF KL has the wrong angular sign","plainExplanation":"The appendix presents a raw directional dot product as the exact KL between equal-concentration von Mises–Fisher laws and inserts it with a negative coefficient in a maximized objective. The exact divergence is \\(\\kappa A_d(\\kappa)(1-\\mu_i^\\top\\mu_j)\\).","primaryUrl":"https://arxiv.org/html/2606.05109","noteUrl":"/review?paper=review-neurips-2026-154626","collection":"neurips-2026","severity":"repairable","issueTypes":["mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154627","title":"What to Forget in Unlearning? Forget Set Curation for Language Models","authors":["Animesh Jha","Arpandeep Khatua","Youssef Allouah","Sanmi Koyejo"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"What to Forget in Unlearning? Forget Set Curation for Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.14855","noteUrl":"/review?paper=review-neurips-2026-154627","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154633","title":"Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation","authors":["Zhiheng Liu","Weiming Ren","Xiaoke Huang","Shoufa Chen","Tianhong Li","Mengzhao Chen","Yatai Ji","Sen He","Jonas Schult","Belinda Zeng","Tao Xiang","Wenhu Chen","Ping Luo","Luke Zettlemoyer","Yuren Cong"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.24763","noteUrl":"/review?paper=review-neurips-2026-154633","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154634","title":"Learning Compositional Latent Structure with Vector Networks","authors":["Niclas Pokel","Benjamin F. Grewe"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Learning Compositional Latent Structure with Vector Networks","plainExplanation":"A large valid penalty deletes genuine support even in noiseless data.","primaryUrl":"https://arxiv.org/html/2605.28007","noteUrl":"/review?paper=review-neurips-2026-154634","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","interpretation_error"],"classifications":["major"]},{"slug":"review-neurips-2026-154636","title":"SLOPE: Optimistic Potential Landscape Shaping for Model-based Reinforcement Learning","authors":["Yao-Hui Li","Zeyu Wang","Xin Li","Wei Pang","Yingfang Yuan","Zhengkun Chen","Boya Zhang","Riashat Islam","Alex Lamb","Yonggang Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Mean-current and minimum-next shaping does not telescope as one potential","plainExplanation":"Let \\(A\\) be the frozen mean potential and \\(B\\) the frozen minimum. Discounted shaping sums to \\(-A(s_0)+\\sum_{t\\geq1}\\gamma^t[B(s_t)-A(s_t)]\\), leaving a policy-dependent disagreement term. In a continuing MDP, action \\(a\\) gives reward one then enters absorbing \\(u\\); action \\(b\\) gives zero then enters absorbing \\(v\\). Ensemble values \\(u=(0,4)\\), \\(v=s_0=(0,0)\\), with \\(\\gamma=0.95\\), give shaped value \\(1-0.95\\cdot2/(1-0.95)=-37\\) for \\(a\\), versus zero for \\(b\\), reversing the original optimum. The Appendix D variant does not inherit the common-potential guarantee. The PBRS proposition and actual repository behavior are not refuted.","primaryUrl":"https://arxiv.org/html/2602.03201","noteUrl":"/review?paper=review-neurips-2026-154636","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154637","title":"Mechanistic Circuit Identification for Controllable Data Generation","authors":["Nakyung Lee","Sangwoo Hong","Jungwoo Lee"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Mechanistic Circuit Identification for Controllable Data Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.24065","noteUrl":"/review?paper=review-neurips-2026-154637","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154638","title":"City-RAG: Stepping Into a City via Spatially-Grounded Video Generation","authors":["Gene Chou","Charles Herrmann","Kyle Genova","Boyang Deng","Songyou Peng","Bharath Hariharan","Jason Y. Zhang","Noah Snavely","Philipp Henzler"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"City-RAG: Stepping Into a City via Spatially-Grounded Video Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.19741","noteUrl":"/review?paper=review-neurips-2026-154638","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154640","title":"RIFLE: Removal of Image Flicker-Banding via Latent Diffusion Enhancement","authors":["Libo Zhu","Zihan Zhou","Xiaoyang Liu","Zhiyi Zhou","Weihang Zhang","Keyu Shi","Yifan Fu","Yulun Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"RIFLE: Removal of Image Flicker-Banding via Latent Diffusion Enhancement","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2509.24644v4","noteUrl":"/review?paper=review-neurips-2026-154640","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154641","title":"Graph-Based Stochastic-Power-UCT: Monte-Carlo Graph Search with Power Mean Estimation","authors":["Tung Tran","Viet Bao","Hoang Ta","Tuan Dam"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"A merged-node value is not an average of terminal evaluations","plainExplanation":"The concentration proof compares one shared physical-node estimator with several depth-specific targets and treats its terminal-depth base case as an empirical mean of leaf evaluations. A one-state, one-action process with reward \\(1\\), discount \\(1/2\\), horizon \\(2\\) and zero cutoff evaluator has only zero leaf draws.","primaryUrl":"https://arxiv.org/html/2609.19956","noteUrl":"/review?paper=review-neurips-2026-154641","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154643","title":"Rethinking Causal Action Tokenization with Conditional Annealing in Flow Matching","authors":["Chenyu Zhang","Yuhang Cao","Yingxi Lu","Daru Du","Jing Shao","Jiajun Liu","Ruoqu Chen","liu cao","Yicheng Liu","Hang Zhao","Mengdi Xu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The commitment loss subtracts vectors in different spaces","plainExplanation":"The quantizer maps \\(q\\in\\mathbb R^D\\) to a bottleneck code \\(e\\in\\mathbb R^{D'}\\), with \\(D'<D\\), then decodes \\(\\widehat q=\\phi(e)\\in\\mathbb R^D\\). The displayed commitment term subtracts a bottleneck vector from \\(\\widehat q\\), whose ambient dimension differs.","primaryUrl":"https://arxiv.org/html/2609.35469","noteUrl":"/review?paper=review-neurips-2026-154643","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154648","title":"Characterizing Trainability of Instantaneous Quantum Polynomial Circuit Born Machine","authors":["Kevin Shen","Susanne Pielawa","Vedran Dunjko","Hao Wang"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Arbitrary initialization does not give the stated derivative-variance comparison","plainExplanation":"The IQP analysis overextends global collision control to every output, a derivative-variance comparison to arbitrary initialization, and a variance formula to empty active sets. One individual four-copy moment identity also fails, although a separate weighted-square argument can preserve its intended lower bound.","primaryUrl":"https://arxiv.org/html/2602.11042","noteUrl":"/review?paper=review-neurips-2026-154648","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154650","title":"Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages","authors":["Vishnu Teja Kunde","Fatemeh Doudi","Mahdi Farahbakhsh","Dileep Kalathil","Krishna Narayanan","JF Chamberland"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.12554","noteUrl":"/review?paper=review-neurips-2026-154650","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154652","title":"Stable Long-Horizon PDE Forecasting via Latent Structured Spectral Propagators","authors":["Xiaoxiao Lu","Ye Yuan","Jiahao Shi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Stable Long-Horizon PDE Forecasting via Latent Structured Spectral Propagators","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10154","noteUrl":"/review?paper=review-neurips-2026-154652","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154653","title":"NGDB-Zoo: Towards Efficient and Scalable Neural Graph Databases Training","authors":["zhongwei xie","Jiaxin Bai","Shujie LIU","Haoyu Huang","LI Yufei","Yisen Gao","Hong Ting Tsang","Yangqiu Song"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The ready-set loop can stop with pooled operations unfinished","plainExplanation":"The scheduler retains ready operations in per-type pools but resets the loop-control ready set to newly enabled successors. Consider independent chains \\(F_1\\to B_1\\) and \\(F_2\\to B_2\\), one-operation execution capacity, and ties favoring forward operations.","primaryUrl":"https://arxiv.org/html/2602.21597","noteUrl":"/review?paper=review-neurips-2026-154653","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154657","title":"Test-Time Learning with an Evolving Library","authors":["Weijia Xu","Alessandro Sordoni","Chandan Singh","Zelalem Gero","Michel Galley","Xingdi Yuan","Jianfeng Gao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Allowed zero scores make the log contrast undefined","plainExplanation":"Consider four equally probable attempts, all with zero synthetic-test success, with a skill present on two attempts. Both conditioning events have positive probability and both score means are zero. The contrast therefore subtracts \\(\\log0\\) from \\(\\log0\\), undefined even under extended-real conventions. Repeated sampling and a threshold do not supply a positive floor. This is a manuscript domain gap. Skills need not be extracted only from successful attempts; actual code smoothing and runtime behavior are unknown.","primaryUrl":"https://arxiv.org/html/2605.14477","noteUrl":"/review?paper=review-neurips-2026-154657","collection":"neurips-2026","severity":"repairable","issueTypes":["mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154658","title":"CoQuant: Covariance-Aware Rotation for 2-bit KV Cache Quantization","authors":["Zhongzhu Zhou","Donglin Zhuang","Jisen Li","Ziyan Chen","Shuaiwen Song","Ben Athiwaratkun","Xiaoxia Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Dense Hadamard mixing prevents exclusive eigenvector placement by permutation","plainExplanation":"The post-Hadamard permutation cannot place individual eigenvectors exclusively into groups: dense mixing has already spread each eigenvector uniformly across coordinates. The grouping rationale needs correction.","primaryUrl":"https://arxiv.org/html/2605.17757v1","noteUrl":"/review?paper=review-neurips-2026-154658","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154659","title":"From Average Sensitivity to Small-Loss Regret Bounds under Random-Order Model","authors":["Shinsaku Sakaue","Yuichi Yoshida"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"From Average Sensitivity to Small-Loss Regret Bounds under Random-Order Model","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09457","noteUrl":"/review?paper=review-neurips-2026-154659","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154662","title":"Shift-Aware Identity-Guided Latent Refinement for Referring Audio–Visual Segmentation","authors":["Kun Li","Sami S Brandt","Michael Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Shift-Aware Identity-Guided Latent Refinement for Referring Audio–Visual Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.12783","noteUrl":"/review?paper=review-neurips-2026-154662","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154663","title":"Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples","authors":["Alexandra Souly","Javier Rando","Ed Chapman","Xander Davies","Burak Hasircioglu","Ezzeldin Shereen","Carlos Mougan","Vasilios Mavroudis","Erik Jones","Chris Hicks","Nicholas Carlini","Yarin Gal","Robert Kirk"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.07192","noteUrl":"/review?paper=review-neurips-2026-154663","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154665","title":"Semantic Motion Anchors: Bridging Motion and Meaning in Co-Speech Gestures","authors":["Varsha Suresh","Mohammad Mahdi Abootorabi","Mohamed Salman","M. Hamza Mughal","Christian Theobalt","Ashwin Ram","Jürgen Steimle","Vera Demberg"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Semantic Motion Anchors: Bridging Motion and Meaning in Co-Speech Gestures","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.30608","noteUrl":"/review?paper=review-neurips-2026-154665","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154666","title":"Best-of-$N$ Guidance for Test-time Diffusion Alignment","authors":["Richard Lee Kim","Yeongmin Kim","Gyuwon Sim","Taekyu Kim","Minsang Park","Il-chul Moon"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Best-of-$N$ Guidance for Test-time Diffusion Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.05108","noteUrl":"/review?paper=review-neurips-2026-154666","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154667","title":"Amortized Molecular Optimization via Group Relative Policy Optimization","authors":["Muhammad Bin Javaid","Hasham Hussain","Ashima Khanna-Reiter","Berke Kisin","Jonathan Pirnay","Alexander Mitsos","Dominik G Grimm","Martin Grohe"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Scaffold heterogeneity alone does not bias an action-independent baseline","plainExplanation":"Scaffold heterogeneity does not itself bias a score-function gradient with an action-independent global baseline. Self-inclusive batch baselines can be biased for a different finite-sample reason.","primaryUrl":"https://arxiv.org/html/2602.12162","noteUrl":"/review?paper=review-neurips-2026-154667","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154669","title":"Provable Pruning for Efficient 3D Gaussian Splatting via Coresets","authors":["Waseem Mousa","Alaa Maalouf"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Provable Pruning for Efficient 3D Gaussian Splatting via Coresets","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.02721","noteUrl":"/review?paper=review-neurips-2026-154669","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154670","title":"Online Resource Allocation With General Constraints","authors":["Eleonora Fidelia Chiefari","Francesco Emanuele Stradi","Matteo Castiglioni","Alberto Marchesi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The OGD penalty lemma omits threshold-separation dependence","plainExplanation":"The standalone OGD lemma claims a fixed cumulative-penalty lower bound for arbitrary threshold constants \\(c_1,c_2\\), although its proof later specializes them. Use learning rate \\(\\eta=0.001\\), unit gradient through step \\(4020\\), then zero, with \\(\\rho=0.5,c_1=2,c_2=2.01\\).","primaryUrl":"https://arxiv.org/html/2605.10519","noteUrl":"/review?paper=review-neurips-2026-154670","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154671","title":"UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation","authors":["Bangwei Guo","Yunhe Gao","Meng Ye","Yang Zhou","Difei Gu","Guoning Zhang","Leon Axel","Dimitris Metaxas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06938","noteUrl":"/review?paper=review-neurips-2026-154671","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154673","title":"Pure Exploration Beyond Reward Feedback: The Role of Post-Action Context","authors":["Mohammad Shahverdikondori","Amir Mohammad Abouei","Alireza Rezaeimoghadam","Negar Kiyavash"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Duplicate action columns invalidate the separator density and lower bound","plainExplanation":"The concentration proof reverses a count substitution, and the stopping threshold is not uniformly logarithmic in confidence as claimed. The separator lower bound also fails for permitted duplicate action columns, where only one arm can ever be uniquely best.","primaryUrl":"https://arxiv.org/html/2502.03061","noteUrl":"/review?paper=review-neurips-2026-154673","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154678","title":"Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate","authors":["Hao Wu","Shoucheng Song","Chang Yao","Huaiyu Wan","Youfang Lin","Kai Lv"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate","plainExplanation":"No issue was found in the checked claims. The current manuscript does not assert a monotonic-return guarantee for clipped policy updates.","primaryUrl":"https://arxiv.org/html/2608.03648","noteUrl":"/review?paper=review-neurips-2026-154678","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154679","title":"Can Entry-Wise Clipping Give Spectral Control of Stochastic Gradients?","authors":["Zitao Song","Cedar Site Bai","Zhe Zhang","Brian Bullins","David Gleich"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The Gaussian ratio rows require quantile rather than expectation rates","plainExplanation":"The Gaussian localization ratio has an infinite mean even for maximally delocalized signal directions. Its median and other fixed quantiles can still have the stated scale, so the expectation label on the ratio rows must be corrected; the convergence theorems are separate.","primaryUrl":"https://arxiv.org/html/2605.27733","noteUrl":"/review?paper=review-neurips-2026-154679","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154681","title":"From Talking Words to Sharing Thoughts: Scalable Multi-LLM Aggregation via Structured Message Passing","authors":["Niloufar Mehrabi","Sayedpedram Haeri Boroujeni","Abolfazl Razi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Resetting the prior does not prevent indirect feedback amplification","plainExplanation":"The update resets each log belief to its initial prior plus current incoming messages and is claimed to bound belief deviation uniformly over iterations. Incoming messages use full neighboring beliefs rather than cavity beliefs, so earlier evidence can return through a two-hop loop.","primaryUrl":"https://arxiv.org/html/2606.00405","noteUrl":"/review?paper=review-neurips-2026-154681","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154682","title":"In-Context Multi-Operator Learning with DeepOSets","authors":["Shao-Ting Chiu","Aditya Nambiar","Ali Syed","Jonathan W. Siegel","Ulisses M. Braga-Neto"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"In-Context Multi-Operator Learning with DeepOSets","plainExplanation":"No issue was found in the checked claims. Centers separated by \\(3\\delta\\) have disjoint closed \\(\\delta\\)-cores: a point in two cores would place their centers at distance at most \\(2\\delta\\). The revised separation condition therefore prevents conflicting prescriptions on overlapping cores.","primaryUrl":"https://arxiv.org/html/2512.16074","noteUrl":"/review?paper=review-neurips-2026-154682","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154683","title":"Perceive, Interact, Reason: Building Tool-Augmented Visual Agents for Spatial Reasoning","authors":["Changye Li","Meng Lu","YI WU","Ligeng Zhu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Perceive, Interact, Reason: Building Tool-Augmented Visual Agents for Spatial Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.12830","noteUrl":"/review?paper=review-neurips-2026-154683","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154688","title":"Learning Robust Reasoning through Guided Adversarial Self-Play","authors":["Shuozhe Li","Vaishnav Tadiparthi","Kwonjoon Lee","Nakul Agarwal","Hossein Nourkhiz Mahjoub","Ehsan Moradi Pari","Lizhang Chen","Du Cheng","Amy Zhang","Liu Leqi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"More successes need not increase fixed-group SNR","plainExplanation":"The fixed-group SNR need not increase with success count. The drift descriptor also needs a common PCA reference: separate centering erases all centroids, while separate uncentered fits leave their coordinate gauges unidentified.","primaryUrl":"https://arxiv.org/html/2602.00173","noteUrl":"/review?paper=review-neurips-2026-154688","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154689","title":"Learning Chance-Constrained MDPs with Bellman Distributional Certificates","authors":["Chenbei Lu","Hongyu Yi"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Learning Chance-Constrained MDPs with Bellman Distributional Certificates","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.30856","noteUrl":"/review?paper=review-neurips-2026-154689","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154690","title":"ROCKET: Rapid Optimization via Calibration-guided Knapsack-Enhanced Truncation for Efficient Model Compression","authors":["Ammar Ali","Baher Mohammad","Denis Makhov","Dmitriy Shopkhoev","Magauiya Zhussip","Stamatios Lefkimmiatis"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"ROCKET: Rapid Optimization via Calibration-guided Knapsack-Enhanced Truncation for Efficient Model Compression","plainExplanation":"The terminal inequality and dominance rule solve the opposite budget direction when costs count retained entries.","primaryUrl":"https://arxiv.org/html/2602.11008","noteUrl":"/review?paper=review-neurips-2026-154690","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154691","title":"TrAction: Action Recognition with Sparse Trajectories","authors":["Jan Frederik Meier","Felix B Mueller","Alexander Ecker","Timo Lüddecke"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TrAction: Action Recognition with Sparse Trajectories","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.03490","noteUrl":"/review?paper=review-neurips-2026-154691","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154694","title":"Embedding Security Properties into AI-Enabled Cyber-Physical Systems","authors":["Ziyan An","John Stankovic","Meiyi Ma"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Sequential rule projections need not project onto their conjunction","plainExplanation":"The reconstruction guidance identifies composition of individual rule projections with projection onto the set satisfying all rules. Let \\(A=\\{x\\geq1\\text{ or }y\\geq1\\}\\), \\(B=\\{x\\leq0\\}\\), and start at \\((1/2,0)\\).","primaryUrl":"https://arxiv.org/html/2609.38381","noteUrl":"/review?paper=review-neurips-2026-154694","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154696","title":"METIS: Multi-Source Egocentric Training for Integrated Dexterous Vision-Language-Action Model","authors":["Yankai Fu","Ning Chen","Junkai Zhao","Shaozhe Shan","Guocai Yao","Pengwei Wang","Zhongyuan Wang","Shanghang Zhang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"METIS: Multi-Source Egocentric Training for Integrated Dexterous Vision-Language-Action Model","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.17366","noteUrl":"/review?paper=review-neurips-2026-154696","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154698","title":"Taking the Road Less Scheduled with Adaptive Polyak Steps","authors":["Dimitris Oikonomou","Matthew Buchholz","Yuen-Man Pun","Robert Gower","Nicolas Loizou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Taking the Road Less Scheduled with Adaptive Polyak Steps","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.07767","noteUrl":"/review?paper=review-neurips-2026-154698","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154699","title":"HPE: Hallucinated Positive Entanglement for Backdoor Attacks in Federated Self-Supervised Learning","authors":["Jiayao Wang","Yang Song","Zhendong Zhao","Jiale Zhang","Xiaoying Lei","Qilin Wu","Wenliang Yuan","Junwu Zhu","Dongfang Zhao"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Mixed-objective convergence does not imply zero benign regret","plainExplanation":"Optimizing a fixed attack/benign mixture can have zero mixed regret while accumulating linear benign regret. The theorem changes comparator without an assumption aligning the two optima.","primaryUrl":"https://arxiv.org/html/2602.02147","noteUrl":"/review?paper=review-neurips-2026-154699","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154700","title":"Adaptive auditing of AI systems with anytime-valid guarantees","authors":["Siyu Zhou","Patrick Vossler","Venkatesh Sivaraman","Yifan Mai","Jean Feng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Adaptive auditing of AI systems with anytime-valid guarantees","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07002","noteUrl":"/review?paper=review-neurips-2026-154700","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154702","title":"Signature Approach for Contextual Bandits with Nonlinear and Path-dependent Rewards","authors":["Xin Guo","Grace He","Xinyu Li"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Raw initial-point augmentation does not give signature universality","plainExplanation":"The proposed signature augmentation cannot approximate nonlinear initial-point rewards on constant paths. The regret proof also transfers diversity through action selection without justification, uses an unclipped leverage inequality outside its range, and drops a factor in the total early-phase count.","primaryUrl":"https://arxiv.org/html/2605.10313","noteUrl":"/review?paper=review-neurips-2026-154702","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154703","title":"Flux Attention: Context-Aware Hybrid Attention for Efficient LLMs Inference","authors":["Quantong Qiu","Zhiyi Hong","Yi Yang","Haitian Wang","Kebin Liu","Qingqing Dang","Juntao Li","Min zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The quadratic dual term enforces a target rather than a sparsity interval","plainExplanation":"Let \\(d=s-t\\), with expected sparsity \\(s\\). Maximizing \\(\\lambda_1d+\\lambda_2d^2\\) over nonnegative multipliers is unbounded for every fixed \\(d\\ne0\\), including sparsities inside the intended interval. The stated max–min order also differs from interval optimization: for \\(f(s)=(s-0.1)^2\\), target \\(0.45\\), and interval \\([0.3,0.6]\\), the quadratic-penalty minimizer approaches \\(0.45\\), while the true interval optimum is \\(0.3\\). The mathematical objective does not encode the stated interval. Finite training can still yield useful routing without proving that equivalence.","primaryUrl":"https://arxiv.org/html/2604.07394","noteUrl":"/review?paper=review-neurips-2026-154703","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154705","title":"Universal Cross-Prompt Adversarial Attacks on Promptable Concept Segmentation","authors":["Ziqi Zhou","Haowen Jiang","Yifan Hu","Xianlong Wang","Yufei Song","Shengshan Hu","Dezhong Yao","Leo Yu Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Universal Cross-Prompt Adversarial Attacks on Promptable Concept Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39265","noteUrl":"/review?paper=review-neurips-2026-154705","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154709","title":"Optimize Once, Execute Fast: Latency-Aware Multi-Agent Workflow Learning for Recurrent Queries","authors":["Enpei Zhang","Feiyu Qu","Zheng Huang","Dawei Zhou","Elynn Chen","Yujun Yan"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Adaptive pool conditioning breaks the independent-error factorization","plainExplanation":"The beam-search bound conditions on the generated candidate pool and then multiplies independent classification-error probabilities as if pool membership were fixed independently of earlier errors. A beam-one chain of tasks with durations \\(1,2,4\\) admits a two-round search in which reaching the later pool already certifies rejection of an earlier top valid candidate.","primaryUrl":"https://arxiv.org/html/2610.14481","noteUrl":"/review?paper=review-neurips-2026-154709","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154711","title":"From I/O to Code with Discovery Agent","authors":["Yihong Dong","Jiaru Qian","Haoran Zhang","Peixu Wang","Binhua Li","Zhi Jin","Yongbin Li","Ge Li","Xiaokang Yang","Xue Jiang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"From I/O to Code with Discovery Agent","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.15334","noteUrl":"/review?paper=review-neurips-2026-154711","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154712","title":"Playing ZendoWorld: Challenging AI Agents on Active Visual Concept Induction","authors":["Sophia Koehler","Antonia Wüst","Inga Ibs","Top Piriyakulkij","Wolfgang Stammer","Constantin Rothkopf","Kevin Ellis","Kristian Kersting"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The DSL can express equal positive red and blue counts without the omitted primitive","plainExplanation":"The recursive DSL can express equal red/blue counts by combining its existing exact-count predicates. Removing SAME_AMOUNT increases synthesis difficulty but does not make the bounded-scene target unrepresentable.","primaryUrl":"https://arxiv.org/html/2607.08233","noteUrl":"/review?paper=review-neurips-2026-154712","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154716","title":"Continuous Latent Diffusion Language Model","authors":["Hongcan Guo","Qinyu Zhao","Yian Zhao","Shen Nie","Rui Zhu","Qiushan Guo","Feng Wang","Tao Yang","Hengshuang Zhao","Guoqiang Wei","Yan Zeng"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The best mismatch and inference gap cannot be minimized independently","plainExplanation":"The population criterion combines separately optimal mismatch and inference-gap terms that share model parameters. Their optima need not coincide, so the stated if-and-only-if ELBO comparison can select an unattainable risk.","primaryUrl":"https://arxiv.org/html/2605.06548","noteUrl":"/review?paper=review-neurips-2026-154716","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154718","title":"Robust Domain Generalization under Divergent Marginal and Conditional Distributions","authors":["Jewon Yeom","Kyubyung Chae","Hyunggyu Lim","Yoonna Oh","Dongyoon Yang","Taesup Kim"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Class conditioning changes domain mixture weights","plainExplanation":"The transport proof conditions a domain mixture on class \\(c\\) but keeps the original domain weights \\(\\omega_j\\). The correct weights are \\(\\omega_j\\pi_j(c)/\\sum_k\\omega_k\\pi_k(c)\\).","primaryUrl":"https://arxiv.org/html/2602.02015","noteUrl":"/review?paper=review-neurips-2026-154718","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154720","title":"Few-Step Boltzmann Generators via Scalable Likelihood Flow Maps","authors":["RuiKang OuYang","Hanlin Yu","Xinyue Ai","Yutong He","Nicholas Boffi","Pradeep Ravikumar","José Miguel Hernández-Lobato","Max Simchowitz","Benjamin K Miller","Omar Chehab"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The squared-loss optimum has a negative Jacobian target","plainExplanation":"Proposition 5 assigns the positive Jacobian target to a loss whose residual requires its negative. The appendix and algorithm support the local minus-sign repair.","primaryUrl":"https://arxiv.org/html/2606.29110","noteUrl":"/review?paper=review-neurips-2026-154720","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154722","title":"Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters","authors":["Hugo Cazaux","Eyjolfur Asgeirsson","Hlynur Stefansson"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06032","noteUrl":"/review?paper=review-neurips-2026-154722","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154723","title":"Learning Multimodal One-step Flow Policy via Value-weighted Optimal Transport","authors":["Jaehun Shon","Jinha Choi","Jongwook Jeon","Jongmin Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Rowwise hard extraction can collapse a valid multimodal soft plan","plainExplanation":"Hard rowwise argmax can send every student to one mode even when the unique soft OT plan allocates positive mass to multiple modes. The soft marginal constraints do not establish the claimed hard-target mode preservation.","primaryUrl":"https://arxiv.org/html/2609.15883","noteUrl":"/review?paper=review-neurips-2026-154723","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154724","title":"Training-Free Task Vectors for LLM Behavioral Control","authors":["Gabriel J Perin","Lucas Boscaini","Andre Araujo","Nina S T Hirata"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Training-Free Task Vectors for LLM Behavioral Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.09054","noteUrl":"/review?paper=review-neurips-2026-154724","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154725","title":"Fast-dLLM++: Fr\\'{e}chet Profile Decoding for Faster Diffusion LLM Inference","authors":["Siva Rajesh Kasa","Yasong Dai","Sumit Negi","Hongdong Li"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Fast-dLLM++: Fr\\'{e}chet Profile Decoding for Faster Diffusion LLM Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.02955v2","noteUrl":"/review?paper=review-neurips-2026-154725","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154726","title":"On the Error Correcting Effects of Stochasticity in Discrete Diffusion","authors":["William Yuan","Sungwon Jeong","Amirali Aghazadeh"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Matching marginals does not identify the minimal-flow CTMC with DDIM","plainExplanation":"The minimal-flow DPF CTMC is not generally the DDIM posterior-mixture channel used for contraction. The discrete update also has a probability-mass sign error, while restart pseudocode restores the wrong state and uses an identity corruption call.","primaryUrl":"https://arxiv.org/html/2605.26582","noteUrl":"/review?paper=review-neurips-2026-154726","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154727","title":"Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants","authors":["Vincent Huang","Dami Choi","Daniel D Johnson","Sarah Schwettmann","Jacob Steinhardt"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.15712","noteUrl":"/review?paper=review-neurips-2026-154727","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154730","title":"Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion","authors":["Kunyang Li","Mubarak Shah","Yuzhang Shang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Residual full-history softmax layers retain growing caches and quadratic work","plainExplanation":"For \\(P\\) layers, \\(R\\) replacements, \\(L\\) tokens per frame, \\(H\\) heads and width \\(D\\), persistent attention state includes \\(RHD^2+2(P-R)(N-1)LHD\\) scalars. Thus any retained full-history layer gives \\(\\Theta(N)\\) memory. Its cumulative cross-frame query/key pairs are \\((P-R)L^2N(N-1)/2\\), giving \\(\\Theta(N^2)\\) work at fixed positive model factors. The evaluated replacements leave fifteen or seven such layers. The inspected public full-history allocation path likewise resizes residual cache capacity with requested sequence length. Per-hybrid-layer bounds, finite coefficient savings and useful speedups remain valid. A fixed-window code option repairs these asymptotics but is not the documented uncapped configuration; no GPU benchmark is reproduced.","primaryUrl":"https://arxiv.org/html/2605.16579","noteUrl":"/review?paper=review-neurips-2026-154730","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154731","title":"Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement","authors":["Lingyu Kong","Ruicheng Li","Ruicheng Wang","Sicheng Xu","Chengtang Yao","Jianfeng Xiang","Jiaolong Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Rounding log depth breaks exact connectivity invariance under arbitrary scale","plainExplanation":"Uniform scaling leaves \\(X/Z,Y/Z\\) fixed but adds real shift \\(D\\log s\\) before rounding log depth. Values \\((0.49,1.51)\\) round to bins \\((0,2)\\). Scaling by \\(e^{0.4/D}\\) moves them to \\((1,2)\\), changing adjacency. Their unrounded gap remains \\(1.02/D>1/D\\), yet the rounded bins become neighbors. A common real shift is not necessarily an integer lattice translation. The exact connectivity and iff claims fail. Distinct image coordinates still preserve the stated occupancy count, and approximate geometric benefits remain possible.","primaryUrl":"https://arxiv.org/html/2607.17967","noteUrl":"/review?paper=review-neurips-2026-154731","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154732","title":"Risk-Controlled Post-Processing of Decision Policies","authors":["Sunay Joshi","Tao Wang","Hamed Hassani","Edgar Dobriban"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Risk-Controlled Post-Processing of Decision Policies","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06479","noteUrl":"/review?paper=review-neurips-2026-154732","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154733","title":"RaPD: Resolution-Agnostic Pixel Diffusion via Semantics-Enriched Implicit Representations","authors":["Yanhao Ge","Shanyan Guan","Weihao Wang","Ying Tai","Mingyu You"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"RaPD: Resolution-Agnostic Pixel Diffusion via Semantics-Enriched Implicit Representations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15908","noteUrl":"/review?paper=review-neurips-2026-154733","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154735","title":"Gradient Descent’s Last Iterate is Often (slightly) Suboptimal","authors":["Guy Kornowski","Ohad Shamir"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The endpoint proof is corrected; the tie-oracle induction still needs repair","plainExplanation":"A zero first plane can keep the specified minimum-index oracle at the origin indefinitely.","primaryUrl":"https://arxiv.org/html/2604.13870","noteUrl":"/review?paper=review-neurips-2026-154735","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154737","title":"Temporal Concentration from Rollout Errors: Implicit Preference Optimization For Text-to-Video Diffusion","authors":["henglin liu","Fangyuan Kong","Jing Wang","Yizhou Lin","Nisha Huang","Chang Liu","Xintao Wang","Pengfei Wan","Kun Gai","Xiu Li"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The composed preference loss reverses its stated update direction","plainExplanation":"The appendix identifies a squared conditional velocity-regression gap with a change in the clean-sample probability induced by the flow. Let independent data and source be standard Gaussian, \\(X_t=(1-t)Y+t\\varepsilon\\), and \\(q(t)=(1-t)^2+t^2\\).","primaryUrl":"https://arxiv.org/html/2607.28058","noteUrl":"/review?paper=review-neurips-2026-154737","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154738","title":"Boosting Brain-to-Image Decoding with TRIBE v2 Data Augmentation","authors":["Yohann Benchetrit","Marlene Careil","Simon Dahan","Hubert Banville","Stéphane d'Ascoli","Jean-Remi King"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Boosting Brain-to-Image Decoding with TRIBE v2 Data Augmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06345","noteUrl":"/review?paper=review-neurips-2026-154738","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154739","title":"CommunityKV: Efficient Long-Context Decoding via Graph Partitioning","authors":["Joe McKenna","Anastasios Alexandridis","Nathan Susanj","Jing Liu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Sparse modularity graphs can have linearly many optimal communities","plainExplanation":"Sparse token graphs need not have only square-root-many communities. A permitted positive-diagonal graph has modularity optimum with one community per token, so centroid cost must retain the actual community count.","primaryUrl":"https://arxiv.org/html/2610.00418","noteUrl":"/review?paper=review-neurips-2026-154739","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154740","title":"TRACER: Token ReAssignment for Concept ERasure in Generative Recommendation","authors":["Ziheng Chen","Jiali Cheng","Zezhong Fan","Diyuan Wu","Hadi Amiri","Gabriele Tolomei","Yang Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The item-generation expansion reverses the history conditional","plainExplanation":"The history marginalization weights tokenizations of an observed item \\(j\\) by \\(\\Pr(j\\mid T)\\), although the needed weights are \\(q(T\\mid j)\\). For equally likely items with token distributions \\((0.8,0.2)\\) and \\((0.6,0.4)\\), the displayed reverse-condition weights for the first item are \\(0.8/1.4\\) and \\(0.2/0.6\\), summing to \\(0.904762\\).","primaryUrl":"https://arxiv.org/html/2606.07688","noteUrl":"/review?paper=review-neurips-2026-154740","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154742","title":"RANSAC Scoring Done Right","authors":["James Pritts","Felix Seegräber","Kevin Köser"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RANSAC Scoring Done Right","plainExplanation":"The inverse-Gamma integral fixes its exponent through the dimension of the specified likelihood. Estimating a parameter does not algebraically replace that dimension after integration.","primaryUrl":"https://arxiv.org/html/2606.27385","noteUrl":"/review?paper=review-neurips-2026-154742","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154743","title":"NeuralBES: A Differentiable, Control-Aware Emulator for Scalable Building Energy Modeling","authors":["Ting-Yu Dai","Takuya Kurihana","Wing Yee Au","YUNG WONG"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The historical chunked scan has a sequential depth bottleneck","plainExplanation":"A dependency chain across a number of chunks proportional to the horizon prevents a logarithmic parallel-depth guarantee for the specified algorithm. Parallel work inside each bounded-size chunk does not remove that chain.","primaryUrl":"https://arxiv.org/abs/2610.10459v1","noteUrl":"/review?paper=review-neurips-2026-154743","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154744","title":"Proteus: Incremental Memory Activation for Long-Context Sequence Modeling","authors":["Reza Bayat","Ali Behrouz","Vahab Mirrokni","Aaron Courville"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Gradient masking alone does not lock AdamW coordinates","plainExplanation":"The capacity schedule requires inactive coordinates to remain exactly fixed, but its concrete AdamW equation masks gradients and moments while retaining unmasked decoupled weight decay. For a never-active coordinate \\(\\theta=1\\) with initial moments zero, the mask keeps both moments zero forever.","primaryUrl":"https://arxiv.org/html/2608.16844","noteUrl":"/review?paper=review-neurips-2026-154744","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154746","title":"Principled Design of Diffusion-based Optimizers for Inverse Problems","authors":["Julio Oscanoa","Irmak Sivgin","Cagan Alkan","Daniel Ennis","John Pauly","Mert Pilanci","Shreyas Vasanawala"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Strictly increasing transformations may have zero derivative","plainExplanation":"At \\((0,0)\\), let \\(f=(x+1)^2,g=(y+1)^2\\), with transformations \\(\\phi_f(s)=(s-1)^3\\) and \\(\\phi_g(s)=s\\). Both original gradient norms are two, but \\(\\phi_f^{\\prime}(f)=0\\), so the transformed normalized first gradient is undefined. A zero-vector convention instead changes the unit-parameter update from \\((-1,-1)\\) to \\((0,-1)\\). Strict monotonicity permits isolated zero derivatives. Positive rescalings and positive-derivative transformations preserve the intended cancellation. A zero-gradient convention alone does not restore the universal theorem.","primaryUrl":"https://arxiv.org/html/2605.11506","noteUrl":"/review?paper=review-neurips-2026-154746","collection":"neurips-2026","severity":"repairable","issueTypes":["mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154747","title":"The sublevel Flood bifiltration: towards scalable 2-parameter persistent homology","authors":["Mattéo Clémot","Julie Digne","Julien Tierny"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The sublevel Flood bifiltration: towards scalable 2-parameter persistent homology","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.05441","noteUrl":"/review?paper=review-neurips-2026-154747","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154748","title":"Post-Selection Distributional Model Evaluation","authors":["Amirmohammad Farzaneh","Osvaldo Simeone"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Post-Selection Distributional Model Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.23055","noteUrl":"/review?paper=review-neurips-2026-154748","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154750","title":"On the Runway Cascade of Transformers for Language Modeling","authors":["Hunjae Lee","Corey Clark"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The common attention component can vanish under the value projection","plainExplanation":"The identity is valid, but the value projection may annihilate the common component.","primaryUrl":"https://arxiv.org/html/2601.14522","noteUrl":"/review?paper=review-neurips-2026-154750","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154751","title":"2D Spatial Reasoning with Adaptive Neural Cellular Automata","authors":["Martin Spitznagel","Janis Keuper"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"2D Spatial Reasoning with Adaptive Neural Cellular Automata","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.08518","noteUrl":"/review?paper=review-neurips-2026-154751","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154752","title":"HiFloat4 Format for End-To-End Reinforcement Learning Post-Training of Large Language Models","authors":["Hei Yi Mak","Shadan Golestan","Hoang Le","Mehran Taghian Jazi","Yunke Peng","Yaoyuan Wang","Yao Wang","JUNSONG WANG","Tianchi Hu","Fengchen He","Guipeng Hu","Tanzila Rahman","Anandharaju D Raju"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"HiFloat4 Format for End-To-End Reinforcement Learning Post-Training of Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.26515","noteUrl":"/review?paper=review-neurips-2026-154752","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154756","title":"Rethinking Entropy Allocation in LLM-based ASR: Understanding the Dynamics between Speech Encoders and Large Language Models","authors":["Yuan Xie","Jiaqi Song","Xianliang Wang","Ming Lei","Jie Gao"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Rethinking Entropy Allocation in LLM-based ASR: Understanding the Dynamics between Speech Encoders and Large Language Models","plainExplanation":"The Gaussian diagnostic metrics are qualified as relative proxies rather than exact information quantities. No proof or formula error was identified in the material examined. Equivalence with the accepted conference manuscript has not been established.","primaryUrl":"https://arxiv.org/html/2604.08003","noteUrl":"/review?paper=review-neurips-2026-154756","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154762","title":"CausalMix: Data Mixture as Causal Inference for Language Model Training","authors":["Zinan Tang","YUKUN ZHANG","Shaomian Zheng","Zhuoshi Pan","Qizhi Pei","Dingnan Jin","Jun Zhou","Yujun wang","Biqing Huang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Signed log weights do not have the claimed finite simplex optimum","plainExplanation":"Negative coefficients make the unsmoothed objective unbounded above near a zero mixture weight. Boundary KKT differentiation cannot identify the claimed unique proportions.","primaryUrl":"https://arxiv.org/html/2607.01104","noteUrl":"/review?paper=review-neurips-2026-154762","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154763","title":"The Era of Agentic Organization: Learning to Organize with Language Models","authors":["Zewen Chi","Li Dong","Qingxiu Dong","Yaru Hao","Xun Wu","Shaohan Huang","Furu Wei"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"The Era of Agentic Organization: Learning to Organize with Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.26658","noteUrl":"/review?paper=review-neurips-2026-154763","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154764","title":"Provable Joint Decontamination for Benchmarking Multiple Large Language Models","authors":["Zhenlong Liu","Hao Zeng","Hongxin Wei"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Provable Joint Decontamination for Benchmarking Multiple Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21543","noteUrl":"/review?paper=review-neurips-2026-154764","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154767","title":"A Topological Sorting Criterion for Random Causal Directed Acyclic Graphs","authors":["Alexander Reisach","Antoine Chambaz","Gilles Blanchard","Sebastian Weichwald"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"A Topological Sorting Criterion for Random Causal Directed Acyclic Graphs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06288","noteUrl":"/review?paper=review-neurips-2026-154767","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154768","title":"On the Rademacher Complexity of Graph Neural Networks: Unifying Expressivity and Geometry","authors":["Martin Carrasco","Caio Deberaldini Netto","Ehimare Okoyomon","Aneeqa Mehrab","Vahan Martirosyan","Caterina Graziani"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The generalization bound needs a loss-range assumption, and two local proof steps need correction","plainExplanation":"Covering numbers increase when the radius decreases. The printed discrete replacement therefore need not be an upper bound. Integrating over the next finer interval repairs the argument without changing the final constants.","primaryUrl":"https://arxiv.org/html/2510.10101","noteUrl":"/review?paper=review-neurips-2026-154768","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154769","title":"Ask Early, Ask Late, Ask Right: When Does Clarification Timing Matter for Long-Horizon Agents?","authors":["Anmol Gulati","Hariom Gupta","Elias Lumer","Sahil Sen","Vamse K Subbiah"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Action-count commitment does not bound remaining information value","plainExplanation":"The timing model bounds remaining value of perfect information using the fraction of already executed actions that depend on the missing dimension. Two irreversible guesses of a fair hidden bit carry rewards \\(1\\) and \\(99\\), and reveal no feedback between guesses.","primaryUrl":"https://arxiv.org/html/2605.07937","noteUrl":"/review?paper=review-neurips-2026-154769","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154770","title":"Solving Max-Cut to Global Optimality via Feasibility-Preserving Graph Neural Networks","authors":["Hao Chen","Chendi Qian","Christopher Morris","Andrea Lodi","Can Li"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The pooled deterministic factor head loses necessary pair information","plainExplanation":"The deterministic row-pooled primal head forces identical factors on a symmetric triangle and cannot approximate its optimal SDP Gram matrix. Pair-feature expressivity alone does not establish expressivity of that pooled factor head.","primaryUrl":"https://arxiv.org/html/2605.07113","noteUrl":"/review?paper=review-neurips-2026-154770","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154773","title":"SceneScaffold: Active Scene-State Construction for Unified 3D Scene Understanding","authors":["Xiangqi Li","Libo Huang","Jiarui Zhao","Weilun Feng","Chuanguang Yang","Zhulin An","Yongjun Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SceneScaffold: Active Scene-State Construction for Unified 3D Scene Understanding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33518","noteUrl":"/review?paper=review-neurips-2026-154773","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154775","title":"Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillation","authors":["Zhixuan Liu","Zhichen Dong","Yuyu Fan","Xiangtian Li","Chao Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.01091","noteUrl":"/review?paper=review-neurips-2026-154775","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154777","title":"A Solvable Model of Chain-of-Thought in In-Context Learning","authors":["Kaito Takanami","Cengiz Pehlevan"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A Solvable Model of Chain-of-Thought in In-Context Learning","plainExplanation":"No issue was found in the checked claims. The nullspace lower bound does not establish those analytic premises. It therefore does not, by itself, contradict the conditional asymptotic claim.","primaryUrl":"https://arxiv.org/html/2606.03217","noteUrl":"/review?paper=review-neurips-2026-154777","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154779","title":"When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning","authors":["Zihao Zhao","Tunyu Zhang","Haizhou Shi","Yusong Zhao","Xinxi Zhang","Hao Wang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04686","noteUrl":"/review?paper=review-neurips-2026-154779","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154780","title":"Beyond Outcome Rewards: Step-Level Self-Distilled Policy Optimization for Deep Search Agents","authors":["Haoze Wu","Chuqiao Kuang","Tianyi Zhuang","Xiaoguang Li"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Beyond Outcome Rewards: Step-Level Self-Distilled Policy Optimization for Deep Search Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.12764","noteUrl":"/review?paper=review-neurips-2026-154780","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154781","title":"A multi-scale information geometry reveals the structure of mutual information in neural populations","authors":["Simone Azeglio","Steeve Laquitaine","Ulisse Ferrari","Matthew Chalk"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The three smoothing axioms do not uniquely select Gaussian kernels","plainExplanation":"Translation invariance, rotational/reflection symmetry and the convolution-semigroup property are claimed to uniquely specify Gaussian smoothing. In one dimension, \\(Z_t=B_t+\\sum_{i=1}^{N_t}S_i\\), with Brownian motion, independent rate-one Poisson jumps and symmetric unit signs, satisfies all three properties.","primaryUrl":"https://arxiv.org/html/2605.06304","noteUrl":"/review?paper=review-neurips-2026-154781","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154783","title":"TopoFisher: Learning Topological Summary Statistics by Maximizing Fisher Information","authors":["Matteo Biagetti","Mathieu Carrière","Francesco Conti","Enrico Maria Ferrari","Sven Heydenreich","Karthik Viswanathan"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"TopoFisher: Learning Topological Summary Statistics by Maximizing Fisher Information","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07720","noteUrl":"/review?paper=review-neurips-2026-154783","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154784","title":"Calibrating LLMs with Semantic-level Reward","authors":["Fengfei Yu","Ruijia Niu","Dongxia Wu","Yian Ma","Rose Yu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A binary equality decision gives infinite BCE on mismatches","plainExplanation":"Let correct-class mass be \\(a\\) and distribute remaining mass uniformly over one hundred wrong classes. The combined mean-field objective is \\(F(a)=a+\\lambda[a\\log a+(1-a)\\log(1-(1-a)/100)]\\). For \\(\\lambda=0.2\\) and sufficiently small positive \\(a\\), its derivative is negative because the \\(a\\log a\\) term dominates the linear improvement. On the wrong-only face, \\(\\sum_w w\\log(1-w)\\) is strictly concave, so uniform wrong mass is a boundary local maximum. All-correct remains globally optimal. The counterexample concerns local degeneracy, not actual GRPO/KL training failure or the empirical curriculum’s effectiveness.","primaryUrl":"https://arxiv.org/html/2605.15588","noteUrl":"/review?paper=review-neurips-2026-154784","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154785","title":"Learning to Explore with Parameter-Space Noise: A Deep Dive into Parameter-Space Noise for Reinforcement Learning with Verifiable Rewards","authors":["Bizhe Bai","Xinyue Wang","Peng Ye","Tao Chen"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The differentiable importance weight counts the current policy twice","plainExplanation":"With frozen noisy behavior \\(q\\), one token and inactive clipping, the expectation is \\(\\sum_aq_a(p_\\theta(a)/q_a)(p_\\theta(a)/p_{\\mathrm{old}}(a))A_a=\\sum_ap_\\theta(a)^2A_a/p_{\\mathrm{old}}(a)\\). Its derivative at policy equality is twice the intended PPO derivative. Correctly transporting the old-policy expectation uses fixed weight \\(p_{\\mathrm{old}}/q\\), not another differentiable current-policy ratio. The printed objective does not equal the intended corrected surrogate. Detaching can fix the expansion-point derivative without proving full multi-epoch equivalence; a factor-two toy alone does not establish instability.","primaryUrl":"https://arxiv.org/html/2602.02555","noteUrl":"/review?paper=review-neurips-2026-154785","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154788","title":"Variance-Averse n-Step Offline Reinforcement Learning for Sparse Long-Horizon Environments","authors":["Guhyeon Kang","Minhae Kwon"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Variance-Averse n-Step Offline Reinforcement Learning for Sparse Long-Horizon Environments","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07899","noteUrl":"/review?paper=review-neurips-2026-154788","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154790","title":"Reasoning via Test-Time Instance-Level Policy Gradient in Latent Space","authors":["Hengli Li","Chenxi Li","Tong Wu","Xuekai Zhu","Yuxuan Wang","Zhaoxin Yu","Eric Hanchen Jiang","Song-Chun Zhu","Zixia Jia","Ying Nian Wu","Zilong Zheng"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A greedy action is not a Monte Carlo draw from the policy objective","plainExplanation":"A grouped prover sees both private questions and can make one answer depend on the other question, which Definition C.7 forbids after splitting. For independent two-bit questions \\(q_1,q_2\\), require answers \\((q_2,q_1)\\). A grouped strategy succeeds surely; local isolated strategies succeed at most \\(1/4\\). Pre-shared randomness only mixes deterministic local strategies and does not improve that maximum. The grouping/splitting proof step fails. This is not a disproof of a complexity-class equality or finite-model expressivity by another valid reduction.","primaryUrl":"https://arxiv.org/html/2505.13308","noteUrl":"/review?paper=review-neurips-2026-154790","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154791","title":"One-Time Soft Alignment Enables Resilient Learning without Weight Transport","authors":["Jeonghwan Cheon","Jaehyuk Bae","Se-Bum Paik"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"One-Time Soft Alignment Enables Resilient Learning without Weight Transport","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.20892","noteUrl":"/review?paper=review-neurips-2026-154791","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154792","title":"Reward Bias Substitution: Single-Axis Bias Mitigations Redirect Optimization Pressure","authors":["Max Lamparth","Daniel Fein","Andreas Haupt","Marcel Hussing","Mykel J Kochenderfer"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The omitted-coordinate conditional expectation is automatically measurable","plainExplanation":"An expectation conditioned on \\((x,\\Phi_{-i})\\) is already measurable in those variables, so merely representing it without \\(\\phi_i\\) imposes no independence requirement. With two independent Rademachers and \\(R=y_1\\), \\(\\mathbb E[R\\mid y_2]=0\\), satisfying that literal form although reward depends entirely on feature one. A fiber kernel setting \\(y_1=1\\) preserves \\(y_2\\) but increases expected reward by one; its epsilon-mixture drift is \\(\\varepsilon\\), not \\(o(\\varepsilon)\\). The written definition and equivalence are defective. The intended bias-substitution phenomena need not disappear under a corrected definition.","primaryUrl":"https://arxiv.org/html/2605.27996","noteUrl":"/review?paper=review-neurips-2026-154792","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154794","title":"Uniform-in-Time Weak Propagation of Chaos in Shallow Neural Networks","authors":["Margalit Glasgow","Joan Bruna"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The empirical interaction discrepancy enters with the opposite sign","plainExplanation":"The exact particle-discrepancy dynamics define \\(\\beta\\) as empirical minus population interaction gradient but insert it with the same sign as the interaction in the velocity. For linear activation, bounded inputs \\(\\pm1\\), zero labels and uniform initialization on \\([-1,1]\\), the mean-field particles stay fixed.","primaryUrl":"https://arxiv.org/html/2605.22010","noteUrl":"/review?paper=review-neurips-2026-154794","collection":"neurips-2026","severity":"repairable","issueTypes":["lemma_error","mathematical_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154796","title":"CoDMD: Copula-aware Distribution Matching Distillation for Fast Video Generation","authors":["Wenhu Zhang","Kun Cheng","Changyuan Wang","Shiyao Li","Yuechen Zhang","Wenbo Li","Jiajun Zha","Jingyi Zhang","Kang Zhao","Jiaya Jia"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Separated components need not give an expected-distance collapse gap","plainExplanation":"The relational argument infers a strictly positive gap in expected pairwise distance from within-component bound \\(\\varepsilon<\\Delta\\), where \\(\\Delta\\) lower-bounds cross-component distances. Let one component put equal mass at \\((-1,0),(1,0)\\), and the other be a point at \\((0,\\sqrt5/2)\\).","primaryUrl":"https://arxiv.org/html/2606.21982","noteUrl":"/review?paper=review-neurips-2026-154796","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154798","title":"Elastic Spectral State Space Models for Train-Once Budgeted Inference","authors":["Dachuan Song","Junyu Yin","Zechen Hu","Xuan Wang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Elastic Spectral State Space Models for Train-Once Budgeted Inference","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.22488","noteUrl":"/review?paper=review-neurips-2026-154798","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154799","title":"Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs","authors":["Mingyu Lee","Akshat Ramachandran","Souvik Kundu","Tushar Krishna"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The first drift update subtracts an empty-history mean","plainExplanation":"The algorithm initializes the drift history \\(\\Theta\\) as empty, computes a centered drift, and appends the new drift only afterward. The first masked-token iteration evaluates \\(\\delta-\\operatorname{mean}(\\Theta)\\) with \\(\\Theta=\\varnothing\\).","primaryUrl":"https://arxiv.org/html/2607.14107","noteUrl":"/review?paper=review-neurips-2026-154799","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154803","title":"ORCA: Hunting Compositional Failures in Text-to-Image Diffusion","authors":["Arshia Hemmat","Amirhossein Vahidi","Amitis Shidani","Mohammad V Sanian","Hesam Asadollahzadeh","Aryan Y Parast","Mohammad Lotfollahi"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"An orthonormal readout cannot realize every Gaussian conditional mean","plainExplanation":"The tightness argument assumes a unit-norm linear readout can achieve the unconstrained linear-Gaussian minimum mean-square error. Let \\(H\\sim N(0,I_2)\\) and \\(V=2H_1+\\epsilon\\), with independent unit-variance Gaussian noise.","primaryUrl":"https://arxiv.org/html/2610.09841","noteUrl":"/review?paper=review-neurips-2026-154803","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154804","title":"The Platonic Defense: Backdoor Defense for Self-Supervised Encoders in the Era of Large Scale Pre-training","authors":["Tuo Chen","Minjing Dong","Benlei Cui","Jian liu","Jie Gui"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Mixed negative sampling changes the BCE density ratio","plainExplanation":"The information interpretation uses negatives from the independent source/reference product law, but practical BCE mixes full and partial mismatch samplers. For positive law \\(P\\) and actual negative law \\(Q\\), optimal BCE energy is \\(\\log(MQ/P)\\).","primaryUrl":"https://arxiv.org/html/2606.29451","noteUrl":"/review?paper=review-neurips-2026-154804","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154807","title":"Interleaved Head Attention","authors":["Sai Surya Duvvuri","Chanakya Ekbote","Rachit Bansal","Rishabh Tiwari","Devvrit Khatri","David Brandfonbrener","Paul Liang","Inderjit Dhillon","Manzil Zaheer"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A magnitude condition does not ensure modular order sensitivity","plainExplanation":"The strictness example collapses to a linear map on repeated tokens. The synthetic modular condition also fails to ensure order sensitivity, and the MHA lower bound needs independence of matrix powers rather than full rank alone.","primaryUrl":"https://arxiv.org/html/2602.21371","noteUrl":"/review?paper=review-neurips-2026-154807","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154809","title":"LEAF: Language-EEG Aligned Foundation Model for Brain-Computer Interfaces","authors":["Muyun Jiang","Shuailei Zhang","Zhenjie Yang","Wu Mengjun","Wei Zhang","Weibang Jiang","Chenyu Liu","Zhiwei Guo","Rui Liu","Shangen Zhang","Yong Li","Yi Ding","Cuntai Guan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"LEAF: Language-EEG Aligned Foundation Model for Brain-Computer Interfaces","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2509.24302","noteUrl":"/review?paper=review-neurips-2026-154809","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154811","title":"Unified Synthesis of Compositional Speech and Sound from Free-Form Text Prompts","authors":["Yuyue Wang","Xihua Wang","Xin Cheng","Yijing Chen","Ruihua Song"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The stated semantic pooling produces five targets rather than six","plainExplanation":"The semantic supervision matches \\(K=6\\) latent positions to segment-wise means of a length-\\(750\\) source sequence with nonoverlapping width-\\(150\\) pooling. The complete blocks are \\([0,150),[150,300),[300,450),[450,600),[600,750)\\): there are exactly five.","primaryUrl":"https://arxiv.org/html/2605.28063","noteUrl":"/review?paper=review-neurips-2026-154811","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154812","title":"Greedy Alignment Principle for Optimizer Selection","authors":["Jaerin Lee","Kyoung Mu Lee"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Greedy Alignment Principle for Optimizer Selection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.06370","noteUrl":"/review?paper=review-neurips-2026-154812","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154814","title":"Unify-Agent: A Unified Multimodal Agent for World-Grounded Image Synthesis","authors":["Shuang Chen","Quanxin Shou","Hangting Chen","Yucheng Zhou","Kaituo Feng","Wenbo Hu","yifan zhang","Yunlong Lin","Wenxuan Huang","Mingyang Song","Dasen Dai","Bolin Jiang","Manyuan Zhang","Yu Cheng","Nanyun Peng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Unify-Agent: A Unified Multimodal Agent for World-Grounded Image Synthesis","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.29620","noteUrl":"/review?paper=review-neurips-2026-154814","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154816","title":"Power Distribution Bridges Sampling, Self-Reward RL, and Self-Distillation","authors":["Akiyoshi Tomihari","Issei Sato"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Power Distribution Bridges Sampling, Self-Reward RL, and Self-Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.04542","noteUrl":"/review?paper=review-neurips-2026-154816","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154817","title":"Pan-FM: A Pan-Organ Foundation Model with Saliency-Guided Masking for Missing Robustness","authors":["Qiangqiang Wu","Grace McIlvain","Zhou Yu","Junhao Wen"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Pan-FM: A Pan-Organ Foundation Model with Saliency-Guided Masking for Missing Robustness","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07055","noteUrl":"/review?paper=review-neurips-2026-154817","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154822","title":"Maxitive Donsker-Varadhan Formulation for Possibilistic Variational Inference","authors":["Jasraj Singh","Shelvia Wongso","Jeremie Houssineau","Badr-Eddine Cherief-Abdellatif"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Local curvature does not ensure a finite global inner optimizer","plainExplanation":"Let \\(A(\\theta)=\\theta^2/4\\), candidate parameter \\(\\eta=0.1\\), loss \\(\\ell(\\theta)=\\theta^2/(1+\\theta^2)\\), and prior possibility \\(\\pi(\\theta)=e^{-0.05\\theta^2}\\). The normalized prior and candidate have attained unit supremum, and the stated local curvature gap at \\(0.1\\) is positive. But the true upper inner objective is \\(0.2\\theta^2-0.05\\theta+0.0025-\\ell(\\theta)\\), tending to \\(+\\infty\\). It has no finite maximum to differentiate. The local Newton surrogate can remain meaningful. Its curvature does not justify the global optimizer or exact CBO gradient under the stated hypotheses.","primaryUrl":"https://arxiv.org/html/2511.21223","noteUrl":"/review?paper=review-neurips-2026-154822","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154825","title":"Learning What's Real: Disentangling Signals and Measurement Artifacts in Multi-Sensor Data, with Applications to Astrophysics","authors":["Pablo Mercader-Perez","Carolina Cuesta Lazaro","Daniel Muthukrishna","Jeroen Audenaert","V Villar","David W Hogg","Marc Huertas-Company","Bill Freeman"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning What's Real: Disentangling Signals and Measurement Artifacts in Multi-Sensor Data, with Applications to Astrophysics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.09787","noteUrl":"/review?paper=review-neurips-2026-154825","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154826","title":"Positive-Unlabeled Preference Optimization For Chest X-ray Report Generation","authors":["Yuta Kobayashi","Pradyun Ramesh","Muhammad Ahmed Chaudhry","Vincent Jeanselme","Judy Wawira","Sanmi Koyejo","Kathleen M Capaccione","Shalmali Joshi"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Clipping requires an activity indicator in the gradient","plainExplanation":"The PU loss is \\(\\alpha\\widehat R_p^++\\max(0,C)\\), with \\(C=\\widehat R_X^--\\alpha\\widehat R_p^-\\), but its sample-gradient derivation omits the clipping branch. At \\(C\\ne0\\), the gradient is \\(\\alpha\\nabla\\widehat R_p^++\\mathbf1_{C>0}(\\nabla\\widehat R_X^--\\alpha\\nabla\\widehat R_p^-)\\).","primaryUrl":"https://arxiv.org/html/2608.05341","noteUrl":"/review?paper=review-neurips-2026-154826","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154827","title":"Remember Your Trace: Memory-Guided Long-Horizon Agentic Framework for Consistent and Hierarchical Repository-Level Code Documentation","authors":["Suyoung Bae","Jaehoon Lee","Changkyu Choi","YunSeok Choi","Jee-Hyong Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A contradiction rate has the wrong direction for a passing verification score","plainExplanation":"Section 4.2.2 averages contradiction rate \\(c\\in[0,1]\\) with positive self-quality \\(s\\in[0,1]\\), then accepts a document when that average exceeds \\(0.9\\). The acceptance score rewards contradictions rather than consistency: averaging the contradiction rate positively makes a conflict-free document unable to reach the stated threshold.","primaryUrl":"https://arxiv.org/html/2605.14563","noteUrl":"/review?paper=review-neurips-2026-154827","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154834","title":"N-vium: Mixture-of-Exits Transformer for Accelerated Exact Generation","authors":["Aleksander Lorenc","Frédéric Berdoz","Joël Mathys","Roger Wattenhofer"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"N-vium: Mixture-of-Exits Transformer for Accelerated Exact Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13190","noteUrl":"/review?paper=review-neurips-2026-154834","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154836","title":"Horizon Adaptive Offline Policy Learning via Value Stitching","authors":["Kexin ZHENG","Xianyuan Zhan","Xintao Yan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Temperature weighting is corrected; return composition needs a bridge law","plainExplanation":"The composition identity needs this bridge distribution. An unconditioned rollout marginal weights the intermediate states incorrectly.","primaryUrl":"https://arxiv.org/html/2606.21136","noteUrl":"/review?paper=review-neurips-2026-154836","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["definition_ambiguity"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-154837","title":"Denoising Time Matters: Diverse Generation in Diffusion Language Models","authors":["jingxuan wu","Zhenglin Wan","Yuzhe YANG","Yiqiao Huang","Chubin Zhang","Xingrui Yu","Ivor Tsang","Yang You"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Denoising Time Matters: Diverse Generation in Diffusion Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.22629","noteUrl":"/review?paper=review-neurips-2026-154837","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154838","title":"Do multimodal models imagine electric sheep?","authors":["Santhosh Kumar Ramakrishnan","Carl Vondrick","Raja Giryes","Philipp Kraehenbuehl","Vladlen Koltun"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Do multimodal models imagine electric sheep?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.09693","noteUrl":"/review?paper=review-neurips-2026-154838","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154839","title":"MUNI: Multimodal Unified Latent Diffusion for Coherent Any-to-Any Generation","authors":["Kyeongmin Yeo","Yunhong Min","Minhyuk Sung"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MUNI: Multimodal Unified Latent Diffusion for Coherent Any-to-Any Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.16408","noteUrl":"/review?paper=review-neurips-2026-154839","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154843","title":"On the Convergence of Success Conditioning for Policy Optimization","authors":["Matthew Brun","Andy Sun"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"On the Convergence of Success Conditioning for Policy Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03642","noteUrl":"/review?paper=review-neurips-2026-154843","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154846","title":"Scaling Linear Mode Connectivity and Merging to Billion Parameter Pretrained Transformers","authors":["Tianyi Li","Zhiqiang Shen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Scaling Linear Mode Connectivity and Merging to Billion Parameter Pretrained Transformers","plainExplanation":"No issue was found in the checked claims. The fixed-chart counterexample does not contradict the group-level absorption identity. Reading “same parameterized form” as an additional assertion that one finite Cayley chart is closed under relative products ignores the source's explicit chart restriction.","primaryUrl":"https://arxiv.org/html/2606.23607","noteUrl":"/review?paper=review-neurips-2026-154846","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154847","title":"Super-Level-Set Regression: Conditional Quantiles via Volume Minimization","authors":["Sacha Braun","Michael Jordan","Francis Bach"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Closed density superlevels need tie handling to minimize volume","plainExplanation":"The exact highest-density-region claim identifies the whole closed density superlevel with a minimum-volume set at target coverage \\(\\tau\\). For uniform \\(Y\\) on \\([0,1]\\) and \\(\\tau=1/2\\), the largest eligible density threshold is \\(1\\).","primaryUrl":"https://arxiv.org/html/2605.06210","noteUrl":"/review?paper=review-neurips-2026-154847","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154851","title":"Risk Horizons: Structured Hypothesis Spaces for Longitudinal Clinical Prediction","authors":["Zhan Qu","Michael Färber"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The learned edge bias can suppress both positive and negative losses without separation","plainExplanation":"Equation (29) learns a type-specific bias \\(\\gamma_r\\) in a negative-sampling edge objective with score equal to minus hyperbolic distance. The learned bias decreases both positive- and negative-edge losses in the same direction, so it can drive the edge objective to zero without separating their embeddings.","primaryUrl":"https://arxiv.org/html/2602.12828","noteUrl":"/review?paper=review-neurips-2026-154851","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154852","title":"The Ringelmann Effect in Multi-Agent LLM Systems: A Scaling Law for Effective Team Size","authors":["Blaz Bertalanic","Carolina Fortuna"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The dependence metric also counts a change in mean accuracy","plainExplanation":"The conditional-correlation numerator subtracts squared initial-round accuracy from final-round cross-agent correctness, intending to isolate social dependence. Conditionally independent final indicators of mean \\(p_3\\) have cross-product expectation \\(p_3^2\\).","primaryUrl":"https://arxiv.org/html/2606.02646","noteUrl":"/review?paper=review-neurips-2026-154852","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154854","title":"AEM: Adaptive Entropy Modulation for Multi-Turn Agentic Reinforcement Learning","authors":["Haotian Zhao","Songlin Zhou","Yuxin Zhang","Stephen S Yau","Wenyu Zhang","Tianlun","Tianshu Zhu","Yifeng Huang","yucheng-zeng","Jingnan gu","Daxiang Dong","Jianmin Wu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"An approximate surprisal proxy does not imply an exact entropy-drift sign","plainExplanation":"Equation (16) asserts an exact drift-change sign from the approximate proxy relation in Equation (15); Appendix B expressly calls that proxy heuristic. The exact entropy-drift sign requires alignment with actual relative surprisal, which the heuristic group-entropy proxy does not guarantee; zero modulation also gives zero drift change.","primaryUrl":"https://arxiv.org/html/2605.00425","noteUrl":"/review?paper=review-neurips-2026-154854","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154856","title":"Enabling Preference-driven Unlearning in Few-step Distilled Text-to-Image Diffusion Models","authors":["Gaurav Patel","Jun Fang","Greg Ver Steeg","Qiang Qiu","Sravan Sripada"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Enabling Preference-driven Unlearning in Few-step Distilled Text-to-Image Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10859","noteUrl":"/review?paper=review-neurips-2026-154856","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154858","title":"What MLLMs Learn about When they Learn about Multimodal Reasoning","authors":["Jiwan Chung","Neel Joshi","Pratyusha Sharma","Youngjae Yu","Vibhav Vineet"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"What MLLMs Learn about When they Learn about Multimodal Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.01719","noteUrl":"/review?paper=review-neurips-2026-154858","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154859","title":"Watermarking Game-Playing Agents in Perfect-Information Extensive-Form Games","authors":["Juho Kim","Fei Fang","Tuomas Sandholm"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Watermarking Game-Playing Agents in Perfect-Information Extensive-Form Games","plainExplanation":"If one action’s utility margin exceeds the largest color-induced change, the entire action law remains unchanged.","primaryUrl":"https://arxiv.org/html/2605.14283","noteUrl":"/review?paper=review-neurips-2026-154859","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154860","title":"LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing","authors":["Yongcong Wang","Chengchao Shen","Guangwei Gao","Wei Wang","Pengwen Dai","Dianjie Lu","Guijuan Zhang","Zhuoran Zheng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11508","noteUrl":"/review?paper=review-neurips-2026-154860","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154863","title":"Hierarchical Variational Policies for Reward-Guided Diffusion","authors":["Kushagra Pandey","Farrin Marouf Sofian","Jan Niklas Groeneveld","Felix Draxler","Stephan Mandt"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Hierarchical Variational Policies for Reward-Guided Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21661","noteUrl":"/review?paper=review-neurips-2026-154863","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154865","title":"Compact SO(3) Equivariant Atomistic Foundation Models via Structural Pruning","authors":["Chen Wang","Siyu Hu","Guangming Tan","Weile Jia"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Unitary cancellation requires the transposed gradient transformation","plainExplanation":"The invariance proof contracts two representation matrices in the order \\(DD\\) and invokes orthogonality to replace that product by the identity. Orthogonality gives \\(D^\\top D=I\\), not \\(D^2=I\\).","primaryUrl":"https://arxiv.org/html/2605.08885","noteUrl":"/review?paper=review-neurips-2026-154865","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154868","title":"Beyond Oversquashing: Understanding Signal Propagation in GNNs Via Observables","authors":["Eden Nagar","Ya-Wei Eileen Lin","Ron Levie"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The modulation-induced momentum has the opposite sign","plainExplanation":"The continuum modulation calculation reverses the stated momentum convention. The broad spectral-GNN location-invariance claim also fails for an allowed two-node adjacency filter that swaps unequal amplitudes.","primaryUrl":"https://arxiv.org/html/2605.13383","noteUrl":"/review?paper=review-neurips-2026-154868","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154870","title":"Near-Optimal Last-Iterate Convergence for Zero-Sum Games with Bandit Feedback and Opponent Actions","authors":["Soumita Hait","Ping Li","Haipeng Luo","Mengxiao Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Entrywise concentration needs both sampling probabilities","plainExplanation":"The proof radius for a fixed game entry \\(A_{ij}\\) divides by a row sampling probability but omits the opponent-column probability. Its sample count is binomial with success probability \\(x_i y_j\\), not \\(x_i\\).","primaryUrl":"https://arxiv.org/html/2605.09363","noteUrl":"/review?paper=review-neurips-2026-154870","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154872","title":"Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation","authors":["Ziheng Li","Zexu Sun","Jinman Zhao","Erxue Min","Yongcheng Zeng","Hui Wu","Hengyi Cai","Shuaiqiang Wang","Xu Chen","Zhihong Deng","Dawei Yin"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The quadratic descent envelope is not an exact global bound","plainExplanation":"The Fisher derivation supports a local reward-variance envelope, not the claimed exact finite-step bound or universal actual-efficiency optimum at 50% accuracy. PSD alone also does not ensure an invertible Fisher or a unique parameter update.","primaryUrl":"https://arxiv.org/html/2509.06923v1","noteUrl":"/review?paper=review-neurips-2026-154872","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154873","title":"Diffusion Language Models Can Approximate Optimal Infilling Lengths Implicitly","authors":["Hengchang Liu","Zhao Yang","Bing Su"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Diffusion Language Models Can Approximate Optimal Infilling Lengths Implicitly","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.00476v1","noteUrl":"/review?paper=review-neurips-2026-154873","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154874","title":"Imperfect World Models are Exploitable","authors":["Logan M Bhamidipaty","Esmeralda S Whitammer","David Abel","Mykel J Kochenderfer","Subramanian Ramamoorthy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Imperfect World Models are Exploitable","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15960","noteUrl":"/review?paper=review-neurips-2026-154874","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154877","title":"SimWorld Studio: Automatic Environment Generation with Evolving Coding Agent for Embodied Agent Learning","authors":["Haoqiang Kang","Xiaokang Ye","Yuhan Liu","Siddhant Hitesh Mantri","Lingjun Mao","James Fleming","Drishti Regmi","Lianhui Qin"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"SimWorld Studio: Automatic Environment Generation with Evolving Coding Agent for Embodied Agent Learning","plainExplanation":"No issue was found in the checked claims. A finite-budget obstruction refutes a stronger guarantee than geometric reachability. The source does not identify these two guarantees.","primaryUrl":"https://arxiv.org/html/2605.09423","noteUrl":"/review?paper=review-neurips-2026-154877","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154878","title":"On the Meta-Design of Allocation Problems","authors":["Unai Fischer Abaigar","Emily Aiken","Christoph Kern","Juan C Perdomo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"On the Meta-Design of Allocation Problems","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.08786","noteUrl":"/review?paper=review-neurips-2026-154878","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154880","title":"Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation","authors":["Haozhan Tang","Xiuqi Zhu","Xinyin Zhang","Boxun Li","Virginia Smith","Kevin Kuo"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07111","noteUrl":"/review?paper=review-neurips-2026-154880","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154882","title":"PATCH: Learnable Tile-Level Hybrid Sparsity for LLMs","authors":["Mohammad Mozaffari","Younes Hourri","Maryam Mehri Dehnavi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"PATCH: Learnable Tile-Level Hybrid Sparsity for LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2509.23410","noteUrl":"/review?paper=review-neurips-2026-154882","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154884","title":"GLOVE: Global Verifier for LLM Memory-Environment Realignment","authors":["XINGKUN YIN","Hongyang Du"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"GLOVE: Global Verifier for LLM Memory-Environment Realignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.19249","noteUrl":"/review?paper=review-neurips-2026-154884","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154885","title":"A Theory of Time-Sensitive Language Generation: Sparse Hallucination Beats Mode Collapse","authors":["Atul Ganju","Travis McVoy","Shaddin Dughmi","Shang-Hua Teng"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"A Theory of Time-Sensitive Language Generation: Sparse Hallucination Beats Mode Collapse","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11302","noteUrl":"/review?paper=review-neurips-2026-154885","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154886","title":"Mapping Uncharted Symmetries: Machine Discovery in Combinatorics","authors":["Eugenio Cainelli","Lorenzo Luccioli","Alessandro Iraci","Michele D&amp;#x27;Adderio","Giovanni Paolini"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Mapping Uncharted Symmetries: Machine Discovery in Combinatorics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19063","noteUrl":"/review?paper=review-neurips-2026-154886","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154887","title":"Deriving Hyperparameter Scaling Laws via Modern Optimization Theory","authors":["Egor Shulgin","Jörg Franke","Dimitri von Rütte","Tianyue Zhang","Niccolò Ajroldi","Korbinian Pöppel","Bernhard Schölkopf","Aaron Klein","Peter Richtarik","Antonio Orvieto"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Deriving Hyperparameter Scaling Laws via Modern Optimization Theory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.15958","noteUrl":"/review?paper=review-neurips-2026-154887","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154888","title":"Towards Differentially Private Reinforcement Learning with General Function Approximation","authors":["Yi He","Xingyu Zhou"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Towards Differentially Private Reinforcement Learning with General Function Approximation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07049","noteUrl":"/review?paper=review-neurips-2026-154888","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154889","title":"Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs","authors":["Seung-Hyun Lee","Dongyoon Han","Sangdoo Yun"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.18639","noteUrl":"/review?paper=review-neurips-2026-154889","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154890","title":"Contribution-Aware Structured Sparsity for Model Merging","authors":["Yan Li","GUIPING CAO","Meng XU","Tao Jiang","Yaguang Song","Ming Tao","Yaowei Wang","Dongmei Jiang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Contribution-Aware Structured Sparsity for Model Merging","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34184","noteUrl":"/review?paper=review-neurips-2026-154890","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154891","title":"WebNavigator: Global Web Navigation via Interaction Graph Retrieval","authors":["Xuanwang Zhang","Yuteng Han","Jinnan Qi","Xinyu Liu","Mulong Xie","Zhen Wu","Xinyu Dai"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"WebNavigator: Global Web Navigation via Interaction Graph Retrieval","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.20366","noteUrl":"/review?paper=review-neurips-2026-154891","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154892","title":"Self-Supervised On-Policy Reinforcement Learning via Contrastive Proximal Policy Optimisation","authors":["Asim Osman","Sasha Abramowitz","Mark Bergh","Ulrich Armel Mbou Sob","Ruan John de Kock","Omayma Mahjoub","Oussama Hidaoui","Noah De Nicola","Arnol M Fokam","Felix Chalumeau","Daniel Rajaonarivonivelomanantsoa","Siddarth Singh","Refiloe Shabe","Juan Formanek","Simon Du Toit","Arnu Pretorius"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A contrastive logit is not an identified occupancy value","plainExplanation":"InfoNCE identifies relative goal logits within each action row, not occupancy values calibrated across actions. An admissible shared-distance construction uses that row freedom to reverse the occupancy-improving policy direction.","primaryUrl":"https://arxiv.org/html/2605.13554","noteUrl":"/review?paper=review-neurips-2026-154892","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154893","title":"Holistic Scaling Laws for Optimal Mixture-of-Experts Architecture Optimization","authors":["Weilin Wan","Jingtao Han","Debing Zhang","Weizhong Zhang","Cheng Jin"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The architectural constraint counts depth twice in the active-parameter term","plainExplanation":"Equation (3) defines \\(N_a\\) as the total active parameter count across layers, then uses \\(6N_aL\\) in the FLOPs constraint. The main constraint multiplies an already all-layer active-parameter count by depth again, contradicting the appendix’s FLOPs identity and the subsequent reduction.","primaryUrl":"https://arxiv.org/html/2603.21862","noteUrl":"/review?paper=review-neurips-2026-154893","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154895","title":"Arbitrarily Conditioned Hierarchical Flows for Spatiotemporal Events","authors":["Keyan Chen","Qiwei Yuan","Zhitong Xu","Bin Shen","Shandian Zhe"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Arbitrarily Conditioned Hierarchical Flows for Spatiotemporal Events","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.01226","noteUrl":"/review?paper=review-neurips-2026-154895","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154896","title":"Learning What Evaluators Value: A Reliable Approach to Modeling Evaluator Preferences","authors":["Madeline Kitch","Nihar Shah"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A nondecreasing link cannot invert weak output order","plainExplanation":"The weak-link ordering lemma incorrectly inverts non-strict inequalities through a link with flat regions. The discrete-to-continuous transfer also turns a lower cell-mass condition into an impossible upper density bound.","primaryUrl":"https://arxiv.org/html/2605.16615","noteUrl":"/review?paper=review-neurips-2026-154896","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154897","title":"fmxcoders: Factorized Masked Crosscoders for Cross-Layer Feature Discovery","authors":["Andreas D Demou","Panagiotis Koromilas","James Oldfield","Yannis Panagakis","Mihalis Nicolaou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Raw Wasserstein margins change under an equivalent latent rescaling","plainExplanation":"The crosscoder comparison interprets larger raw Wasserstein separation between activation distributions as a robustness improvement across methods without fixing latent scale. ReLU and global BatchTopK are positively homogeneous.","primaryUrl":"https://arxiv.org/html/2605.09438","noteUrl":"/review?paper=review-neurips-2026-154897","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154900","title":"The Geometry of Linear Program Compression: An Exact Characterization and Learning Algorithm","authors":["Yuhan Ye","Omar Bennouna"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Geometry of Linear Program Compression: An Exact Characterization and Learning Algorithm","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25635","noteUrl":"/review?paper=review-neurips-2026-154900","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154901","title":"MemReward: Graph-Based Experience Memory for LLM Reward Prediction with Limited Labels","authors":["Tianyang Luo","Tao Feng","Zhigang Hua","Yan Xie","Shuang Yang","Ge Liu","Jiaxuan You"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"MemReward: Graph-Based Experience Memory for LLM Reward Prediction with Limited Labels","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.19310","noteUrl":"/review?paper=review-neurips-2026-154901","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154904","title":"Learning From Failures: Efficient Reinforcement Learning Control with Episodic Memory","authors":["Chenyang Miao","Bingchen Zhou","Yu Jianliang","Zeyang Liu","Xingyu Chen","Qiongjie Cui","Xuguang Lan"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Learning From Failures: Efficient Reinforcement Learning Control with Episodic Memory","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.07110","noteUrl":"/review?paper=review-neurips-2026-154904","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154906","title":"Robust PAC Learning of Concurrent Stochastic Games","authors":["Angel Y He","David Parker"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Eventual reachability does not ensure coverage within the sampling horizon","plainExplanation":"Eventual reachability does not ensure that reset trajectories of length \\(H\\) sample every row used by the stopping rule. A within-exploration-horizon coverage condition repairs that gap. Separately, a finite-horizon game can have no stationary \\(\\varepsilon\\)-equilibrium for small \\(\\varepsilon\\), even though time-dependent equilibria exist. The paper acknowledges the latter distinction, but its universal stationary-existence and global-margin statements do not carry the qualification through. Existence-conditional value transfer and non-existence certification are not refuted by that example.","primaryUrl":"https://arxiv.org/html/2609.04189","noteUrl":"/review?paper=review-neurips-2026-154906","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement","incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154911","title":"Beyond Ground Truth: Evaluating Non-Verifiable Reasoning in LLMs through Moral Robustness","authors":["Elizaveta Tennant","Benjamin Henke","Anita Keshmirian","Murray Shanahan","Verena Rieser","Kristian Lum","Sydney Levine","Julia Haas"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Beyond Ground Truth: Evaluating Non-Verifiable Reasoning in LLMs through Moral Robustness","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.12731v1","noteUrl":"/review?paper=review-neurips-2026-154911","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154912","title":"A Single-Sample Polylogarithmic Regret Bound for Nonstationary Online Linear Programming","authors":["Haoran Xu","Owen Shen","Peter W Glynn","Yinyu Ye","Patrick Jaillet"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"A Single-Sample Polylogarithmic Regret Bound for Nonstationary Online Linear Programming","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.14673","noteUrl":"/review?paper=review-neurips-2026-154912","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154913","title":"Reliable Chain-of-Thought via Prefix Consistency","authors":["Naoto Iwase","Yuki Ichihara","Mohammad Atif Quamar","Junpei Komiyama"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Reliable Chain-of-Thought via Prefix Consistency","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07654","noteUrl":"/review?paper=review-neurips-2026-154913","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154916","title":"Distributionally Robust Token Optimization in RLHF","authors":["Yeping Jin","Jiaming Hu","Yannis Paschalidis"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The KL dual optimum can occur only at infinite temperature parameter","plainExplanation":"Algorithm 2 requests an attained minimizer over finite positive \\(\\eta\\) for every finite span-loss minibatch. The algorithm requests a finite attained dual minimizer even when the robust optimum saturates at the maximum-loss support and the dual infimum occurs only at infinite parameter.","primaryUrl":"https://arxiv.org/html/2604.08577","noteUrl":"/review?paper=review-neurips-2026-154916","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154919","title":"When Attribution Patching Lies: Diagnosis and a Second-Order Correction","authors":["Luyang Zhang","Jialu Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Sampled Hessian secants are lower estimates, not remainder certificates","plainExplanation":"The empirical remainder quantity substitutes the maximum of finitely sampled Hessian secants for an upper Lipschitz-Hessian constant. A sampled maximum is bounded above by the true path constant, so substitution can shrink an upper certificate.","primaryUrl":"https://arxiv.org/html/2606.09899","noteUrl":"/review?paper=review-neurips-2026-154919","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154920","title":"Path-independent Flow Matching for Multi-parameter Generative Dynamics","authors":["Francisco Téllez","Amirhossein Zamani","Philippe Martin","Shuang Ni","Guy Wolf","Eugene Belilovsky","Sina Sanjari","Yanlei Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The printed Lie-bracket sign reverses flow compatibility","plainExplanation":"The pointwise compatibility condition uses the wrong bracket sign. The general-parameter formula independently uses own-coordinate derivatives where cross derivatives are required; either defect can produce path-dependent endpoints.","primaryUrl":"https://arxiv.org/html/2605.13487","noteUrl":"/review?paper=review-neurips-2026-154920","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154921","title":"Improved techniques for fine-tuning flow models via adjoint matching: a deterministic control pipeline","authors":["Zhengyi Guo","Jiayuan Sheng","David Yao","Wenpin Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strict convexity alone does not ensure a finite radial-control optimum","plainExplanation":"The general radial-control theorem needs coercivity and boundary cases beyond strict convexity. The symmetric-mixture example is not generated or normalized by its printed equations, and the VE toy coefficient has an extra time factor that changes its peak.","primaryUrl":"https://arxiv.org/html/2605.06583","noteUrl":"/review?paper=review-neurips-2026-154921","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154924","title":"DROGO: Default Representation Objective via Graph Optimization in Reinforcement Learning","authors":["Hon Tik (Rick) Tse","Marlos C. Machado"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Symmetrization does not preserve unit graph degree","plainExplanation":"The objective replaces the degree matrix of \\(\\widetilde P=(P+P^\\top)/2\\) by the identity, assuming only row-stochastic \\(P\\). Expanding the symmetric edge energy gives \\(\\sum_i d_i u_i^2-u^\\top\\widetilde P u\\), with \\(d=\\widetilde P\\mathbf1\\).","primaryUrl":"https://arxiv.org/html/2602.00403","noteUrl":"/review?paper=review-neurips-2026-154924","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154925","title":"Synthetic Pre-Pre-Training Improves Language Model Robustness to Noisy Pre-Training Data","authors":["Xu Guo","Runyu Peng","Jian Tong","Yunhua Zhou","haijun Lv","Zhihui Lu","Qipeng Guo"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Synthetic Pre-Pre-Training Improves Language Model Robustness to Noisy Pre-Training Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10129","noteUrl":"/review?paper=review-neurips-2026-154925","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154926","title":"Generative Cross-Entropy: A Strictly Proper Loss for Data-Efficient Classification","authors":["Qipeng Zhan","Zhuoping Zhou","Li Shen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Generative Cross-Entropy: A Strictly Proper Loss for Data-Efficient Classification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.06689","noteUrl":"/review?paper=review-neurips-2026-154926","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154930","title":"In-Context Black-Box Optimization with Unreliable Feedback","authors":["Nicolas Samuel Blumer","Julien Martinelli","Samuel Kaski"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The reward must compare against the previous incumbent","plainExplanation":"The printed reward compares a candidate to an incumbent that already includes it. Its clipped improvement is therefore identically zero, unlike the previous-incumbent reward defined elsewhere. This is a pseudocode subscript typo. Other passages provide the corrected reward; the paper does not establish that training used an identically zero signal.","primaryUrl":"https://arxiv.org/html/2605.06187","noteUrl":"/review?paper=review-neurips-2026-154930","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-154931","title":"KVBuffer: IO-aware Serving for Linear Attention","authors":["Longwei Zou","Lin Zhong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The checkpoint transformed key has noncommuting factors in the wrong order","plainExplanation":"The chunkwise identity uses cumulative decay matrix \\(G\\), step matrix \\(B=\\operatorname{Diag}(\\beta)\\) and triangular inverse \\(A\\), placing the checkpoint transformed key at \\(GABK\\). After correcting the token Gram, recurrence expansion gives \\(U=ABV-ABGK S_*\\), where \\(A=[I+\\operatorname{strictLower}(B(\\Gamma\\odot KK^\\top))]^{-1}\\).","primaryUrl":"https://arxiv.org/html/2605.19049","noteUrl":"/review?paper=review-neurips-2026-154931","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154932","title":"ECHO: Terminal Agents Learn World Models for Free","authors":["Vaishnavi Shrivastava","Ahmed Awadallah","Dimitris Papailiopoulos"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ECHO: Terminal Agents Learn World Models for Free","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.24517","noteUrl":"/review?paper=review-neurips-2026-154932","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154934","title":"TACO: Towards Task-Consistent Open-Vocabulary Adaptation in Video Recognition","authors":["Minghao Zhu","Xiao Lin","Mengxian Hu","Xun Zhou","Liuyi Wang","Xiaoyan Qi","Chengju Liu","Qijun Chen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TACO: Towards Task-Consistent Open-Vocabulary Adaptation in Video Recognition","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.25478","noteUrl":"/review?paper=review-neurips-2026-154934","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154939","title":"On Time, Within Budget: Constraint-Driven Online Resource Allocation for Agentic Workflows","authors":["Xinglin Wang","Zishen Liu","Shaoxiong Feng","Peiwen Yuan","Yiwei Li","Jiayi Shi","Yueqi Zhang","Chuyi Tan","Ji Zhang","Boyuan Pan","Yao Hu","Prof. Kan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"On Time, Within Budget: Constraint-Driven Online Resource Allocation for Agentic Workflows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06110","noteUrl":"/review?paper=review-neurips-2026-154939","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154940","title":"MoME: Mixture-of-Memory Embeddings for Context-Aware Sparse Lookup","authors":["Muchen Li","Leonid Sigal","Renjie Liao"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"MoME: Mixture-of-Memory Embeddings for Context-Aware Sparse Lookup","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.15126","noteUrl":"/review?paper=review-neurips-2026-154940","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154942","title":"Adversary-Robust Learning from Fully Asynchronous Directional Derivative Estimates","authors":["Anik Kumar Paul","Nibedita Roy","Nagesh Talagani","Swetha Ganesh","Gugan Chandrashekhar Mallika Thoppe","Alexandre Reiffers-Masson"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The sampled second moment needs the sampling factor and finite-difference bias","plainExplanation":"Sampling one of \\(mN\\) coordinates and multiplying its residual by \\(mN\\) gives \\(mN\\sum_j\\mathbb E[(Y_j-y_j)^2]\\), not an unamplified noise sum. With one direction, four workers, independent unit Rademacher evaluation noise and a quadratic objective at zero, the actual noise contribution is \\(2\\), while the printed bound gives \\(1\\).\n\nA noiseless finite difference can also have nonzero bias at a zero true gradient: \\(f(x)=x^2/2+\\sin x-x\\) supplies such a case. The bias cannot be absorbed into a term that is declared to vanish with observation noise. The stated moment bound fails. Fixed-dimension convergence may survive corrected constants and a renewed proof.","primaryUrl":"https://arxiv.org/html/2605.09337","noteUrl":"/review?paper=review-neurips-2026-154942","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154944","title":"Differentiable Belief-based Opponent Shaping","authors":["Aarav G Sane","Karthik Sivachandran","Rohan Paleja"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Differentiable Belief-based Opponent Shaping","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29042","noteUrl":"/review?paper=review-neurips-2026-154944","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154945","title":"FLUX: Geometry-Aware Longitudinal Flow Matching with Mixture of Experts","authors":["Josue Ortega Caro","Yongxu Zhang","Hannah M Batchelor","Sizhuang He","Jessica Cardin","Shreya Saxena"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The flow target omits the conversion from local to global time","plainExplanation":"The chain rule gives \\(dx/dt=(T-1)dx/d\\alpha\\). For consecutive Gaussian marginals with means \\(k\\) and common variance \\(0.01\\), a shared-offset coupling gives local velocity \\(1\\). That velocity has zero printed flow-matching loss, but integration over global time \\([0,1]\\) translates the mean by \\(1\\), rather than \\(T-1\\). This is a time-parameterization inconsistency; it does not establish a failure of the actual implementation.","primaryUrl":"https://arxiv.org/html/2605.08648","noteUrl":"/review?paper=review-neurips-2026-154945","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154948","title":"LESSViT: Robust Hyperspectral Representation Learning under Spectral Configuration Shift","authors":["Haozhe Si","Yuxuan Wan","Yuqing Wang","Minh Do","Han Zhao"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"LESSViT: Robust Hyperspectral Representation Learning under Spectral Configuration Shift","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.18541","noteUrl":"/review?paper=review-neurips-2026-154948","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154949","title":"Agentic Neural Architecture Search","authors":["Seokhoon Jeong","Mijung Kim","Taehwan Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Agentic Neural Architecture Search","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.07984","noteUrl":"/review?paper=review-neurips-2026-154949","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154951","title":"Hide to Guide: Learning via Semantic Masking","authors":["Ruitao Liu","Qinghao Hu","Alex Hu","Yecheng Wu","Shang Yang","Luke J Huang","Zhuoyang Zhang","Han Cai","Song Han"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Hide to Guide: Learning via Semantic Masking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25198","noteUrl":"/review?paper=review-neurips-2026-154951","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154952","title":"Memory Determines Learning Direction: A Theory of Gradient-Based Optimization in State Space Models","authors":["JingChuan Guan","Tomoyuki Kubota","Yasuo Kuniyoshi","Kohei Nakajima"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Teacher coefficients are not normalized memory functions","plainExplanation":"For independent equal-variance inputs and \\(y=\\sum_k\\alpha_kU_k\\), the projection-based memory function is \\(\\operatorname{MF}_y(\\tau)=\\alpha_\\tau^2/\\sum_k\\alpha_k^2\\). It is nonnegative and unchanged by rescaling all coefficients. A raw coefficient \\(\\alpha_\\tau\\) has neither property. The single unit impulse is only a special case where the two agree. The issue is the memory interpretation in Appendix D.3, not the entire learning-direction analysis.","primaryUrl":"https://arxiv.org/html/2510.00563","noteUrl":"/review?paper=review-neurips-2026-154952","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154955","title":"Interpreting Latent Protein Language Model Features with Geometric Annotations","authors":["Siddharth Setlur","Djordje Mihajlovic","Darrick Lee"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Interpreting Latent Protein Language Model Features with Geometric Annotations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.26419","noteUrl":"/review?paper=review-neurips-2026-154955","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154956","title":"On the Limits of Latent Reuse in Diffusion Models","authors":["Yifeng Yu","Lu Yu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"On the Limits of Latent Reuse in Diffusion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13448","noteUrl":"/review?paper=review-neurips-2026-154956","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154960","title":"Manifold Drift in Flow Preference Optimization: A Root Cause of Reward Hacking","authors":["Yansen Han","Shengyi Liao","Yuanxing Zhang","Pengfei Wan","Tao Lin"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Manifold Drift in Flow Preference Optimization: A Root Cause of Reward Hacking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.20011","noteUrl":"/review?paper=review-neurips-2026-154960","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154961","title":"A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning","authors":["Hyosoon Jang","Hyunjin Seo","Honghui Kim","Taewon Kim","Yunhui Jang","Seonghyun Park","Sungsoo Ahn"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.13249","noteUrl":"/review?paper=review-neurips-2026-154961","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154962","title":"IRPO: Boosting Image Restoration via Post-training GRPO","authors":["Haoxuan Xu","Yi Liu","Tianfu Li","Ruolin Shen","Boyuan Jiang","Jinlong Peng","Donghao Luo","Xiaobin Hu","Shuicheng Yan","Haoang Li"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"IRPO: Boosting Image Restoration via Post-training GRPO","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.00814","noteUrl":"/review?paper=review-neurips-2026-154962","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154967","title":"TimeES: Probabilistic and Deterministic Time Series Forecasting via Evolutionary Spectra","authors":["Weiwei Ye","Renhe Jiang","Hangchen Liu","Dongyuan Li","Yoshihide Sekimoto"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Time-varying Fourier coefficients and STFT normalization need narrower statements","plainExplanation":"Real synthesis does not force rowwise Hermitian symmetry when coefficients vary with time, though imposing that symmetry remains sufficient. The appendix’s STFT approximation also drops the frequency integral’s window mass, preventing the stated identification with local amplitude.","primaryUrl":"https://arxiv.org/html/2609.32384","noteUrl":"/review?paper=review-neurips-2026-154967","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference"],"classifications":["major"]},{"slug":"review-neurips-2026-154970","title":"Efficient Transferable Optimal Transport via Min-Sliced Transport Plans","authors":["Xinran Liu","Elaheh Akbari","Rocio Diaz Martin","Navid NaderiAlizadeh","Soheil Kolouri"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The smoothed final quantile is infinite","plainExplanation":"Injectivity does not guarantee uniqueness of all one-dimensional optimal plans at $p=1$. The smoothed sorting formula also evaluates a full-support inverse CDF at probability one, producing infinity instead of a finite sorted endpoint.","primaryUrl":"https://arxiv.org/html/2511.19741","noteUrl":"/review?paper=review-neurips-2026-154970","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154972","title":"Information-Directed Offline-to-Online Reinforcement Learning","authors":["Keru Chen"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Information-Directed Offline-to-Online Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29405","noteUrl":"/review?paper=review-neurips-2026-154972","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154973","title":"EditBridge: Towards Faithful and Efficient Ultra-High-Resolution Image Editing","authors":["Jiayi Song","Shijie Huang","Fangtai Wu","Yubo Huang","Zhenxiong Tan","Songhua Liu","Jiaming Liu","Ruihua Huang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"EditBridge: Towards Faithful and Efficient Ultra-High-Resolution Image Editing","plainExplanation":"No issue was found in the checked claims. Treating the inherited low-resolution block label as an unscaled high-resolution coordinate omits the conversion stated in the same paragraph. It does not establish that the specified attention window is miscentered.","primaryUrl":"https://arxiv.org/html/2608.18063","noteUrl":"/review?paper=review-neurips-2026-154973","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154974","title":"BayesRAG: Probabilistic Mutual Evidence Corroboration for Multimodal Retrieval-Augmented Generation","authors":["Xuan Li","Yining Wang","Haocai Luo","Shengping Liu","Jiaen Liang","Ying Fu","Wei Huang","Jun Yu","Junnan Zhu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The printed layout threshold accepts every same-page pair","plainExplanation":"Any two points on one page are at distance at most its diagonal. Thus \\(d\\leq 2\\,\\mathrm{diag}\\) accepts even opposite corners once the same-source and same-page tests pass. The distance test supplies no additional locality restriction. The finding concerns the printed threshold \\(\\tau=2\\). It does not determine which threshold the experiments actually used.","primaryUrl":"https://arxiv.org/html/2601.07329","noteUrl":"/review?paper=review-neurips-2026-154974","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154976","title":"Generalization at the Edge of Stability","authors":["Mario Tuci","Caner Korkmaz","Umut Simsekli","Tolga Birdal"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Zero box dimension does not imply zero covering entropy","plainExplanation":"The covering corollary fails on a finite set of box dimension zero. The dynamical proof separately assumes powers of an ergodic base are ergodic, which need not hold and changes the relevant Birkhoff limit.","primaryUrl":"https://arxiv.org/html/2604.19740","noteUrl":"/review?paper=review-neurips-2026-154976","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154977","title":"Efficient and Simple Data Mixing All The Time","authors":["Michael Hu","Apurva Gandhi","Kyunghyun Cho","Tal Linzen","Pratyusha Sharma"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Efficient and Simple Data Mixing All The Time","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15220v1","noteUrl":"/review?paper=review-neurips-2026-154977","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154979","title":"FTerViT: Fully Ternary Vision Transformer","authors":["Szymon Ruciński","Pietro Bonazzi","Engin Turetken","Simon Narduzzi","Michele Magno","Nadim Maamari"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"FTerViT: Fully Ternary Vision Transformer","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.21171","noteUrl":"/review?paper=review-neurips-2026-154979","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154981","title":"Constrained Goal-directed Planar Graph Generation with Grammar-based Reinforcement Learning","authors":["Nicolas Hochuli","Lorenzo Miele","Kristina Shea","Tino Stankovic"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Constrained Goal-directed Planar Graph Generation with Grammar-based Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.06244","noteUrl":"/review?paper=review-neurips-2026-154981","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154983","title":"LOFT: Low-Rank Orthogonal Fine-Tuning via Task-Aware Support Selection","authors":["Lanxin Zhao","Bamdev Mishra","Pratik Kumar Jawanpuria","Lequan Lin","Dai Shi","Junbin Gao","Andi Han"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LOFT: Low-Rank Orthogonal Fine-Tuning via Task-Aware Support Selection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11872","noteUrl":"/review?paper=review-neurips-2026-154983","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154984","title":"Learning to Sample From Diffusion Models via Inverse Reinforcement Learning","authors":["Constant Bourdrez","Alexandre Verine","Olivier Cappé"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Learning to Sample From Diffusion Models via Inverse Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.08689","noteUrl":"/review?paper=review-neurips-2026-154984","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154986","title":"Curvature-Dependent Lower Bounds for Frank-Wolfe","authors":["Jannis Halbey","Christophe Roux","Sebastian Pokutta"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Curvature-Dependent Lower Bounds for Frank-Wolfe","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10595","noteUrl":"/review?paper=review-neurips-2026-154986","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154987","title":"Pooling Versus Ensembling for Ridge Regression Under Covariate Shift","authors":["Maya Ramchandran","Rajarshi Mukherjee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The resolvent equivalent is missing a factor of inverse lambda","plainExplanation":"The fixed-effect formula double counts signal scale, a full-risk ratio omits test noise, and an appendix resolvent equivalent loses an inverse-regularization factor. The main resolvent normalization is already correct.","primaryUrl":"https://arxiv.org/html/2608.01641","noteUrl":"/review?paper=review-neurips-2026-154987","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-154990","title":"Representation Fréchet Loss for Visual Generation","authors":["Jiawei Yang","Zhengyang Geng","Xuan Ju","Yonglong Tian","Yue Wang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Representation Fréchet Loss for Visual Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.28190","noteUrl":"/review?paper=review-neurips-2026-154990","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-154992","title":"Gradient Regularized Newton Boosting Trees with Global Convergence","authors":["Nikita Zozoulenko","Daniel Falkowski","Thomas Cass","Lukas Gonon"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Regularized cross-entropy does not satisfy the unshifted loss domination","plainExplanation":"The convergence application compares the decrement to an unshifted regularized cross-entropy loss with positive minimum. It needs loss excess rather than loss itself to apply the stated domination condition.","primaryUrl":"https://arxiv.org/html/2605.00581","noteUrl":"/review?paper=review-neurips-2026-154992","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-154994","title":"Seeing Together: Multi-Robot Cooperative Egocentric Spatial Reasoning with Multimodal Large Language Models","authors":["Kunyu Peng","Zhikun Zhou","Kailun Yang","Di Wen","Ruiping Liu","Yufan Chen","Junwei Zheng","Hao Shi","Yi Zhou","M. Saquib Sarfraz","Danda Pani Paudel","Luc V Gool"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Seeing Together: Multi-Robot Cooperative Egocentric Spatial Reasoning with Multimodal Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.18431","noteUrl":"/review?paper=review-neurips-2026-154994","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-154995","title":"Motif-Mamba: network motif improved mamba for long-range sequence modeling","authors":["Hao Chonghe","Yue Sun","Jian Zhang","Yansong Wang","Wangzi Yao","Yao Yunjie","Tielin Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The projected flow and the relaxed motif formula need corrections","plainExplanation":"Differentiating the defined coordinate gives a negative decay term and a learned coupling term. The printed field reverses the former sign and substitutes $-I$ for the latter without an assumption that permits either substitution.","primaryUrl":"https://arxiv.org/html/2608.00027","noteUrl":"/review?paper=review-neurips-2026-154995","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-154998","title":"Reflective Prompted Policy Optimization: Trajectory-Grounded Revision and Salience Bias","authors":["Rahaf Abu Hara","Vaibbhav Murarri","Claudio Zito"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The abstention predicate reverses the stated threshold behavior","plainExplanation":"For finite reward, \\(\\tau=-\\infty\\) makes the reward predicate always true; \\(\\tau=+\\infty\\) makes it false. Likewise, lowering a positive threshold from \\(480\\) to \\(240\\) expands the set of rewards that cause abstention. It cannot weaken abstention under the printed predicate. This is a specification inconsistency. An unreported sentinel branch could change execution, so actual run behavior is not inferred.","primaryUrl":"https://arxiv.org/html/2605.08315","noteUrl":"/review?paper=review-neurips-2026-154998","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155000","title":"VI-CuRL: Stabilizing Verifier-Independent RL Reasoning via Confidence-Guided Variance Reduction","authors":["Xin-Qiang Cai","Masashi Sugiyama"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A decreasing envelope reverses the threshold comparison","plainExplanation":"The threshold rule does not enforce exact retention, and the decreasing entropy-envelope comparison is reversed. The variance argument also drops a selected-mean term that can diverge as retention tends to zero.","primaryUrl":"https://arxiv.org/html/2602.12579","noteUrl":"/review?paper=review-neurips-2026-155000","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155001","title":"Theory on Attention Dynamics for Out-of-Distribution In-Context Learning","authors":["Junze Deng","Daouda Sow","Sen Lin","Yingbin Liang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The two-feature loss inequality and its induction need correction","plainExplanation":"The inequality already has the wrong direction at \\(T_1\\). Separately, \\(\\eta_t\\le1/\\alpha_{\\max}\\) does not imply \\(1-\\eta_te^{\\alpha_{\\max}}\\ge0\\), so the displayed induction does not supply the intended all-time ordering.","primaryUrl":"https://arxiv.org/html/2609.36448","noteUrl":"/review?paper=review-neurips-2026-155001","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"memory-retrieval-saliency-identity","title":"Memory Retrieval for Changing Preferences","authors":["Yuehan Qin","Li Li","Linxin Song","Jiate Li","Wei Yang","Yuqing Yang","Yue Zhao"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The saliency identity adds an unsupported query term","plainExplanation":"Under conditional sufficiency, answer-saliency differences equal marginal answer-evidence differences. The added query term is unsupported, and the approximation bound needs an answer–query alignment term. Positive joint distributions disprove the identities and can reverse the rankings.","primaryUrl":"https://arxiv.org/abs/2606.02976v1","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155004","title":"Reward Modeling for Multi-Agent Orchestration","authors":["King Yeung Tsang","Zihao Zhao","Vishal Venkataramani","Haizhou Shi","Zixuan Ke","Semih Yavuz","Shafiq Joty","Hao Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Reward Modeling for Multi-Agent Orchestration","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.13598","noteUrl":"/review?paper=review-neurips-2026-155004","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155006","title":"SocialDirector: Training-Free Social Interaction Control for Multi-Person Video Generation","authors":["LIANGYANG OUYANG","Ruicong Liu","Caixin Kang","Yifei Huang","Yoichi Sato"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The post-normalization bias still grows with head dimension","plainExplanation":"For equal baseline logits, the log odds between a biased and an unbiased key shift by \\(\\gamma\\sqrt d\\). With \\(\\gamma=0.5\\), dimensions \\(4,16,64,256\\) give shifts \\(1,2,4,8\\). Setting the value contribution to zero isolates this logit effect. The directional attention effect remains; the claimed dimension-independent scaling does not follow from the displayed placement.","primaryUrl":"https://arxiv.org/html/2605.10079","noteUrl":"/review?paper=review-neurips-2026-155006","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155008","title":"StructLens: A Structural Lens for Language Models via Maximum Spanning Trees","authors":["Haruki Sakajo","Frederikus Hudi","Yusuke Sakai","Hidetaka Kamigaito","Taro Watanabe"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"StructLens: A Structural Lens for Language Models via Maximum Spanning Trees","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.03328","noteUrl":"/review?paper=review-neurips-2026-155008","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155011","title":"Deep Ensembles for Epistemic Uncertainty: A Frequentist Perspective","authors":["Anchit Jain","Stephen Bates"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The BvM display needs estimator centering after repairing regularity assumptions","plainExplanation":"Appendix A.3.1 states conditional convergence of \\(\\sqrt n(\\theta-\\theta_0)\\) to a fixed zero-mean Gaussian. Its literal Assumption 6 is inconsistent: it requires both a test’s true-parameter rejection probability and its complementary acceptance probability to tend to zero. The following centering analysis concerns an explicitly repaired formulation using the usual coherent regular Bernstein–von Mises assumptions. The original testing condition is logically inconsistent. Replacing it by coherent standard BvM assumptions still does not justify the truth-centered conditional display, because the posterior’s random sampling shift remains. This is an original assumption/display defect and a centering obligation for an intended nonvacuous repair. The Bernoulli example is not a counterexample satisfying the literal inconsistent assumptions. Centered covariance and mutual-information expansions may remain valid.","primaryUrl":"https://arxiv.org/html/2510.22063","noteUrl":"/review?paper=review-neurips-2026-155011","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155013","title":"Joint Optimization of Tool Creation and Use for Large Language Model Agents","authors":["Zhi Rui Tam","Chieh-Yen Lin","Yun-Nung (Vivian) Chen","Shao-Hua Sun","Hung-yi Lee"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The floor activates at an earlier schedule threshold","plainExplanation":"The stated floor-activation threshold does not solve the displayed schedule equation. The floor starts near normalized turn \\(0.673\\), and at the fourth of five discrete turns, rather than the claimed later activation. This is an explanatory calculation for the defined schedule. It does not establish which schedule was implemented or its empirical effect.","primaryUrl":"https://arxiv.org/html/2608.24571","noteUrl":"/review?paper=review-neurips-2026-155013","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-155016","title":"Lie Generator Networks for Nonlinear Partial Differential Equations","authors":["Shafayeth Jamil","Rehan Kapadia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Latent contraction does not imply monotone decoded enstrophy","plainExplanation":"Latent dissipativity does not force monotone physical enstrophy through an unconstrained decoder. Damping anisotropy also changes imaginary eigenvalue branches, and separate skew/damping profiles can change under admissible similarities even when the full generator family is equivalent.","primaryUrl":"https://arxiv.org/pdf/2603.29264v2","noteUrl":"/review?paper=review-neurips-2026-155016","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","interpretation_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155017","title":"Uniform Stability and Generalization Error of GD and SGD on Fixed-Point Parameters","authors":["Jonghyun Shin","Sejun Park"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Uniform Stability and Generalization Error of GD and SGD on Fixed-Point Parameters","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06934","noteUrl":"/review?paper=review-neurips-2026-155017","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155020","title":"Pose6DAug: Physically Plausible Multi-View Object Swapping for Robot Data Augmentation","authors":["Jonghoon Lee","Seong Hyeon Park","Minha Lee","Byungwoo Jeon","Jinwoo Shin"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"A source robot mask need not preserve visibility after object replacement","plainExplanation":"Section 3.3 restores the original robot/gripper mask above every rendered replacement and claims correct occlusion ordering. Restoring source robot pixels on top assumes their visibility remains unchanged under every replacement, which does not follow from compatible grasp contacts or collision freedom.","primaryUrl":"https://arxiv.org/html/2606.20118v2","noteUrl":"/review?paper=review-neurips-2026-155020","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155022","title":"Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models","authors":["Boxin Wang","Chankyu Lee","Nayeon Lee","Sheng-Chieh Lin","Wenliang Dai","Yang Chen","Yangyi Chen","Zhuolin Yang","Zihan Liu","Mohammad Shoeybi","Bryan Catanzaro","Wei Ping"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Setting an on-policy ratio to one before differentiation deletes its gradient","plainExplanation":"Section 4.1.2 replaces the likelihood ratio by one and presents an expectation of group-centered advantages as the REINFORCE objective. An on-policy likelihood ratio has value one but nonzero derivative; replacing it by a constant makes the equal-length group-centered surrogate identically zero. The printed surrogate loses the intended policy gradient. A single on-policy update is compatible with a correctly implemented GRPO estimator.","primaryUrl":"https://arxiv.org/html/2512.13607","noteUrl":"/review?paper=review-neurips-2026-155022","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155024","title":"ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both","authors":["Ziyu Guo","Yu Liu","Xinyan Chen","Pheng-Ann Heng"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The minimized loss rewards KL divergence with the printed sign","plainExplanation":"With zero advantages, the loss becomes \\(-\\beta D_{\\mathrm{KL}}\\). A policy farther from the reference has a smaller loss: a Bernoulli policy with parameter \\(0.99\\) is preferred to its reference with parameter \\(0.5\\). The other stated loss signs fix the minimization convention, so they do not cancel this reversal. This is a manuscript objective-sign error, not evidence that the training code used that sign.","primaryUrl":"https://arxiv.org/html/2605.15198","noteUrl":"/review?paper=review-neurips-2026-155024","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155027","title":"MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning","authors":["Ruilin Tong","Dong Gong"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Pointwise KL constraints need state-dependent dual multipliers","plainExplanation":"A separate constraint at each context has a multiplier \\(\\beta(x)\\). A scalar multiplier instead enforces an averaged KL constraint. With a uniform binary reference and reward gaps \\(0.1\\) and \\(1\\) at two contexts, one shared temperature gives different KL values and can violate one pointwise budget while meeting the average. Equal pointwise budgets require different temperatures, by a factor of ten in this example. The pointwise-to-averaged dual step is invalid. The fixed-penalty main method and BCE heuristic are not refuted.","primaryUrl":"https://arxiv.org/html/2607.06974","noteUrl":"/review?paper=review-neurips-2026-155027","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155028","title":"SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema","authors":["Shuang Wu","Jiude Wei","Cewu Lu","Jianhua Sun"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.18374","noteUrl":"/review?paper=review-neurips-2026-155028","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155029","title":"BrainWhisperer: Leveraging Whisper for Speech Decoding in Neuroprosthetics","authors":["Tommaso Boccato","Michał Olak","Moein Khajehnejad","Matteo Ferrante"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BrainWhisperer: Leveraging Whisper for Speech Decoding in Neuroprosthetics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2603.13321v1","noteUrl":"/review?paper=review-neurips-2026-155029","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155034","title":"Slowly Annealed Langevin Dynamics: Theory and Applications to Training-Free Guided Generation","authors":["Atsushi Nitanda","Dake Bu","Yueming LYU","Tanya Veeravalli"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Slowly Annealed Langevin Dynamics: Theory and Applications to Training-Free Guided Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07950","noteUrl":"/review?paper=review-neurips-2026-155034","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155037","title":"Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models","authors":["Zhiqi Zhang","Xinhao Zhong","Yi Sun","Shuoyang Sun","Jiahui Wang","Bin Chen","Shu-Tao Xia","Xuan Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Low-rank posterior-mean Jacobians need not give low-rank velocity Jacobians","plainExplanation":"The printed velocity target reverses its path derivative. The confinement proof also omits the moving-projector term, and its low-rank velocity motivation does not follow from the cited low-rank posterior-mean Jacobian result.","primaryUrl":"https://arxiv.org/html/2602.01089","noteUrl":"/review?paper=review-neurips-2026-155037","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","invalid_inference","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155039","title":"Efficient Pre-Training with Token Superposition","authors":["Bowen Peng","Théo Gigant","Jeffrey Quesnelle"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Repeated tokens invalidate the uniform-vocabulary entropy correction","plainExplanation":"Appendix C treats a bag of \\(s\\) token positions as a uniform vocabulary target with entropy \\(\\log s\\), while Listing 3 averages ordinary cross-entropy over positions without deduplication. Repeated bag positions produce a frequency-weighted vocabulary target, whose entropy need not be the logarithm of bag size; the printed KL correction therefore fails.","primaryUrl":"https://arxiv.org/pdf/2605.06546v2","noteUrl":"/review?paper=review-neurips-2026-155039","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155041","title":"What are Key Factors for Updates in RL for LLM Reasoning?","authors":["Peidong Wang","Ruohan Wang","Xufang Luo","Jiahang Xu","Xiaocui Yang","Shi Feng","Yuqing Yang","Dongsheng Li"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A positive low-probability coupling alone does not ensure the sign condition","plainExplanation":"The precise expansions replace a conditional shape at leading order and suppress a permitted cubic noise term without justification. The ancillary sign condition also needs control of the high-probability group, not merely a positive low-group coupling.","primaryUrl":"https://arxiv.org/html/2606.22570","noteUrl":"/review?paper=review-neurips-2026-155041","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155042","title":"Beyond Prediction: Steering VLM Agents with Retrospective World Modeling","authors":["Yongjiang Liu","Jie ZHANG","Haoyue Zhang","Jingcai Guo","Deze Zeng","Song Guo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Bounded integer scores cannot produce the claimed unbounded reward tail","plainExplanation":"With \\(m\\) actions, every action probability lies between \\(1/[1+(m-1)e^{10}]\\) and \\(1/[1+(m-1)e^{-10}]\\). The resulting log reward, including the \\(\\log m\\) term, therefore has finite extrema. An enclosure extending to \\(-\\infty\\) can still be a valid loose bound; the claim that this tail is attainable is the unsupported step. This narrows the reward-range rationale. It does not refute empirical instability or all reward asymmetry.","primaryUrl":"https://ar5iv.labs.arxiv.org/html/2609.39101","noteUrl":"/review?paper=review-neurips-2026-155042","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155043","title":"MinMax Recurrent Neural Cascades","authors":["Alessandro Ronca"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Bounded inputs and parameters do not uniformly bound gradients over unbounded labels","plainExplanation":"The general cascade claims omit target bounds for uniform gradients and bounded-on-bounded assumptions for arbitrary continuous input maps. Local construction typos separately repeat a coordinate, omit an initialization row and swap identity/reset extrema; the neural subclass and intended corrected constructions have narrower valid claims.","primaryUrl":"https://arxiv.org/pdf/2605.06384v3","noteUrl":"/review?paper=review-neurips-2026-155043","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155045","title":"Inpainting physics: self-supervised learning for context-driven fluid simulation","authors":["Jonas Weidner","Yeray Martin-Ruisanchez","Daniel Rueckert","Benedikt Wiestler","Julian Suk"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Inpainting physics: self-supervised learning for context-driven fluid simulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08832","noteUrl":"/review?paper=review-neurips-2026-155045","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155046","title":"Transcoder Adapters for Reasoning-Model Diffing","authors":["Nathan Hu","Jake Ward","Thomas Icard","Chris Potts"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Transcoder Adapters for Reasoning-Model Diffing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2602.20904v1","noteUrl":"/review?paper=review-neurips-2026-155046","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155047","title":"Training Deliberative Monitors for Black-Box Scheming Detection","authors":["Aditya Sinha","Akshat Naik","Victor Gillioz","Simon Storf","Kilian Merkelbach","Rich Barton-Cooper","Axel Højmark","Marius Hobbhahn"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Training Deliberative Monitors for Black-Box Scheming Detection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.29601","noteUrl":"/review?paper=review-neurips-2026-155047","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155048","title":"Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity","authors":["Loic Cabannes","Pierre-Emmanuel Mazare","Ilze Amanda Auzina","Gergely Szilvasy","Maria Lomeli","Matthijs Douze","Justin Carpentier","Gabriel Synnaeve","Herve Jegou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The sparse row update does not equal the stated dense delta rule","plainExplanation":"Writing \\(z=\\alpha M\\), the dense update is \\(\\beta k(v-k^\\top z)\\), whereas the printed row update is \\(\\beta k_i(v-z_i)\\). With four rows, uniform key entries \\(1/4\\), first-coordinate values \\((1,3,2,2)\\) and target \\(2\\), the shared prediction is already \\(2\\), so the dense update vanishes. The row update changes the first two rows. The discrepancy persists with admissible square product-key shapes. The appendix triangular solve matches the dense rule, not equation (4). A one-hot unit key is a special case where the updates agree.","primaryUrl":"https://arxiv.org/html/2607.07386","noteUrl":"/review?paper=review-neurips-2026-155048","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155049","title":"Transcoda: End-to-End Zero-Shot Optical Music Recognition via Data-Centric Synthetic Training","authors":["Daniel Dratschuk","Paul Swoboda"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Transcoda: End-to-End Zero-Shot Optical Music Recognition via Data-Centric Synthetic Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10835","noteUrl":"/review?paper=review-neurips-2026-155049","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155050","title":"H2G: Hierarchy-Aware Hyperbolic Grouping for 3D Scenes","authors":["ByungHa Ko","Youngmin Lee","Dong Hwan Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Coincident hyperbolic prototypes leave the angle and projection undefined","plainExplanation":"For a singleton, the Einstein midpoint equals its feature. At curvature \\(1\\), take \\(p=s=(1.25,0.75,0)\\); the Lorentz product is \\(-1\\), and the angular expression has numerator and denominator zero. Inward and outward limiting approaches give different angles, so there is no unique continuous extension. Distinct leaves can also coincide, making the Klein-segment projection denominator zero. The geometric loss needs boundary definitions; large batches do not logically exclude these cases. No observed training failure is claimed.","primaryUrl":"https://arxiv.org/html/2605.11967","noteUrl":"/review?paper=review-neurips-2026-155050","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155051","title":"Drifting Fields are not Conservative","authors":["Leonard T. Franz","Sebastian Hoffmann","Tim Weiland","Bernhard Schölkopf","Georg Martius"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Drifting Fields are not Conservative","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.06333","noteUrl":"/review?paper=review-neurips-2026-155051","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155052","title":"STEP: Learning STructured Embeddings for Progressive Time Series","authors":["Lucas Thil","Jesse Read","Rim Kaddah","Guillaume Doquet"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"STEP: Learning STructured Embeddings for Progressive Time Series","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.31061","noteUrl":"/review?paper=review-neurips-2026-155052","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155053","title":"AdaPreLoRA: Adafactor Preconditioned Low-Rank Adaptation","authors":["Ziyun Liu","Fengmiao Bian","Jian-Feng CAI"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The displayed mixed projector uses the wrong Gram inverses","plainExplanation":"Let \\(P=B(B^\\top L^{1/2}B)^{-1}B^\\top L^{1/2}\\) and \\(Q=R^{1/2}A^\\top(AR^{1/2}A^\\top)^{-1}A\\). The projector is \\(PZ+ZQ-PZQ\\). With \\(B=e_1\\), \\(A=e_1^\\top\\), \\(L=R=4I\\) and \\(Z=e_1e_1^\\top\\), a tangent vector must remain \\(Z\\). The displayed mixed term is \\(4Z\\), so the formula returns \\(-2Z\\) and is not idempotent. The proposition display is wrong locally. The proof-end formula and the main half-projector update remain valid on this account.","primaryUrl":"https://arxiv.org/html/2605.08734","noteUrl":"/review?paper=review-neurips-2026-155053","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155054","title":"ZNO: Stable Rational Neural Operators in the Z-Domain for Discrete-Time Dynamics","authors":["Xianli Zhu","Jia Yin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ZNO: Stable Rational Neural Operators in the Z-Domain for Discrete-Time Dynamics","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.02356","noteUrl":"/review?paper=review-neurips-2026-155054","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155056","title":"Training a Predictive Coding Network on ImageNet using Equilibrium Propagation","authors":["Tugdual Kerjan","Rasmus Høier","Benjamin Scellier"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The printed PGD step adds the upper-layer derivative with the wrong sign","plainExplanation":"The stated PGD update reverses its upper-layer derivative sign. Even after that repair, the modified update can converge to an interior point that is not stationary for the original energy, leaving its proposed exact-equilibrium interpretation unjustified.","primaryUrl":"https://arxiv.org/html/2606.03584","noteUrl":"/review?paper=review-neurips-2026-155056","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155058","title":"Platonic Task Arithmetic","authors":["Junghwan Park","Woojin Cho"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Platonic Task Arithmetic","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.00929","noteUrl":"/review?paper=review-neurips-2026-155058","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155059","title":"BONSAI: Bayesian Optimization with Natural Simplicity and Interpretability","authors":["Samuel Daulton","David Eriksson","Maximilian Balandat","Eytan Bakshy"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"BONSAI: Bayesian Optimization with Natural Simplicity and Interpretability","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.07144","noteUrl":"/review?paper=review-neurips-2026-155059","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155060","title":"Segment and Select: Vision-Language Segmentation in 3D Scenarios","authors":["Yulin Chen","Zhihang Zhong","Yuenan Hou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Segment and Select: Vision-Language Segmentation in 3D Scenarios","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.10594","noteUrl":"/review?paper=review-neurips-2026-155060","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155067","title":"Hi-Q: Hierarchical Evidence-guided Query Refinement for Multi-Hop Question Answering","authors":["Jueun Kim","Sungho Park","WOOK SHIN HAN"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Hi-Q: Hierarchical Evidence-guided Query Refinement for Multi-Hop Question Answering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.30468","noteUrl":"/review?paper=review-neurips-2026-155067","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155070","title":"Structure-agnostic Causal Representation Learning","authors":["Arman Behnam","Binghui Wang"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"The HSIC concentration exponent has the wrong kernel scale","plainExplanation":"The SoftMin value identity omits entropy, the HSIC concentration formula loses kernel scaling and bias control, and the CVaR indicator overallocates threshold mass. The convergence proof also reverses the smoothness inequality and needs an additional error-bound condition to control violation values.","primaryUrl":"https://arxiv.org/html/2610.00968","noteUrl":"/review?paper=review-neurips-2026-155070","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155073","title":"The Expressivity Boundary of Probabilistic Circuits: A Comparison with Large Language Models","authors":["Zhiyu Zhao","Xuejie Liu","Muhan Zhang","Anji Liu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The separation-rank upper bound does not establish attainment","plainExplanation":"An upper bound does not establish a matching lower bound. In particular, a circuit can have full interface dimension \\(d\\) and separation rank only \\(d\\) on every nontrivial cut.","primaryUrl":"https://arxiv.org/html/2605.12940","noteUrl":"/review?paper=review-neurips-2026-155073","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155074","title":"Interpreting Neural Combinatorial Optimization via Evolving Programmatic Bottlenecks","authors":["Haocheng Duan","Yuxin Guo","Jieyi Bi","Anqi Xie","Sirui Li","Yining Ma","Cathy Wu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Interpreting Neural Combinatorial Optimization via Evolving Programmatic Bottlenecks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.19741","noteUrl":"/review?paper=review-neurips-2026-155074","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155075","title":"NuMuon: Nuclear-Norm-Constrained Muon for Compressible LLM Training","authors":["Hadi Mohaghegh Dolatabadi","Thalaiyasingam Ajanthan","Sameera Ramasinghe","Chamin Hewa Koneputugodage","Shamane Siriwardhana","Violetta Shevchenko","Karol Pajak","James Snewin","Gil Avraham","Alexander Long"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"NuMuon: Nuclear-Norm-Constrained Muon for Compressible LLM Training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.03597","noteUrl":"/review?paper=review-neurips-2026-155075","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155078","title":"Delayed homomorphic reinforcement learning for environments with delayed feedback","authors":["Jongsoo Lee","Jangwon Kim","Soohee Han"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Delayed homomorphic reinforcement learning for environments with delayed feedback","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.03641","noteUrl":"/review?paper=review-neurips-2026-155078","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155081","title":"Constraint-Aware Influence Estimation in Deep Constrained Learning","authors":["Xin Wang","R.Tyrrell Rockafellar","Xuegang Ban"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The exact constrained quadratic uses a Hessian-metric projection","plainExplanation":"The exact toy constrained solution needs the quadratic’s anisotropic metric rather than Euclidean projection. The deletion influence formulas also use the upweighting sign despite defining change as removed minus original parameters.","primaryUrl":"https://arxiv.org/html/2607.23388","noteUrl":"/review?paper=review-neurips-2026-155081","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155082","title":"Backdoor Attacks under  Lossy Compression: From Failure to Reactivation and Adaptation","authors":["Qian Li","Yunuo Chen","Yuntian Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Backdoor Attacks under  Lossy Compression: From Failure to Reactivation and Adaptation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.13864v2","noteUrl":"/review?paper=review-neurips-2026-155082","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155084","title":"Learning When to Stop: Selective Imitation Learning Under Arbitrary Dynamics Shift","authors":["Surbhi Goel","Jonathan Pei","James Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning When to Stop: Selective Imitation Learning Under Arbitrary Dynamics Shift","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09183","noteUrl":"/review?paper=review-neurips-2026-155084","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155086","title":"Proper Scoring Rules for Agentic Uncertainty Quantification","authors":["Suresh Raghu","Satwik Pandey","Shashwat Pandey"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Proper Scoring Rules for Agentic Uncertainty Quantification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24756","noteUrl":"/review?paper=review-neurips-2026-155086","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155087","title":"Latent Introspection: Models Can Detect Prior Concept Injections","authors":["Theia Pearson-Vogel","Martin Vanek","Raymond Douglas","Jan Kulveit"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Latent Introspection: Models Can Detect Prior Concept Injections","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.20031","noteUrl":"/review?paper=review-neurips-2026-155087","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155088","title":"The two clocks and the innovation window: When and how generative models learn rules","authors":["Binxu Wang","Emma Finn","Bingbin Liu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Permuting one Latin square samples one isotopy orbit","plainExplanation":"If exactly \\(K\\) of \\(d\\) entries are positive, their sum is \\(2K-d\\). For \\(d=36,K=6\\), a valid assignment has sum \\(-24\\) and printed penalty \\(450\\); an invalid assignment with \\(21\\) positives has sum \\(6\\) and penalty zero. This is a local main-text formula error; the cited appendix and implementation convention already avoid it.","primaryUrl":"https://arxiv.org/html/2605.10019","noteUrl":"/review?paper=review-neurips-2026-155088","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155095","title":"Causal Attribution via Activation Patching","authors":["Amirmohammad Izadi","Mohammadali Banayeeanzade","Alireza Mirrokni","Hosein Hasani","Mobin Bagherian","Faridoun Mehri","Mahdieh Baghshah"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Cached patch states do not reproduce evolving states after the intervention cutoff","plainExplanation":"Section 3.3 claims exact equality between Algorithm 1’s patch-wise intervention and a parallel construction that propagates only dedicated classification tokens. The classification-token-only construction freezes patch states that Algorithm 1 continues updating after the intervention cutoff, so their later attention contexts need not agree exactly.","primaryUrl":"https://arxiv.org/html/2603.13652","noteUrl":"/review?paper=review-neurips-2026-155095","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155096","title":"Transformers Linearly Represent Highly Structured World Models","authors":["Roman Kniazev","Nathanaël Fijalkow"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Adding probe directions does not independently overwrite their coordinates","plainExplanation":"For unit probe columns \\(W\\), the intervention \\(WW^\\top\\delta\\) changes probe coordinates by \\((W^\\top W)W^\\top\\delta\\). This equals the intended coordinate change only under additional geometry. Take directions \\(e_1,(-0.8,0.6,0),e_3\\) and \\(\\delta=(1,3,1)\\). The intended changes are \\((1,1,1)\\), but the actual changes are \\((0.2,0.2,1)\\). A threshold bias of \\(-0.5\\) then flips only the third probe. The algebraic overwrite guarantee fails for nonorthogonal probes. The actual probe Gram matrices and empirical intervention effects are not determined here.","primaryUrl":"https://arxiv.org/html/2605.18847","noteUrl":"/review?paper=review-neurips-2026-155096","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155097","title":"Graph Cascades: Contagion-Based Mesoscopic Rewiring for Structure-Aware Graph Machine Learning","authors":["Meher Chaitanya Pindiprolu","My Le","Luana Ruiz"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Graph Cascades: Contagion-Based Mesoscopic Rewiring for Structure-Aware Graph Machine Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05046","noteUrl":"/review?paper=review-neurips-2026-155097","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155101","title":"Thinking in Boxes: 3D Editing in Real Images Made Easy","authors":["Pradhaan Bhat","Naveen Chandra R","Rishubh Parihar","Vaibhav Vavilala","Venkatesh Babu Radhakrishnan","David Forsyth","Anand Bhattad"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Thinking in Boxes: 3D Editing in Real Images Made Easy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.20556","noteUrl":"/review?paper=review-neurips-2026-155101","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155102","title":"Probability-Conserving Flow Guidance","authors":["Parsa Esmati","Junha Hyung","Amirhossein Dadashzadeh","Jaegul Choo","Majid Mirmehdi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Radial guidance can be entirely score-parallel in every dimension","plainExplanation":"Isotropic Jacobians do not ensure dimension-suppressed divergence of score-parallel guidance. A realizable Gaussian pair gives entirely parallel residual guidance and divergence ratio one in every dimension.","primaryUrl":"https://arxiv.org/html/2605.20079","noteUrl":"/review?paper=review-neurips-2026-155102","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-155104","title":"DiffRatio: Training One-Step Diffusion Models Without Teacher Supervision","authors":["Wenlin Chen","Mingtian Zhang","Jiajun He","Zijing Ou","José Miguel Hernández-Lobato","Bernhard Schölkopf","David Barber"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"More candidates can increase noisy-ratio target bias","plainExplanation":"The noisy density ratio can make two-candidate resampling farther from the target than one-candidate sampling. The counterexample persists for smooth full-support mixtures and arbitrarily small absolute noise after scaling, so the monotonicity guarantee needs stronger conditions.","primaryUrl":"https://arxiv.org/html/2502.08005","noteUrl":"/review?paper=review-neurips-2026-155104","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155105","title":"Network Learning with Semi-relaxed Gromov-Wasserstein","authors":["Charles Dufour","Ulysse Naepels","Leonardo V Santoro"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Heaviest-block rounding can exceed the stated objective-gap bound","plainExplanation":"Heaviest-block rounding can exceed the stated objective-gap bound. Stationarity does not imply the global statistical oracle bound.","primaryUrl":"https://arxiv.org/html/2606.02223","noteUrl":"/review?paper=review-neurips-2026-155105","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-155106","title":"Mixture-Trained Merging for Unified Multi-Objective Models","authors":["SeongHyeon Kim","Chaeyun Jang","Seungyoo Lee","Jiyeon Ham","Yunju Bak","Boseop Kim","Juho Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The constrained specialist update needs an empty-set rule","plainExplanation":"The specialist argmax can have an empty feasible set, so the next mixture needs a fallback or an explicit feasibility condition. For branches already trained on mixtures, local merge weights correspond to composed dataset weights; Eq. (7) already implements that composition correctly. The formal mean-score promotion rule is well defined, but it does not imply the appendix’s weakest-objective and boundary-avoidance description.","primaryUrl":"https://arxiv.org/html/2610.01238","noteUrl":"/review?paper=review-neurips-2026-155106","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155107","title":"VGB for Masked Diffusion Model: Efficient Test-time Scaling for Reward Satisfaction and Sample Editing","authors":["Kijung Jeon","Thuy-Duong Vuong","Molei Tao"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Conductance minima use different admissible cut sets","plainExplanation":"Conductance minima use different admissible cut sets. The un-lazified lifted chain can have period two. Exact verifier accuracy does not imply balanced move rates.","primaryUrl":"https://arxiv.org/html/2606.28301","noteUrl":"/review?paper=review-neurips-2026-155107","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155109","title":"Zeroth-Order Sharpness-Aware Learning with Exponential Tilting","authors":["Xuchen Gong","Tian Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The ball-to-sphere gradient approximation can retain unit bias","plainExplanation":"The ball-to-sphere gradient approximation can retain unit bias. The sensitivity formula uses a gradient coordinate in place of an eigenvalue. A zero-gradient eigendirection can determine the sharpness optimum.","primaryUrl":"https://arxiv.org/html/2510.16157","noteUrl":"/review?paper=review-neurips-2026-155109","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155110","title":"Online Set Learning from Precision and Recall Feedback","authors":["Lee Cohen","Yishay Mansour","Shay Moran","Han Shao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A nonempty filtered class can contain only a suboptimal set","plainExplanation":"A nonempty filtered class can contain only a suboptimal set.","primaryUrl":"https://arxiv.org/html/2605.09565","noteUrl":"/review?paper=review-neurips-2026-155110","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155112","title":"Estimating Implicit Regularization in Deep Learning","authors":["Joseph H Rudoler","Kevin Tan","Giles Hooker","Konrad Kording"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The PAC theorem compares an unsquared error with a squared-scale bound","plainExplanation":"The inverse-design factor scales as the inverse square root of the eigenvalue. The PAC theorem compares an unsquared error with a squared-scale bound. Pairwise martingale differences do not control the empirical noise floor.","primaryUrl":"https://arxiv.org/html/2605.05436","noteUrl":"/review?paper=review-neurips-2026-155112","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155113","title":"Beyond Modality Fusion: Deep Ensembles for Multimodal Classification","authors":["Ilya Burenko","Dmitry Vetrov"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Beyond Modality Fusion: Deep Ensembles for Multimodal Classification","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.05019","noteUrl":"/review?paper=review-neurips-2026-155113","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155114","title":"Forgetting is Not Always Bad: A Neuro-Inspired Memory Repair Mechanism for Poisoned LLM Agents","authors":["Lei Liu","Yunji Liang","Xiaowen Zhang","Jingqi Liu","Qi Li","Bin Guo","Zhiwen Yu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Forgetting is Not Always Bad: A Neuro-Inspired Memory Repair Mechanism for Poisoned LLM Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.13267","noteUrl":"/review?paper=review-neurips-2026-155114","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155118","title":"Selective Critique for Cost-Aware LLM Agents in Long-Horizon Decision Making","authors":["Heewon Park","Somin Im","Minhae Kwon"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Selective Critique for Cost-Aware LLM Agents in Long-Horizon Decision Making","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07335","noteUrl":"/review?paper=review-neurips-2026-155118","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155119","title":"VLA-Pro: Cross-Task Procedural Memory Transfer for Vision-Language-Action Models","authors":["Shengyu Si","Yuanzhuo Lu","Ruimeng Yang","Ziyi Ye","Zuxuan Wu","Yu-Gang Jiang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"VLA-Pro: Cross-Task Procedural Memory Transfer for Vision-Language-Action Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.29562","noteUrl":"/review?paper=review-neurips-2026-155119","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155120","title":"SAIR: Cost-Efficient Multi-Stage ML Pipeline Autoscaling via In-Context Reinforcement Learning","authors":["Jianchang Su","Yifan Zhang","Shengkai Lin","Shizhen Zhao","YUSHENG ZHENG","Yiwei Yang","Wei Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"An absolute reward bound gives twice the stated one-step gap","plainExplanation":"The difference between two allowed rewards can be \\(2R_{\\max}\\). With four actions having rewards \\((2,-2,-2,-2)\\), uniform exploration has expected regret \\(3\\). At exploration rate \\(0.15\\), its contribution is \\(0.45\\), exceeding the printed \\(0.30\\) term even when exploitation is optimal and the other errors vanish. The constant in the theorem is unsupported under the absolute bound. The qualitative rates are not disproved.","primaryUrl":"https://arxiv.org/html/2601.22397","noteUrl":"/review?paper=review-neurips-2026-155120","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155121","title":"SCION: Scene Composition via Instanced Neural Primitives","authors":["William Koch","Amogh Joshi","Cyrus Vachha","Cheng Zheng","Felix Heide"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SCION: Scene Composition via Instanced Neural Primitives","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02322","noteUrl":"/review?paper=review-neurips-2026-155121","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155122","title":"Block Sparse Flash Attention","authors":["Daniel Ohayon","Itay Lamprecht","Itay Hubara","Israel Cohen","Daniel Soudry","Noam Elata"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Rectangular attention tiles do not share a single same-index causal boundary","plainExplanation":"Algorithm 1 accepts generic query/key tile sizes \\(B_M,B_N\\), enumerates past blocks by \\(j<i\\), and masks only the same-index block \\(j=i\\). For rectangular tiles, equal block indices do not identify the causal boundary; the printed loop can omit past keys or leave future keys unmasked.","primaryUrl":"https://arxiv.org/html/2512.07011","noteUrl":"/review?paper=review-neurips-2026-155122","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155125","title":"Tracing Moral Foundations in Large Language Models","authors":["Chenxiao Yu","Bowen Yi","Farzan K Malekabadi","Suhaib Abdurahman","Jinyi Ye","Shrikanth Narayanan","Yue Zhao","Morteza Dehghani"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Tracing Moral Foundations in Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.05437","noteUrl":"/review?paper=review-neurips-2026-155125","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155126","title":"Geometry Conflict: Explaining and Controlling Forgetting in LLM Continual Post-Training","authors":["Yuanyi Wang","Yifan Yang","Su Lu","Yanggan Gu","Pengkai Wang","Wenjun Wang","Zhaoyi Yan","Congkai Xie","Jianmin Wu","Jialun Cao","Shing-Chi Cheung","Hongxia Yang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Geometry Conflict: Explaining and Controlling Forgetting in LLM Continual Post-Training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.09608","noteUrl":"/review?paper=review-neurips-2026-155126","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155131","title":"ViT-AdaLA: Adapting Vision Transformers with Linear Attention","authors":["Yifan Li","Seunghyun Yoon","Viet Lai","Franck Dernoncourt","Jason Kuen","Yu Kong","Trung Bui"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ViT-AdaLA: Adapting Vision Transformers with Linear Attention","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.16063","noteUrl":"/review?paper=review-neurips-2026-155131","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155132","title":"Non-Asymptotic Best Policy Identification Guarantees in Online Reinforcement Learning","authors":["Joseph Lazzaro","Alessio Russo","Aldo Pacchiano"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The finite-confidence proof loses factors in four intermediate bounds","plainExplanation":"Theorem 33’s final allowance replaces this contribution by $2t\\ell(t)\\log L$. It is smaller than the term established by the preceding proof.","primaryUrl":"https://arxiv.org/html/2607.17201","noteUrl":"/review?paper=review-neurips-2026-155132","collection":"neurips-2026","severity":"repairable","issueTypes":["mathematical_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155133","title":"Computing All Optimal Partial $p$-Wasserstein Matchings on the Line","authors":["Sebastian Angrick","Jacobus Conradi","Mónika Csikós","Niko Mertens","Danny Mittal","André Nusser","Krzysztof Onak","Sharath Raghvendra"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The public summary extends an integer-power algorithm to all real powers","plainExplanation":"With the printed coordinates and \\(m\\leq n\\), the queried right interval misses the shifted gadget, so the balanced subproblem used by the reduction is not the intended one. The independent check recovers the intended gadget costs after consistent translation; the error is the mismatch between point coordinates, intervals and the associated affine cost. This is a repairable reduction-coordinate error. The strict-convexity separation for \\(p>1\\) and the corrected lower-bound construction require the stated consistent bookkeeping.","primaryUrl":"https://arxiv.org/html/2608.18875","noteUrl":"/review?paper=review-neurips-2026-155133","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155134","title":"CaMeLs Can Use Computers Too: System-level Security for Computer Use Agents","authors":["Hanna Foerster","Tom Blanchard","Kristina Nikolić","I Shumailov","Cheng Zhang","Robert Mullins","Nicolas Papernot","Florian Tramer","Yiren Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"CaMeLs Can Use Computers Too: System-level Security for Computer Use Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2601.09923","noteUrl":"/review?paper=review-neurips-2026-155134","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155137","title":"Auteur: Language-Driven Cinematographic Framing for Human-Centric Video Generation","authors":["Muhammed Burak Kızıl","Enes Sanli","Niloy Mitra","Xuelin Chen","Erkut Erdem","Aykut Erdem","Duygu Ceylan"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Auteur: Language-Driven Cinematographic Framing for Human-Centric Video Generation","plainExplanation":"A variable needed at geometric decode time need not be sampled first in a joint distribution.","primaryUrl":"https://arxiv.org/html/2606.01900","noteUrl":"/review?paper=review-neurips-2026-155137","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","implementation_mismatch"],"classifications":["major"]},{"slug":"review-neurips-2026-155138","title":"SphMind: Towards Robust, Training-Free VLM-based Spatial Reasoning with a 360 Camera","authors":["Shriram Damodaran","Soumyaratna Debnath","Cheston Tan","Lin Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The softmax partition cancels from the log-probability-ratio gradient","plainExplanation":"Appendix B.1 calls its probability-dependent contrastive-energy vector the exact gradient and uses that dependence to justify automatic scaling with violation. The two log-softmax partition terms cancel exactly, so the printed probability-dependent gradient contains an extra term and does not justify the claimed self-scaling. The printed derivative and its self-scaling justification fail; the claim does not identify the gradient used in the evaluated implementation.","primaryUrl":"https://arxiv.org/html/2609.33462","noteUrl":"/review?paper=review-neurips-2026-155138","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155140","title":"EmpathyChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue","authors":["Dingdong WANG","Shujie LIU","Jinyu Li","Yuxuan Hu","Yayue Deng","Yunrui Cai","Jincenzi Wu","Jianwei Yu","Helen Meng"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"EmpathyChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04826","noteUrl":"/review?paper=review-neurips-2026-155140","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155143","title":"EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts","authors":["Minseo Kim","Minjae Lee","Seunghyuk Oh","Kevin Galim","Donghoon Kim","Coleman Hooper","Harman Singh","Amir Gholami","HYUNG IL KOO","Wonjun Kang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.18967","noteUrl":"/review?paper=review-neurips-2026-155143","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155144","title":"You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning","authors":["Yian Wang","Ali Ebrahimpour-Boroojeny","Hari Sundaram","Varun Chandrasekaran"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning","plainExplanation":"Fattened geometric overlap need not carry a positive displacement floor. A valid lower bound needs actual non-target probability mass and quantitative control of displacement near the target region; calibration of the sample overlap statistic requires a separate sampling-measure argument.","primaryUrl":"https://arxiv.org/html/2609.34137","noteUrl":"/review?paper=review-neurips-2026-155144","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","incorrect_bound","missing_assumption","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-155145","title":"Auditing Instruction Robustness in Vision-Language-Action Models via Diversity-Aware Red Teaming","authors":["Baoshun Tong","Haoran He","Yang Liu","Ling Pan","Liang Lin"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The printed group-centering formula cancels every reward","plainExplanation":"The group-centering formula repeats the current reward inside its mean, so subtraction cancels every reward exactly. The averaging summand needs the group index instead. This is a printed centering-index typo, visible in the primary PDF as well as the text. Other training terms may still update the policy; no full-training stagnation or empirical failure follows.","primaryUrl":"https://arxiv.org/html/2604.05595","noteUrl":"/review?paper=review-neurips-2026-155145","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-155146","title":"Retrieval Heads Meet Vision: Uncovering How VLMs Locate and Extract Visual Information","authors":["Chanho Park","Daehyeon Choi","Jihyun Lee","Minhyuk Sung"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Retrieval Heads Meet Vision: Uncovering How VLMs Locate and Extract Visual Information","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.27417","noteUrl":"/review?paper=review-neurips-2026-155146","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155147","title":"Soft geometric inductive bias for object centric dynamics","authors":["Hampus Linander","Conor Heins","Marco Perin","Alexander Tschanz","Christopher L Buckley"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The rotation example has a sign mismatch; the equivariance qualification is explicit","plainExplanation":"With the stated sandwich convention, the positive bivector exponent rotates in the opposite direction to the displayed coordinate formula.","primaryUrl":"https://arxiv.org/html/2512.15493","noteUrl":"/review?paper=review-neurips-2026-155147","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error","notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-155149","title":"Trapping Attacker in Dilemma: Examining Internal Correlations and External Influences of Trigger for Defending GNN Backdoors","authors":["Fan Yang","Binyan Xu","Di Tang","Kehuan Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Trapping Attacker in Dilemma: Examining Internal Correlations and External Influences of Trigger for Defending GNN Backdoors","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08278","noteUrl":"/review?paper=review-neurips-2026-155149","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155150","title":"CAST: Causal Anchored Simplex Transport for Distribution-Valued Time Series","authors":["Jiecheng Lu","Jieqi Di","Runhua Wu","Yuwei Zhou"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The drift bound confuses raw and attenuated mixing weights","plainExplanation":"The actual drift is \\(\\rho_{\\mathrm{raw}}g|\\delta|\\leq\\rho_{\\mathrm{raw}}B\\). If \\(\\rho_{\\mathrm{eff}}=\\rho_{\\mathrm{raw}}g\\), it is not generally at most \\(\\rho_{\\mathrm{eff}}B\\), since that would require \\(|\\delta|\\leq B\\). Local transport within a unit radius can still have displacement \\(0.95\\), so locality alone does not impose an arbitrary smaller budget. The drift constant is misstated. Simplex closure and the separate Wasserstein locality property survive.","primaryUrl":"https://arxiv.org/html/2605.16919","noteUrl":"/review?paper=review-neurips-2026-155150","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155151","title":"Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs","authors":["Jonathan Zheng","Zirui Shao","Alan Ritter","Wei &amp;quot;Coco&amp;quot; Xu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.00184","noteUrl":"/review?paper=review-neurips-2026-155151","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155152","title":"Can Linguistic Reasoning Vectors Enhance Multimodal Reasoning Ability?","authors":["Ziyi Wang","Li Li","Aolin Zhou","Yankun Shen","Liu Chonghan","Shuxia Lin","Xu Yang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Can Linguistic Reasoning Vectors Enhance Multimodal Reasoning Ability?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31140","noteUrl":"/review?paper=review-neurips-2026-155152","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155153","title":"ChunkFT: Byte-Streamed Optimization for Memory-Efficient Full Fine-Tuning","authors":["Yongkang Liu","Zijing Wang","Mengjie Zhao","Ercong Nie","Mingyang Wang","Qian Li","Feiliang Ren","Shi Feng","Daling Wang","Hinrich Schuetze"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Weight decay can create a fixed point with nonzero loss gradient","plainExplanation":"Weight decay can create a fixed point with nonzero loss gradient. Unbiased gradient noise still adds a nonzero descent remainder.","primaryUrl":"https://arxiv.org/html/2605.21177","noteUrl":"/review?paper=review-neurips-2026-155153","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155154","title":"Exact Posterior Score Estimation for Solving Linear Inverse Problems","authors":["Abbas Mammadov","Ozgur Kara","Kaan Oktay","Iskander Azangulov","Adil K Akan","Hyungjin Chung","James Rehg","Yee Whye Teh"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A finite highest-noise denoiser is not exactly the posterior mean","plainExplanation":"For a standard Gaussian prior in two dimensions and an observation of only coordinate one, the unobserved coordinate satisfies \\(\\mathbb E[x_{0,2}\\mid x_{t,2},y]=x_{t,2}/(1+\\sigma^2)\\), whereas \\(\\mathbb E[x_{0,2}\\mid y]=0\\). At \\(\\sigma=80\\) these differ whenever the noisy coordinate is nonzero; the excess squared risk under the noisy marginal is \\(1/6401\\). The asymptotic limit and fixed-noise score identity can remain valid. The example does not imply a large practical error.","primaryUrl":"https://arxiv.org/html/2606.17048","noteUrl":"/review?paper=review-neurips-2026-155154","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155156","title":"NP-LoRA: Null Space Projection for Subject-Style LoRA Fusion","authors":["Chuheng Chen","Xiaofei Zhou","Geyuan Zhang","Yong Huang","Gang Xiong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A right-singular-space projector must multiply the update on the right","plainExplanation":"Equation (20) defines overlap by left-multiplying a content update \\(C\\in\\mathbb R^{m\\times n}\\) with the style right-space projector \\(P=VV^{\\top}\\in\\mathbb R^{n\\times n}\\). The style right-singular projector acts in the input space and must multiply the content update on the right; the displayed left product is generally dimensionally undefined.","primaryUrl":"https://arxiv.org/html/2511.11051","noteUrl":"/review?paper=review-neurips-2026-155156","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155157","title":"SURF: Steering the Scalarization Weight to Uniformly Traverse the Pareto Front","authors":["Liuyuan Jiang","Chentong Huang","Lisha Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An entropy gradient singularity does not create a boundary value barrier","plainExplanation":"Appendix D.2 infers that every scalarized sublevel set initialized in the simplex interior stays uniformly away from the boundary because the entropy gradient diverges there. A diverging entropy gradient does not make the entropy value diverge at the boundary, so the initial sublevel set need not stay in a compact interior region. The sublevel-set and inner-trajectory regularity argument fails. This does not refute the exact optimal-path geometry or exact-solver results.","primaryUrl":"https://arxiv.org/html/2605.20619","noteUrl":"/review?paper=review-neurips-2026-155157","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155158","title":"Enhancing Gaze Reasoning in Vision Foundation Models for Gaze Following","authors":["SHIJING WANG","Yihua Cheng","CHAOQUN CUI","Yaping Huang","David Wong","Alexandros Neophytou","Hyung Jin Chang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Enhancing Gaze Reasoning in Vision Foundation Models for Gaze Following","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22607","noteUrl":"/review?paper=review-neurips-2026-155158","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155159","title":"ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models","authors":["Arash Akbari","Arman Akbari","Masih Eskandar","Qitao Tan","Yixiao Chen","Jingwu Luo","Bertha Pangaribuan","Liyun Zhang","Jennifer Dy","Geng Yuan","Xue Lin","Gaowen Liu","Stratis Ioannidis","Yanzhi Wang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Signed tensor sensitivities can cancel in the squared layer-error proxy","plainExplanation":"Equation (5) subtracts two HSIC terms, producing signed sensitivities, while Equation (6) squares their noise-weighted sum and interprets it as a noncancelling reconstruction-error penalty with beneficial precision upgrades. Signed HSIC sensitivities can cancel before the layer sum is squared, so reducing one tensor’s distortion can increase the supposed reconstruction-error penalty.","primaryUrl":"https://arxiv.org/html/2605.24011","noteUrl":"/review?paper=review-neurips-2026-155159","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155161","title":"MoRSE: Task-Oriented Multi-Agent System with Mixture of Role-Subtask Experts","authors":["Peiwen Li","Shiyang Zhang","Yangtian Zhang","Sizhuang He","David van Dijk","Rex Ying"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Conditional reward centering does not eliminate score-weighted covariance","plainExplanation":"Conditional reward centering does not eliminate score-weighted covariance. Activation variance does not identify optimal adapter directions.","primaryUrl":"https://arxiv.org/html/2608.09251","noteUrl":"/review?paper=review-neurips-2026-155161","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155167","title":"SPACE: Unifying Symmetric and Asymmetric Routing Problems for Generalist Neural Solver","authors":["Rongsheng Chen","Changliang Zhou","Canhong Yu","Yuanyao Chen","Yu Zhou","Zhuo Chen","Zhenkun Wang"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"SPACE: Unifying Symmetric and Asymmetric Routing Problems for Generalist Neural Solver","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24484","noteUrl":"/review?paper=review-neurips-2026-155167","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155168","title":"Bandits via Additive Quantized Representations","authors":["Ami Tavory","Noam Touitou","Tal Sarig","Frank Cheng","Ido Guy"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The three posterior axioms do not force score concentration","plainExplanation":"With one centroid and constant contexts, give each arm a Gaussian score with its empirical centered reward as mean and variance \\(1\\). Rewards \\(1\\) and \\(0\\) satisfy the stated mean and starvation conditions. Both arms are sampled infinitely often, but their limiting scores still differ by a Gaussian with mean \\(1\\) and variance \\(2\\). The inferior arm is selected with positive limiting probability, even though the clipped predicted means tend to the correct extremes and the claimed limiting bound is zero. The axioms are insufficient. The concrete implementations with shrinking posterior variance are not refuted by this nonconcentrating construction.","primaryUrl":"https://arxiv.org/html/2610.02440","noteUrl":"/review?paper=review-neurips-2026-155168","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155173","title":"Bilevel Optimization of Synthetic Trajectories for Multi-Turn LLM Fine-Tuning","authors":["Shresth Verma","Mauricio Tec","Cheol Woo Kim","Kai Wang","Milind Tambe"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Graded attribute reward is not a hitting-time objective","plainExplanation":"Graded attribute reward is not a hitting-time objective.","primaryUrl":"https://arxiv.org/html/2605.24743","noteUrl":"/review?paper=review-neurips-2026-155173","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155174","title":"Efficient Test-Time Adaptation For Robot Policies","authors":["Motasem Alfarra","Pietro Mazzaglia","Markus Peschl","Daniel Dijkman","Christos Louizos"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The reciprocal hardness score reverses order across its pole","plainExplanation":"The derivative is negative on each side of the pole, but the function is not globally decreasing across it. An allowed distributional critic on support \\([-1,1]\\) can have values \\(-0.5\\) and \\(0.5\\). With \\(\\varepsilon=0.01\\), their scores are approximately \\(-2.0408\\) and \\(1.9608\\), so minimum-score eviction removes the harder state. Minimum-subtracted sampling retains this reversed order. The allowed value \\(Q=-\\varepsilon\\) is also undefined. The general hardness guarantee needs a domain condition. A particular implementation with positive shifted support could already avoid the issue.","primaryUrl":"https://arxiv.org/html/2610.20318v1","noteUrl":"/review?paper=review-neurips-2026-155174","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155181","title":"Detecting RLVR Training Data via Structural Convergence of Reasoning","authors":["Hongbo Zhang","Leyang Cui","Jianhao Yan","Guangsheng Bao","Yue Zhang","Shuo Yang","Yue Zhang"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Detecting RLVR Training Data via Structural Convergence of Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.11792","noteUrl":"/review?paper=review-neurips-2026-155181","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155182","title":"Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis","authors":["Thong Bach","Dung Nguyen","Thao Le","Truyen Tran"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The aggregate token ratio is not a sufficient statistic for the retained logits","plainExplanation":"The retained token logits can distinguish cases that have the same aggregate refusal ratio. Factorization also does not make all masked positions identical, and constant logit norms do not make cosine displacement monotone in KL divergence. The moving-mask energy bound needs a separate jump term.","primaryUrl":"https://arxiv.org/html/2609.30841","noteUrl":"/review?paper=review-neurips-2026-155182","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155185","title":"The Geometric Wall: Manifold Structure Predicts Layerwise Sparse Autoencoder Scaling Laws","authors":["Eslam Zaher","Maciej Trzaskowski","Quan Nguyen","Fred Roosta"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Dense atoms do not imply a positive curvature reconstruction floor","plainExplanation":"On \\(\\mathbb R^2\\), take logits \\(z=(x,y,0)\\) and probabilities \\(p=\\operatorname{softmax}(z)\\). The logit Jacobian gives a positive-definite metric \\(G=J^\\top(\\operatorname{diag}p-pp^\\top)J\\). Under \\(s=2\\sqrt p\\), this is the induced metric of the radius-two sphere, with Gaussian curvature \\(1/4\\). The Euclidean metric on the same coordinate domain has curvature zero. Curvature depends on metric derivatives, not merely its pointwise matrix congruence. The background justification fails; no numerical fit depends on this equality. The weaker intrinsic-dimension statement need not change.","primaryUrl":"https://arxiv.org/html/2605.09887","noteUrl":"/review?paper=review-neurips-2026-155185","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155186","title":"AR1-ZO: Topology-Aware Rank-1 Zeroth-Order Queries for High-Rank LoRA Fine-Tuning","authors":["Ziye Chen","Hongbin Lin","Chenyu Zhang","Xiangda Yan","Yongjie Yang","Yao SHU"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Endpoint-noise cancellation can leave a positive limiting cosine","plainExplanation":"Endpoint-noise cancellation can leave a positive limiting cosine. Gradient Lipschitzness alone does not give quadratic central-estimator bias. The displayed noise term prevents the claimed smoothing-and-rate limit.","primaryUrl":"https://arxiv.org/html/2605.19767","noteUrl":"/review?paper=review-neurips-2026-155186","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155188","title":"Posterior-First Neural PDE Simulation: Inferring Hidden Problem State from a Single Field","authors":["Wenshuo Wang","Fan Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Posterior-First Neural PDE Simulation: Inferring Hidden Problem State from a Single Field","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.03247","noteUrl":"/review?paper=review-neurips-2026-155188","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155192","title":"Resolving Representation Ambiguity in Feedforward Novel View Synthesis Transformer via Semantic-Spatial Decoupling","authors":["Yihang Wu","Yihang Sun","Shaofeng Zhang","Zuxuan Wu","Junchi Yan","Xiaosong Jia","Yu-Gang Jiang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Resolving Representation Ambiguity in Feedforward Novel View Synthesis Transformer via Semantic-Spatial Decoupling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18599","noteUrl":"/review?paper=review-neurips-2026-155192","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155193","title":"Zeroth-Order Stackelberg Control in Combinatorial Congestion Games","authors":["Saeed Masiha","Sepehr Elahi","Negar Kiyavash","Patrick Thiran"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A ZDD traversal can visit many excluded elements","plainExplanation":"Coordinatewise parameter sensitivity accumulates across cost coordinates. A ZDD traversal can visit many excluded elements. Ambient interiorization can be empty on an equality-constrained domain.","primaryUrl":"https://arxiv.org/html/2602.23277","noteUrl":"/review?paper=review-neurips-2026-155193","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155196","title":"Geometric Inductive Biases for Semi-Supervised Equalization: The Constellation-Aware Transformer","authors":["Avi Caciularu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Geometric Inductive Biases for Semi-Supervised Equalization: The Constellation-Aware Transformer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.33695","noteUrl":"/review?paper=review-neurips-2026-155196","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155197","title":"Inference and Uncertainty Quantification for Streaming $r$-PCA","authors":["Haoshu Xu","Hongzhe Li"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Inference and Uncertainty Quantification for Streaming $r$-PCA","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.18374","noteUrl":"/review?paper=review-neurips-2026-155197","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155198","title":"StableHand: Quality-Aware Flow Matching for World-Space Dual-Hand Motion Estimation from Egocentric Video","authors":["Huajian Zeng","Chaohua Yao","Yuantai Zhang","Jiaqi Yang","Rolandos Alexandros Potamias","Xingxing Zuo"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Noise matching and pooled calibration need narrower claims","plainExplanation":"The inference initialization matches the noise amplitudes but changes the signal anchor, and a pooled error-distribution loss does not determine separate wrist and finger calibration percentiles. Appendix E also misdescribes the rotation-aware wrist error as translation-only; its angular attenuation is already identified as a variant.","primaryUrl":"https://arxiv.org/html/2605.18553","noteUrl":"/review?paper=review-neurips-2026-155198","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","false_statement","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155199","title":"OntoPlan: An Ontology-Grounded Scene Representation and Agentic Framework for Scalable Robot Task Planning","authors":["Hyeongwoo Nam","Woongje Cho","Juwon Kim","Jongeun Choi"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"OntoPlan: An Ontology-Grounded Scene Representation and Agentic Framework for Scalable Robot Task Planning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07649","noteUrl":"/review?paper=review-neurips-2026-155199","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155200","title":"Prediction-Powered Active Testing","authors":["Kianoosh Ashouritaklimi","Valentin Kilian","Daolang Huang","Thomas Rainforth","Francois Caron"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Prediction-Powered Active Testing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.08347","noteUrl":"/review?paper=review-neurips-2026-155200","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155201","title":"Learning Rate Transfer in Normalized Transformers","authors":["Boris Shigida","Boris Hanin","Andrey Gromov"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Learning Rate Transfer in Normalized Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.27077","noteUrl":"/review?paper=review-neurips-2026-155201","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155203","title":"Non-Linear Pricing Restores Tractability for a Data Seller","authors":["Bhaskar Ray Chaudhury","Jugal Garg","Eklavya Sharma","Jiaxin Song"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Non-Linear Pricing Restores Tractability for a Data Seller","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.36589","noteUrl":"/review?paper=review-neurips-2026-155203","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155208","title":"MineEvolve: Self-Evolution with Accumulated Knowledge for Long-Horizon Embodied Minecraft Agents","authors":["Zhengwei Xie","Zhisheng Chen","Ziyan Weng","Jinhan Li","Chenglong Li","Zikai Xiao","Jinhao Jing","Jingwei Song","Guibin Zhang","Kun Wang"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"MineEvolve: Self-Evolution with Accumulated Knowledge for Long-Horizon Embodied Minecraft Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.13131","noteUrl":"/review?paper=review-neurips-2026-155208","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155209","title":"GOLD PANNING: Strategic Context Shuffling for Needle-in-Haystack Reasoning","authors":["Adam Byerly","Daniel Khashabi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Detector separation and KL drift can order positions differently","plainExplanation":"Detector separation and KL drift can order positions differently. Top-rank membership alone does not control expected information drift.","primaryUrl":"https://arxiv.org/html/2510.09770","noteUrl":"/review?paper=review-neurips-2026-155209","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155211","title":"Verifiers in the Loop: Decoding Time Verification for Code Translation","authors":["Tianyang Zhou","Somesh Jha","Mihai Christodorescu","Kirill Levchenko","Varun Chandrasekaran"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Passing currently applicable checks does not prove a prefix has a valid completion","plainExplanation":"Section 2.2 says unanimous applicable-oracle acceptance enforces that every committed prefix can be extended to a program passing all oracles. Passing all currently applicable oracles establishes local admissibility, not existence of a full completion satisfying later-scope checks.","primaryUrl":"https://arxiv.org/html/2605.17626","noteUrl":"/review?paper=review-neurips-2026-155211","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155213","title":"Neural Galerkin Normalizing Flows for Bayesian Inference of Diffusions with Inaccessible Boundaries","authors":["Riccardo Saporiti","Fabio Nobile"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Neural Galerkin Normalizing Flows for Bayesian Inference of Diffusions with Inaccessible Boundaries","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04324","noteUrl":"/review?paper=review-neurips-2026-155213","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155215","title":"IGGT4D: Streaming 4D Instance-Grounded Geometry Transformer","authors":["Zhengyu Zou","Hao Li","kuixuan jiao","Liu.Liu","TingyangXiao","xiaolin.zhou","Fangzhou Hong","Zhizhong Su","Dingwen Zhang","Ziwei Liu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"IGGT4D: Streaming 4D Instance-Grounded Geometry Transformer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.19228","noteUrl":"/review?paper=review-neurips-2026-155215","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"tokenpb-average-calibration-guarantee","title":"TokenPB: Enabling VLA Action Self-Verification via VLM Token Probability Bucketing","authors":["Chen Zhao","Zhuoran Wang","Haoyang Li","Shifeng Bao","Guanlin Li","Youhe Feng","Yang Li","Jie Tang","Jing Zhang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Calibration and score normalization support narrower guarantees","plainExplanation":"Mean calibration error does not guarantee correct ranking at every observation, and the Gibbs utility bound omits the score-to-utility scale. Two additional gaps concern transferring fixed-length sequence-ranking and properness results to scores and losses normalized by each sequence’s own length.","primaryUrl":"https://arxiv.org/html/2603.18091v2","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","invalid_inference","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-155217","title":"Support-Safe Variational Hybrid Filtering for Contact-Mode and Sparse-Law Recovery","authors":["Marios Papamichalis","Regina Ruane"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Support-Safe Variational Hybrid Filtering for Contact-Mode and Sparse-Law Recovery","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16398","noteUrl":"/review?paper=review-neurips-2026-155217","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155218","title":"Fixed-Point Masked Generative Modeling","authors":["Andrea Miele","Yiming QIN","Alba Carballo Castro","Justin Deschenaux","Pascal Frossard"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Fixed-Point Masked Generative Modeling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.31215","noteUrl":"/review?paper=review-neurips-2026-155218","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155222","title":"Physics-Informed Video Generation via Mixture-of-Experts Latent Alignment","authors":["Cong Wang","Hanxin Zhu","Jiayi Luo","Yonglin Tian","Xiaoqian Cheng","Peiyan Tu","Xin Jin","Long Chen","Zhibo Chen"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Soft label-biased routing can leave the loss normalizer zero","plainExplanation":"Let the label set be \\(\\{0\\}\\), the finite label bias be \\(2\\), and learned logits be \\((0,5,6,7,8,0,0,0)\\). After biasing, the top four experts are \\(4,3,2,1\\), none label-consistent. Both the selected residual numerator and its normalizer are empty sums, hence zero. Positive softmax weights cannot make an empty intersection nonempty. This is a boundary-definition gap in the written objective; actual NaNs or missing code guards are not established.","primaryUrl":"https://arxiv.org/html/2606.04737","noteUrl":"/review?paper=review-neurips-2026-155222","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155223","title":"Low-Dimensional Adaptation of Rectified Flow: A  Diffusion and Stochastic Localization Perspective","authors":["Saptarshi Roy","Alessandro Rinaldo","Purnamrita Sarkar"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The posterior-covariance coefficient fails on a bounded two-point target","plainExplanation":"The posterior-covariance coefficient fails on a bounded two-point target.","primaryUrl":"https://arxiv.org/html/2601.15500","noteUrl":"/review?paper=review-neurips-2026-155223","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-155224","title":"$\\mathbf{D^{3}S^{2}}$: Diffusion-Guided Dataset Distillation for Semantic Segmentation","authors":["Wenjie Zheng","Haoji Hu","Jiali Lu","Xingze Zou","Jing Wang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"$\\mathbf{D^{3}S^{2}}$: Diffusion-Guided Dataset Distillation for Semantic Segmentation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.25022","noteUrl":"/review?paper=review-neurips-2026-155224","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155225","title":"BLARM: Animating 3D Objects from Video via Blending LAtent Rigid Motion Primitives","authors":["Pradyumn Goyal","Yizhak Ben-Shabat","Hsueh-Ti Derek Liu","Haomiao Jiang","Snehasish Mukherjee","Kyle Spence","Mark Stauber","Evangelos Kalogerakis","Yunze Zeng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"BLARM: Animating 3D Objects from Video via Blending LAtent Rigid Motion Primitives","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.31113","noteUrl":"/review?paper=review-neurips-2026-155225","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155226","title":"How to Train Your Latent Diffusion Language Model Jointly With the Latent Space","authors":["Viacheslav Meshchaninov","Alexander Shabalin","Egor Chimbulatov","Nikita Gushchin","Ilya Koziev","Aleksandr Korotin","Dmitry Vetrov"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Learned denoising losses need not yield nonnegative timestep masses","plainExplanation":"Use \\(\\bar\\alpha(u)=1-u\\), unit-variance Rademacher clean data and independent Gaussian noise. The time-conditioned predictor \\(\\sqrt{1-u}\\,z_u+\\sin(\\pi u)\\) has loss \\(L(u)=u+\\sin^2(\\pi u)\\). Its losses at \\(0,1/2,1\\) are \\(0,1.5,1\\), giving normalized two-bin masses \\(1.5,-0.5\\). A zero predictor instead gives constant loss and a zero normalizing denominator. Optimal Bayes-risk monotonicity does not order errors of an arbitrary learned predictor. The adaptive sampling specification needs a guard. The ELBO and generation results are not counterexamples here.","primaryUrl":"https://arxiv.org/html/2605.07933","noteUrl":"/review?paper=review-neurips-2026-155226","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155230","title":"Video-FLAIR: Not Whether to Reason, But How","authors":["Yogesh Kulkarni","Pooyan Fazli"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Video-FLAIR: Not Whether to Reason, But How","plainExplanation":"Positive component scaling cancels in normalization apart from its numerical stabilizer.","primaryUrl":"https://arxiv.org/html/2608.26495","noteUrl":"/review?paper=review-neurips-2026-155230","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155231","title":"MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation","authors":["Yawen Shao","Jie Xiao","Kai Zhu","Yu Liu","Hongchen Luo","Xueyang Fu","Yang Cao","Wei Zhai","Zheng-Jun Zha"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.17009","noteUrl":"/review?paper=review-neurips-2026-155231","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155234","title":"FactorizedHMR: A Hybrid Framework for Video Human Mesh Recovery","authors":["Patrick Kwon","Chen Chen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FactorizedHMR: A Hybrid Framework for Video Human Mesh Recovery","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14854","noteUrl":"/review?paper=review-neurips-2026-155234","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155236","title":"Does Your Neural Network Extrapolate? Feature Engineering as Identifiability Bias for OOD Generalization","authors":["Leonel Aguilar","Jan Nagler","Christoph Hoelscher","Nino Antulov-Fantulin"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A local growth condition cannot be used beyond its stated neighborhood","plainExplanation":"A local growth condition cannot be used beyond its stated neighborhood.","primaryUrl":"https://arxiv.org/html/2605.07483","noteUrl":"/review?paper=review-neurips-2026-155236","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155237","title":"Parameterized Complexity of Stationarity Testing for Piecewise-Affine Functions and Shallow CNN Losses","authors":["Yuhan Ye"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Parameterized Complexity of Stationarity Testing for Piecewise-Affine Functions and Shallow CNN Losses","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10219","noteUrl":"/review?paper=review-neurips-2026-155237","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155238","title":"Focusing Influence Mechanism for Multi-Agent Reinforcement Learning","authors":["Yisak Park","Sunwoo Lee","Seungyul Han"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Fixed-device parallelism does not make growing counterfactual work constant-time","plainExplanation":"Appendix G.4 claims constant runtime in agent count \\(n\\), action count \\(A\\) and state dimension \\(D\\) because the counterfactual evaluations run on a GPU. GPU batching does not eliminate input-dependent scalar work on a fixed device; constant asymptotic runtime requires an explicit scaling-resource model.","primaryUrl":"https://arxiv.org/html/2506.19417","noteUrl":"/review?paper=review-neurips-2026-155238","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155243","title":"Codebook-Guided Cross-Modal Knowledge Distillation for Structurally Heterogeneous Features","authors":["Dae Ung Jo","Jongin Lim","YoungJoon Yoo","Daeho Um"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Codebook-Guided Cross-Modal Knowledge Distillation for Structurally Heterogeneous Features","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.37243","noteUrl":"/review?paper=review-neurips-2026-155243","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155248","title":"Muon Does Not Converge on Convex Lipschitz Functions","authors":["Tetiana Parshakova","Ahmed Khaled","Michael Crawshaw","Guillaume Garrigos","Robert Gower"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Muon Does Not Converge on Convex Lipschitz Functions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08980","noteUrl":"/review?paper=review-neurips-2026-155248","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155251","title":"Embeddings for Preferences, Not Semantics","authors":["Carter Blair","Ariel Procaccia","Milind Tambe"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Embeddings for Preferences, Not Semantics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08360","noteUrl":"/review?paper=review-neurips-2026-155251","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155254","title":"From Experts to Sub-experts: Fine-grained Parameter-Efficient Fine-Tuning for MoE LLMs","authors":["Zhentao Tan","Chang Liu","Yao Liu","Yue Wu","Jieping Ye"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The one-group entropy exponent contains an undefined ratio","plainExplanation":"For one positive-energy group, its normalized weight is \\(1\\), entropy is \\(-\\log1=0\\), and maximum entropy is \\(\\log1=0\\). The exponent therefore contains \\(0/0\\), even though the base scale in equation (6) is \\(1\\). A numerical convention such as \\(1^{\\mathrm{NaN}}=1\\) does not define the printed exponent. This is a removable boundary-definition gap, not an observed training failure.","primaryUrl":"https://arxiv.org/html/2609.25655","noteUrl":"/review?paper=review-neurips-2026-155254","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155257","title":"Logit-Gap Steering: A Forward-Pass Diagnostic for Alignment Robustness","authors":["Tung-Ling Li","Hongliang Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Logit-Gap Steering: A Forward-Pass Diagnostic for Alignment Robustness","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2506.24056","noteUrl":"/review?paper=review-neurips-2026-155257","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155259","title":"SPERA: Spherical Prior EEG Foundation Model with Geometry- and Frequency-Aware Latent Prediction","authors":["Minsu Kim","Ye-Sung Kim","Hyeseong Jeon","Wooseok Hyung","Joshua Lee","Chang-Hwan Im"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SPERA: Spherical Prior EEG Foundation Model with Geometry- and Frequency-Aware Latent Prediction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10571","noteUrl":"/review?paper=review-neurips-2026-155259","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155267","title":"Massively Parallel Exact Inference for Hawkes Processes","authors":["Ahmer Raza","Hudson Smith"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Interaction orientation and the constant-memory algorithm need precise corrections","plainExplanation":"Keeping those $N$ scalar intensities costs $\\Omega(N)$ auxiliary storage, even if the much larger prefix-scan buffers are released batch by batch. This is a defect in the stated constant-memory algorithm, rather than evidence about measured GPU memory.","primaryUrl":"https://arxiv.org/html/2604.01342","noteUrl":"/review?paper=review-neurips-2026-155267","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","complexity_error","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155268","title":"Mimicking the Physicist's Eye: A VLM-centric Approach for Physics Formula Discovery","authors":["Jiaqi Liu","Songning Lai","Pengze Li","Di Yu","Zhou wenjie","Yiyang Zhou","Peng Xia","Zijun Wang","Xi Chen","SHIXIANG TANG","LEI BAI","Wanli Ouyang","Mingyu Ding","Huaxiu Yao","Aoran Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mimicking the Physicist's Eye: A VLM-centric Approach for Physics Formula Discovery","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2508.17380","noteUrl":"/review?paper=review-neurips-2026-155268","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155269","title":"Sparse Video Generation Propels Real-World Beyond-the-View Vision-Language Navigation","authors":["Hai Zhang","Siqi Liang","Li Chen","Yuxian LI","Yukuan XU","Yichao Zhong","Fu Zhang","Hongyang Li"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The action corruption puts signed actions inside a square root","plainExplanation":"At an interior noise parameter \\(\\beta=1/2\\), any negative action coordinate makes \\(\\beta a_0<0\\), so \\(\\sqrt{\\beta a_0}\\) is not real. Even for \\(a_0=1/4\\), the clean endpoint \\(\\beta=1\\) returns \\(1/2\\), rather than \\(a_0\\). The labels are not squared or nonnegative encodings. This is a printed corruption-formula error. The robot implementation and experiments are not independently refuted.","primaryUrl":"https://arxiv.org/html/2602.05827","noteUrl":"/review?paper=review-neurips-2026-155269","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155270","title":"AtomWorld-Mem: Memory-Restored World States for Long-Horizon Atomistic Evolution","authors":["Tian Luo","Ruge Zhang","Haozhi Han","Yifeng Chen","Yunquan Zhang","Ting Cao","Yunxin Liu","Kun Li"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"AtomWorld-Mem: Memory-Restored World States for Long-Horizon Atomistic Evolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31133v2","noteUrl":"/review?paper=review-neurips-2026-155270","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155271","title":"Short-Context Dominance: How Much Local Context Natural Language Actually Needs?","authors":["Vala Vakilian","Zimeng Wang","Ankit Rawat","Christos Thrampoulidis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Short-Context Dominance: How Much Local Context Natural Language Actually Needs?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2512.08082","noteUrl":"/review?paper=review-neurips-2026-155271","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155273","title":"Reinforcement Learning from Rich Feedback with Distributional DAgger","authors":["Rishabh Agrawal","Jacob Fein-Ashley","Paria Rashidinejad"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Visited-state updates do not imply an all-state per-iteration contraction","plainExplanation":"A normalized score direction has zero mean, unlike a nontrivial log ratio. The future-token average divides by one fewer term than it sums. Visited-state updates do not imply an all-state per-iteration contraction.","primaryUrl":"https://arxiv.org/html/2606.05152","noteUrl":"/review?paper=review-neurips-2026-155273","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155274","title":"Valid Best-Model Identification for LLM Evaluation via Low-Rank Factorization","authors":["Elad Tolochinsky","Yaniv Tenzer","Yaniv Romano"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The variance-minimizing coefficient must respect its interval constraint","plainExplanation":"The variance-minimizing coefficient must respect its interval constraint.","primaryUrl":"https://arxiv.org/html/2605.10405","noteUrl":"/review?paper=review-neurips-2026-155274","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155280","title":"Crafting Reversible SFT Behaviors in Large Language Models","authors":["Yuping Lin","Pengfei He","Yue XING","Yingqian Cui","Jiayuan Ding","Subhabrata Mukherjee","Hui Liu","Zhen Xiang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Crafting Reversible SFT Behaviors in Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06632","noteUrl":"/review?paper=review-neurips-2026-155280","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155281","title":"Epistemic Pairwise Maximin Share","authors":["Michal Feldman","Amos Fiat","Yael Nissan","Tomasz Ponitka"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Epistemic Pairwise Maximin Share","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.18921","noteUrl":"/review?paper=review-neurips-2026-155281","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155282","title":"From Baselines to Transport Geodesics: Axiomatic Attribution via Optimal Generative Flows","authors":["Cenwei Zhang","Lin Zhu","Manxi Lin","Lei You"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"From Baselines to Transport Geodesics: Axiomatic Attribution via Optimal Generative Flows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.05093","noteUrl":"/review?paper=review-neurips-2026-155282","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155283","title":"Statistical Query Lower Bounds for Smoothed Agnostic Learning","authors":["Ilias Diakonikolas","Daniel Kane"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Large smoothing makes the sign function close to a constant","plainExplanation":"The Gaussian reverse-norm constant is too small. The classification-risk conversion loses a factor of two. Large smoothing makes the sign function close to a constant. The projection identity has the wrong dimension. The conditioning substitution divides where it must multiply. The displayed SQ exponent contains an extra reciprocal. The small-smoothing expansion has the reciprocal power reversed.","primaryUrl":"https://arxiv.org/html/2602.21191","noteUrl":"/review?paper=review-neurips-2026-155283","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","lemma_error","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155285","title":"When Reasoning Meets Its Laws","authors":["Junyu Zhang","Yifan Sun","Tianang Leng","Jingyan Shen","Liu Ziyin","Paul Liang","Huan Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Sublinear composition error can accumulate into an unbounded slope","plainExplanation":"Sublinear composition error can accumulate into an unbounded slope.","primaryUrl":"https://arxiv.org/html/2512.17901","noteUrl":"/review?paper=review-neurips-2026-155285","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-155286","title":"No Coin Left Behind: Maximizing Strategic Surplus Against No-Regret Dynamics","authors":["Yiheng Su","Emmanouil-Vasileios Vlatakis-Gkaragkounis"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The odd-horizon proof does not cover the first round","plainExplanation":"Theorem 4.4 takes an unrestricted infimum over step size and time after a constrained supremum over games and strategies meeting terminal accuracy. The proof inserts a fixed alternating witness into constrained suprema where that witness may fail the accuracy constraint, then takes an unrestricted outer infimum. The fixed-witness quantifier step is unsupported. The bounded-price observation applies only to a bounded payoff class; the bare unrestricted game display does not justify extending it to unbounded payoffs.","primaryUrl":"https://arxiv.org/html/2604.05129","noteUrl":"/review?paper=review-neurips-2026-155286","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155289","title":"Posterior Sampling-based Online Learning for Episodic POMDPs","authors":["Dengwang Tang","Dongze Ye","Rahul Jain","Ashutosh Nayyar","Pierluigi Nuzzo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Posterior Sampling-based Online Learning for Episodic POMDPs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2310.10107","noteUrl":"/review?paper=review-neurips-2026-155289","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155291","title":"Harnessing Textual Refusal Directions for Multimodal Safety","authors":["Moreno D&amp;#x27;Incà","Nicu Sebe","Massimiliano Mancini"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Harnessing Textual Refusal Directions for Multimodal Safety","plainExplanation":"No issue was found in the checked claims. The unsafe-class reversal in the historical formula is removed.","primaryUrl":"https://arxiv.org/html/2606.31876","noteUrl":"/review?paper=review-neurips-2026-155291","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155297","title":"Virtual Head Attention","authors":["Xibo Ding","Guoxia Wang","Jinle Zeng","Jiabin Yang","Sirui Tao","Chao Yang","Dianhai Yu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Virtual Head Attention","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.17959","noteUrl":"/review?paper=review-neurips-2026-155297","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155298","title":"Tikhonov-Stabilized Bezier Representation Forecasting for Training-free Diffusion Acceleration","authors":["Lei Zhu","Mujie Lin","Ruochong Zheng","Guangyi Wang","Li Hao","Peng Jin","Chang Liu","Jie Chen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Tikhonov-Stabilized Bezier Representation Forecasting for Training-free Diffusion Acceleration","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16448","noteUrl":"/review?paper=review-neurips-2026-155298","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155300","title":"How to Scale Mixture-of-Experts: From muP to the Maximally Scale-Stable Parameterization","authors":["Leena Chennuru Vankadara","Moritz Haas","Luke Hayward","Sebastian Bordt","Alessandro Breccia"],"year":2026,"category":"Machine learning","status":"unresolved","issueTitle":"The router-kernel constraints need a local sign and sum correction","plainExplanation":"The plus sign imposes the negative of the defined kernel, and placing \\(G^\\phi\\) inside the expert sum repeats it \\(M\\) times. The two time arguments also need to agree with the defined cross-time kernel.","primaryUrl":"https://arxiv.org/html/2605.14200","noteUrl":"/review?paper=review-neurips-2026-155300","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-155304","title":"Nearly Optimal Fixed-Confidence Best-Arm Identification with 1-Bit Feedback","authors":["Khang Luong","Thai Son Dinh","Hoang Ta","Hung TRAN","Tuan Dam"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Nearly Optimal Fixed-Confidence Best-Arm Identification with 1-Bit Feedback","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.02771","noteUrl":"/review?paper=review-neurips-2026-155304","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155306","title":"Learning Orthonormal Bases for Function Spaces","authors":["Hamidreza Kamkari","Mohammad S Nabizadeh","Justin Solomon"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Hilbert–Schmidt rotations need not contain an eigenbasis","plainExplanation":"Every admissible rotation satisfies \\(Q-I\\) Hilbert–Schmidt. Let \\(A=V\\operatorname{diag}(2^{-i})V^*\\), where \\(V\\) rotates every consecutive basis pair by \\(\\pi/4\\). This is positive trace-class with simple spectrum. Every coordinate of every eigenvector has magnitude at most \\(1/\\sqrt2\\), so any signed and permuted eigenbasis satisfies \\(\\|(Q-I)e_i\\|^2\\geq2-\\sqrt2\\) for every \\(i\\), contradicting the Hilbert–Schmidt requirement.\n\nFinite-prefix rotations are admissible and approach objective \\(\\sum_i4^{-i}=1/3\\). An attained optimum would have to diagonalize \\(A\\): admissible planar variations eliminate off-diagonal entries and order the diagonal entries. Hence this example has a supremum without an admissible maximizer. The unconditional existence and diagonalization claims fail. A characterization explicitly conditional on an existing maximizer is not contradicted.","primaryUrl":"https://arxiv.org/html/2605.19959","noteUrl":"/review?paper=review-neurips-2026-155306","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155314","title":"DiffeoMorph: Learning to Morph 3D Shapes Using Differentiable Agent-Based Simulations","authors":["Seong Ho Pahng","Guoye Guan","Benjamin Fefferman","Sahand Hormoz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DiffeoMorph: Learning to Morph 3D Shapes Using Differentiable Agent-Based Simulations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.17129","noteUrl":"/review?paper=review-neurips-2026-155314","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155316","title":"Sparsely-Supervised Data Assimilation via Physics-Informed Schrödinger Bridge","authors":["Dohyun Bu","Chanho Kim","Seokun Choi","Jong-Seok Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Burgers denominator is not a valid inverse-residual bound","plainExplanation":"Noise added after projection leaves the observation constraint. The Burgers denominator is not a valid inverse-residual bound. The localization proof neither closes the exterior nor separates equality cases. Unsquared drift regression elicits a median rather than the required mean.","primaryUrl":"https://arxiv.org/html/2603.22319","noteUrl":"/review?paper=review-neurips-2026-155316","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155321","title":"Measuring Safety Alignment Effects in Autonomous Security Agents","authors":["Isaac David","Arthur Gervais"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Measuring Safety Alignment Effects in Autonomous Security Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.19722","noteUrl":"/review?paper=review-neurips-2026-155321","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155323","title":"Understanding Sample Efficiency in Predictive Coding","authors":["Gaspard Oliviers","Elene Lominadze","Rafal Bogacz"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The decorrelation inverse has its factors transposed in the wrong order","plainExplanation":"The decorrelation inverse has its factors transposed in the wrong order. The equilibrium formula uses the wrong layer-product indices. The alignment rescaling can divide by a zero activity product.","primaryUrl":"https://arxiv.org/html/2605.11911","noteUrl":"/review?paper=review-neurips-2026-155323","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155324","title":"Geometric Latent Reasoning Induces Shorter Generations in LLMs","authors":["Shashi Kumar","Yacouba Kaloga","Petr Motlicek","Ina Kodrasi","Andrea Cavallaro"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Masking latent-token losses does not isolate the transition head","plainExplanation":"If answer CE is computed on the modified second-pass sequence, the chain rule gives \\(\\nabla_\\phi L_{\\mathrm{answer}}=(\\partial L_{\\mathrm{answer}}/\\partial\\hat e)(\\partial g_\\phi/\\partial\\phi)\\), which need not vanish. A scalar replacement \\(\\hat e=0.2+1.3\\phi\\) feeding a positive-target logistic answer loss already has nonzero derivative. Masking direct token losses does not block downstream answer gradients. Only the claimed sufficiency of masking is false. The main-text first-pass wording leaves actual routing ambiguous; freezing the embedding table does not resolve that ambiguity.","primaryUrl":"https://arxiv.org/html/2606.02248","noteUrl":"/review?paper=review-neurips-2026-155324","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155325","title":"Tackling the Data-Parallel Load Balancing Bottleneck in LLM Serving: Practical Online Routing at Scale","authors":["Tianci Bu","Yuan Lyu","Zixi Chen","Chendong Song","Hong Liang","Tsepten Gurung","Yuwei Fan","Yinyu Ye","Zijie Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Free request slots do not identify a safe workload regime","plainExplanation":"Section 3.2 infers that a worker with the most free request slots is lightest-loaded and can accept the largest waiting request with score equal to its size. Free slots count requests rather than workload, so the most free-capacity worker need not be lightest-loaded or in the safe regime for the largest request. The shortcut’s justification fails. The full-score rule in the appendix and algorithm can avoid this particular inference.","primaryUrl":"https://arxiv.org/html/2605.06113","noteUrl":"/review?paper=review-neurips-2026-155325","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155327","title":"How Long Does Infinite Width Last? Signal Propagation in Long-Range Linear Recurrences","authors":["Mariia Seleznova"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"How Long Does Infinite Width Last? Signal Propagation in Long-Range Linear Recurrences","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05113","noteUrl":"/review?paper=review-neurips-2026-155327","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155328","title":"SCOT: Multi-Source Cross-City Transfer with Optimal-Transport Soft-Correspondence Objectives","authors":["Yuyao Wang","Min Yang","Meng Chen","Weiming Huang","Yilong Yin","Yongshun Gong"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The contrastive-loss rewrite changes its weights and normalization","plainExplanation":"The contrastive-loss rewrite changes its weights and normalization. All-one marginals are incompatible for rectangular transport.","primaryUrl":"https://arxiv.org/html/2604.07383","noteUrl":"/review?paper=review-neurips-2026-155328","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155330","title":"Words That Make Language Models Perceive","authors":["Sophie L. Wang","Phillip Isola","Brian Cheung"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Words That Make Language Models Perceive","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.02425","noteUrl":"/review?paper=review-neurips-2026-155330","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155333","title":"Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models","authors":["Xiangyu Zhou","Saleh Zare Zade","Rafi Ibn Sultan","Alexander Kotov","Dongxiao Zhu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.36254","noteUrl":"/review?paper=review-neurips-2026-155333","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155334","title":"Agentic AI-Empowered Dynamic Survey Framework","authors":["Furkan Mumcu","Lokman Bekit","Michael Jones","Anoop Cherian","Yasin Yilmaz"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Agentic AI-Empowered Dynamic Survey Framework","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.04071","noteUrl":"/review?paper=review-neurips-2026-155334","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155335","title":"Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language","authors":["Max Bennett","Thomas Zollo","Richard Zemel"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.23201","noteUrl":"/review?paper=review-neurips-2026-155335","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155336","title":"Do We Need Asynchronous SGD? On the Near-Optimality of Synchronous Methods","authors":["Grigory Begunov","Alexander Tyurin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Worker-count near-optimality and timing independence need precise statements","plainExplanation":"The proposed worker count can fail to minimize runtime exactly, while a factor-four guarantee follows under the stated timing scale condition. The random-time lower-bound proof needs independent resampling across attempts, beyond the marginal and gradient-independence conditions in Assumption 3.1.","primaryUrl":"https://arxiv.org/html/2602.03802","noteUrl":"/review?paper=review-neurips-2026-155336","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","invalid_inference","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155338","title":"BEAGLE: Behavior-Enforced Agent for Grounded Learner Emulation","authors":["David Wang","Clayton Cohn","Zifan Xu","Siyuan Guo","Gautam Biswas","Meiyi Ma"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Pooled human judgments do not determine uncertainty under participant dependence","plainExplanation":"Appendix H.2 uses a pooled standard error for a two-one-sided equivalence test, although each of 71 participants judges six real and six generated traces. Pooled totals do not identify within-participant covariance, so the independent-trial standard error does not establish population equivalence without a sampling-model justification.","primaryUrl":"https://arxiv.org/html/2602.13280","noteUrl":"/review?paper=review-neurips-2026-155338","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155341","title":"It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs","authors":["Pierre-Carl Langlais","Pieter Delobelle","Yannick Detrois","Pavel Chizhov","Carlos Rosas-Hinostroza","Neil S Smail","Benjamin Burtin","Hanna Shcharbakova","Ivan Yamshchikov","Anastasia Stasenko"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.37891","noteUrl":"/review?paper=review-neurips-2026-155341","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155342","title":"PISCO: Precise Video Instance Insertion with Sparse Control","authors":["Xiangbo Gao","Renjie Li","Xinghao Chen","Yuheng Wu","Suofei Feng","Jie Yang","Qing Yin","Zhengzhong Tu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PISCO: Precise Video Instance Insertion with Sparse Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.08277","noteUrl":"/review?paper=review-neurips-2026-155342","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155346","title":"LLM-ACES: Closed-Loop Discovery of Dynamic Systems with LLM-Guided Adaptive Search","authors":["Nikhil Abhyankar","Sha Li","Sanchit Kabra","Naren Ramakrishnan","Yulia Gel","Chandan Reddy"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LLM-ACES: Closed-Loop Discovery of Dynamic Systems with LLM-Guided Adaptive Search","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.25039","noteUrl":"/review?paper=review-neurips-2026-155346","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155348","title":"Making LoRA Identifiable: Orthogonal Alignment for Continual Learning","authors":["Pratik Rakesh Singh","Mohammadi Zaki","Akash Saha","Aneesh Mukkamala","Pankaj Wasnik"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Making LoRA Identifiable: Orthogonal Alignment for Continual Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16664","noteUrl":"/review?paper=review-neurips-2026-155348","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155350","title":"Bridging Compute- and Data-Optimal Pretraining","authors":["Tian Qin","Kimia Hamidieh","David Alvarez-Melis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A coordinate gradient path is not a resource-allocation optimum","plainExplanation":"With \\(D_{\\mathrm{eff}}=D+DR(1-e^{-D^{\\prime}/(DR)})\\), its derivative with respect to derived data is \\(e^{-D^{\\prime}/(DR)}>0\\) at finite inputs. Near-saturation is not zero marginal benefit. More fundamentally, replacing a coordinate \\(x\\) by \\(1000x\\) changes the physical gradient-path slope by a factor \\(10^6\\). Without a resource cost or target-loss constraint, this coordinate-dependent path does not identify an optimal allocation. The optimal-allocation interpretation is unsupported. Fitted losses, near-saturation and useful empirical heuristics can remain valid.","primaryUrl":"https://arxiv.org/html/2607.25271","noteUrl":"/review?paper=review-neurips-2026-155350","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155352","title":"D-VLA: A High-Concurrency Distributed Asynchronous RL Framework for Vision-Language-Action Models","authors":["Guo Yucheng","Yongjian Guo","Zhong Guan","Wen Huang","Haodong Yue","shuai di","Junwu Xiong","Yicheng Gong"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"D-VLA: A High-Concurrency Distributed Asynchronous RL Framework for Vision-Language-Action Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13276v2","noteUrl":"/review?paper=review-neurips-2026-155352","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155353","title":"Latent Abstraction for Retrieval-Augmented Generation","authors":["Thi Ha Lan Nguyen","Nguyen Minh-Anh","Dung Le"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Latent Abstraction for Retrieval-Augmented Generation","plainExplanation":"The label can say stop because of a document obtained only after the stop decision. It therefore does not express evidence sufficiency in the input context.","primaryUrl":"https://arxiv.org/html/2604.17866","noteUrl":"/review?paper=review-neurips-2026-155353","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155354","title":"PhysFormer: Learning to Simulate Mechanics in World Space","authors":["Yiming Chen","Yushi LAN","Andrea Vedaldi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PhysFormer: Learning to Simulate Mechanics in World Space","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.27364","noteUrl":"/review?paper=review-neurips-2026-155354","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155357","title":"NEAT: Neighborhood-Guided, Efficient, Autoregressive Set Transformer for 3D Molecular Generation","authors":["Daniel Rose","Roxane Axel Jacob","Johannes Kirchmair","Thierry Langer"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"The velocity-to-score identity omits source variance","plainExplanation":"The velocity-to-score identity omits source variance.","primaryUrl":"https://arxiv.org/html/2512.05844","noteUrl":"/review?paper=review-neurips-2026-155357","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155358","title":"Actionable Hallucination Detection: Translating Latent Uncertainty into Agentic Critique","authors":["Sanidhya Vijayvargiya","Rahul Lokesh"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Actionable Hallucination Detection: Translating Latent Uncertainty into Agentic Critique","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.10430","noteUrl":"/review?paper=review-neurips-2026-155358","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155361","title":"Every Sample Counts: Supervised Fine-Tuning of Language Models with Pointwise Constraints","authors":["Ignacio Hounie","Ignacio Boero","Alejandro Ribeiro"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The relaxed dual supergradient divides one term by an extra factor","plainExplanation":"Slack elimination gives \\(g_R(\\lambda)=g_A(\\beta\\lambda/(\\alpha+\\beta),\\alpha\\beta/(\\alpha+\\beta))-\\lambda^2/[4(\\alpha+\\beta)]\\). The active-coordinate derivative is \\(\\beta v/(\\alpha+\\beta)-\\lambda/[2(\\alpha+\\beta)]\\). The printed second term contains an additional division by \\(\\beta\\). A singleton model removes minimizer ambiguity; rescaling the whole multiplier or step cannot change only this term. This is a local supergradient error for \\(\\beta\\ne1\\); no implementation or empirical failure follows.","primaryUrl":"https://arxiv.org/html/2607.08968","noteUrl":"/review?paper=review-neurips-2026-155361","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155362","title":"Masked Diffusion Language Agents for Tool-Integrated Chemical Reasoning","authors":["Mengdi Liu","Chenghao Jia","Hong Chang","Shiguang Shan"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Masked Diffusion Language Agents for Tool-Integrated Chemical Reasoning","plainExplanation":"No issue was found in the checked claims. The two objectives differ, but the derivation explicitly changes the masking law. Its support statement concerns block sets \\(A\\subseteq[K]\\), not every subset of individual tokens. The stronger equality and token-support assertions attributed to it are not established as source claims.","primaryUrl":"https://arxiv.org/html/2610.13330","noteUrl":"/review?paper=review-neurips-2026-155362","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155367","title":"TransmissiveGS: Residual-Guided Disentangled Gaussian Splatting for Transmissive Scene Reconstruction and Rendering","authors":["Zhenyu Liang","Xiao Zhang","Tianchao Li","Jack C Cheng","Chi-Keung Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A directional environment dictionary need not identify geometry","plainExplanation":"A camera ray \\((0,0,-1)\\) reflects to \\((0,0,1)\\) for normal \\((0,0,1)\\), but to \\((1,0,0)\\) for normal \\((1,0,1)/\\sqrt2\\). The allowed constant environment map \\(E=0.8\\) returns the same reflected color \\(0.4\\) with \\(f_0=0.5\\) for both directions. Taking reflectance \\(0.5\\), scattering \\(0.2\\) and target \\(0.3\\) leaves the same nonzero residual \\(0.1\\) under both normals. A direction-indexed map is not necessarily injective. This disproves sufficiency of the cited Stage II argument, not the full regularized pipeline or its benchmark quality.","primaryUrl":"https://arxiv.org/html/2605.10705","noteUrl":"/review?paper=review-neurips-2026-155367","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155369","title":"Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal Editing","authors":["Bryce Hinkley","Peyman Najafirad"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal Editing","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.20262","noteUrl":"/review?paper=review-neurips-2026-155369","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155370","title":"PAIR-CI: Calibrated Conditional Independence Testing for Causal Discovery with Incomplete Data","authors":["Thomas S. Robinson","Ranjit Lall"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"PAIR-CI: Calibrated Conditional Independence Testing for Causal Discovery with Incomplete Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.04838","noteUrl":"/review?paper=review-neurips-2026-155370","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155371","title":"Attention Alignment Between Humans and Vision-Language Models","authors":["Isaac Christian","Udith Haputhanthri","Declan Campbell","Samuel Nastase","Taylor Webb","Michael S Graziano"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Mean human–human correlation is not a universal model ceiling","plainExplanation":"Take the stated finite-width Gaussian fixation maps centered at horizontal positions \\(120.5\\) and \\(136.5\\), with a fixed model map centered halfway at \\(128.5\\). On the stated image grid, model–human Spearman correlation is about \\(0.97024\\), while cross-human correlation is about \\(0.92337\\). With \\(13\\) and \\(14\\) observers of the two types, including all distinct same-type pairs gives mean human–human correlation about \\(0.96026\\), still below the fixed model’s mean alignment. The universal upper-bound interpretation fails; empirical rankings, normalization values and the actual subjects’ attainable ceiling are not determined by this construction.","primaryUrl":"https://arxiv.org/html/2606.17410","noteUrl":"/review?paper=review-neurips-2026-155371","collection":"neurips-2026","severity":"repairable","issueTypes":["interpretation_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155372","title":"Continually Evolving Skill Knowledge in Vision Language Action Model","authors":["Yuxuan Wu","Guangming Wang","Zhiheng Yang","Tianchen Deng","Maoqing Yao","Brian Sheil","Hesheng Wang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Continually Evolving Skill Knowledge in Vision Language Action Model","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2511.18085","noteUrl":"/review?paper=review-neurips-2026-155372","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155376","title":"Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization","authors":["Guangyu Yang","Jingbiao Mei","Mingsheng Sun","Jinghong Chen","Yingtong Bu","Pengda Qin","Da Chen","Bill Byrne"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A correctly predicted safe video makes the Tversky reward undefined","plainExplanation":"For a safe video, the ground-truth unsafe set is empty. A correct empty prediction has \\(TP=FP=FN=0\\), so the displayed quotient is \\(0/0\\) for all positive Tversky parameters. The evaluated datasets explicitly include safe videos. A separate format reward cannot define an undefined overlap quotient. The manuscript lacks a boundary convention. A library or code guard may already handle it; no runtime failure was reproduced.","primaryUrl":"https://arxiv.org/html/2610.02019","noteUrl":"/review?paper=review-neurips-2026-155376","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155378","title":"The Unreasonable Effectiveness of Text Embedding Interpolation for Continuous Image Steering","authors":["Yiğit Ekin","Yossi Gandelsman"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Absolute semantic increments discard direction reversals","plainExplanation":"Absolute semantic increments discard direction reversals.","primaryUrl":"https://arxiv.org/html/2603.17998","noteUrl":"/review?paper=review-neurips-2026-155378","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155380","title":"Precision As You Need: Stochastic Computing Is a Dense Adaptive Quantizer","authors":["Haoran Jin","Kangqi Zhang","Jirong Yang","Barry Lyu","Qiuyi Ding","Ruijie Gao","Nathan Bleier"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Probability independence does not ensure exact finite-stream multiplication","plainExplanation":"Probability independence does not ensure exact finite-stream multiplication. Stream-length totals omit operator-dependent gate multiplicities.","primaryUrl":"https://arxiv.org/html/2609.32922","noteUrl":"/review?paper=review-neurips-2026-155380","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155381","title":"Neural Signals Generate Clinical Notes in the Wild","authors":["Jathurshan Pradeepkumar","Zheng Chen","Jimeng Sun"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Neural Signals Generate Clinical Notes in the Wild","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.22197","noteUrl":"/review?paper=review-neurips-2026-155381","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155382","title":"XTC: Head-Aware Sampling by Excluding Top Choices","authors":["Philipp E Weidmann","Allen Roush","Judah Goldfeder","Sanjay Basu","Ravid Shwartz-Ziv"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"XTC: Head-Aware Sampling by Excluding Top Choices","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.22758","noteUrl":"/review?paper=review-neurips-2026-155382","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155387","title":"Hypothesis generation and updating in large language models","authors":["Hua-Dong Xiong"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Hypothesis generation and updating in large language models","plainExplanation":"Domain expansion changes compatible hypotheses’ likelihood ratios and can give interval hypotheses positive extension mass.","primaryUrl":"https://arxiv.org/html/2605.05851","noteUrl":"/review?paper=review-neurips-2026-155387","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155388","title":"Generative Actor-Critic with Soft Bridge Policies","authors":["Ke He","Le He","Shunpu Tang","Yafei Wang","Lisheng Fan"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Generative Actor-Critic with Soft Bridge Policies","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08733","noteUrl":"/review?paper=review-neurips-2026-155388","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155392","title":"Agentic Reward Modeling: Verifying GUI Agent via Progressive Trajectory-Grounded Interaction","authors":["CHAOQUN CUI","Jing Huang","SHIJING WANG","Liming Zheng","Qingchao Kong","Zhixiong Zeng"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Agentic Reward Modeling: Verifying GUI Agent via Progressive Trajectory-Grounded Interaction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.00575","noteUrl":"/review?paper=review-neurips-2026-155392","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155393","title":"Inner Product Aware Quantization: Provably Fast, Accurate, and Adaptive Algorithms","authors":["Nathan White","Krish Singal"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The candidate grid omits each cluster’s translation","plainExplanation":"The required quadrangle inequality is \\(C(a,c)+C(b,d)\\leq C(a,d)+C(b,c)\\). An admissible four-entry instance has values \\(1,1,4,0\\). For grid width \\(\\delta>4\\), ceiling-rounding gives \\(\\delta,\\delta,\\delta,0\\), requiring \\(2\\delta\\leq\\delta\\). The receipt checks that a legitimate warm-start choice can produce such a grid. The claimed fast rounded solver lacks a required hypothesis. A generic dynamic program remains possible but does not establish the stated linear-time primitive.","primaryUrl":"https://arxiv.org/html/2606.00289","noteUrl":"/review?paper=review-neurips-2026-155393","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155394","title":"Read It Back: Pretrained MLLMs Are Zero-shot Reward Models for Text-to-Image Generation","authors":["Runhui Huang","Qihui Zhang","Zhe Liu","Yu Gao","Jie Wu","Hengshuang Zhao"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Read It Back: Pretrained MLLMs Are Zero-shot Reward Models for Text-to-Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.11886","noteUrl":"/review?paper=review-neurips-2026-155394","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155395","title":"Time series analysis with Gumbel dynamics","authors":["Yiliu Wang","Timothy Kim","Eric Todd SheaBrown","Uygar Sümbül"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Time series analysis with Gumbel dynamics","plainExplanation":"No issue was found in the checked claims. The current formula contains the quadratic weights and cross terms missing from the retained manuscript. The historical soft-state covariance objection is resolved.","primaryUrl":"https://arxiv.org/html/2509.21578","noteUrl":"/review?paper=review-neurips-2026-155395","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155396","title":"Articraft: An Agentic System for Scalable Articulated 3D Asset Generation","authors":["Matt Zhou","Ruining Li","Xiaoyang Lyu","Zhaomou Song","Zhening Huang","Chuanxia Zheng","Christian Rupprecht","Andrea Vedaldi","Shangzhe Wu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Articraft: An Agentic System for Scalable Articulated 3D Asset Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.15187","noteUrl":"/review?paper=review-neurips-2026-155396","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155399","title":"Finite-Sample Performance of Gradient Descent in Logistic Regression with Gaussian Design","authors":["Junren Chen","Arya Mazumdar"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Finite-Sample Performance of Gradient Descent in Logistic Regression with Gaussian Design","plainExplanation":"The first relative error is independent of sample size.","primaryUrl":"https://arxiv.org/html/2606.21683","noteUrl":"/review?paper=review-neurips-2026-155399","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-155400","title":"Toward Better Geometric Representations for Molecule Generative Models","authors":["Shaoheng Yan","Zian Li","Cai Zhou","Qiaojing Huang","Kai Liu","Muhan Zhang"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Toward Better Geometric Representations for Molecule Generative Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07693","noteUrl":"/review?paper=review-neurips-2026-155400","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155403","title":"The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation","authors":["Aaryaman Kalani","Murari Mandal","Dhruv Kumar","Mohan Kankanhalli","Yash Sinha"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.08901","noteUrl":"/review?paper=review-neurips-2026-155403","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155404","title":"AgentKVShift: Efficient KV Cache Reuse for Agentic Memory Systems","authors":["Nilesh Prasad Pandey","Jason Kong","Lanxiang Hu","Quanling Zhao","Yujie Zhao","Onat Gungor","Hao Zhang","Tajana S Rosing"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The downstream attention constant uses the wrong softmax operator norm","plainExplanation":"The downstream attention constant uses the wrong softmax operator norm.","primaryUrl":"https://arxiv.org/html/2607.21604","noteUrl":"/review?paper=review-neurips-2026-155404","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155405","title":"Minimax Rates and Spectral Distillation for Tree Ensembles","authors":["Binh Vu","David Watson"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Training eigenvectors need not optimize the test cross-kernel","plainExplanation":"The displayed variance bound does not justify replacing subsample size by total size. The signed spectral packing is not a valid bounded-response mean family. Training eigenvectors need not optimize the test cross-kernel.","primaryUrl":"https://arxiv.org/html/2605.11841","noteUrl":"/review?paper=review-neurips-2026-155405","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155407","title":"UniCustom: Unified Visual Conditioning for Multi-reference Image Generation","authors":["Yiyan Xu","Qiulin Wang","Wenjie Wang","Yunyao Mao","Xintao Wang","Pengfei Wan","Kun Gai","Fuli Feng"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"UniCustom: Unified Visual Conditioning for Multi-reference Image Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12088","noteUrl":"/review?paper=review-neurips-2026-155407","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155409","title":"Near-Linear Time Generalized Sinkhorn Algorithms for Bounded Genus Graphs","authors":["Krzysztof M Choromanski","Derek Long","Ananya Parashar","Dwaipayan Saha"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Generalized entropy stationarity does not give the claimed fixed-kernel scaling","plainExplanation":"Generalized entropy stationarity does not give the claimed fixed-kernel scaling. Tied separator values break both the expansion and the halving argument. The feature signs reconstruct the reflected function and omit complex phases.","primaryUrl":"https://arxiv.org/html/2605.09782","noteUrl":"/review?paper=review-neurips-2026-155409","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155410","title":"Data-Driven Covariate Selection for Nonparametric and Cycle-Agnostic Causal Effect Estimation","authors":["Ana L Vicente","Gijs van Seeventer","Saber Salehkaleybar"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Data-Driven Covariate Selection for Nonparametric and Cycle-Agnostic Causal Effect Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06385","noteUrl":"/review?paper=review-neurips-2026-155410","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155412","title":"BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics","authors":["Junfeng Xia","Wenhao Ye","Junxiang Zhang","Xuanye Pan","Mo WANG","Quanying Liu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.17742","noteUrl":"/review?paper=review-neurips-2026-155412","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155414","title":"M$^\\star$: Every Task Deserves Its Own Memory Harness","authors":["Wenbo Pan","Shujie LIU","Xiangyang Zhou","Xianlong Wang","Shiwei Zhang","Bingjing Xu","Wanlu Shi","Xiaohua Jia"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"M$^\\star$: Every Task Deserves Its Own Memory Harness","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.11811","noteUrl":"/review?paper=review-neurips-2026-155414","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155419","title":"DIGS: Distribution-Informed Gaussian Splatting for Training-Free Open-Vocabulary 3D Segmentation","authors":["ZHIFAN ZENG","Wenbo Xiao","Arcot Sowmya","Changming Sun"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"DIGS: Distribution-Informed Gaussian Splatting for Training-Free Open-Vocabulary 3D Segmentation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.16177","noteUrl":"/review?paper=review-neurips-2026-155419","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155421","title":"Closed-Form Last Layer Optimization","authors":["Alexandre Galashov","Nathaël Da Costa","Liyuan Xu","Philipp Hennig","Arthur Gretton"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The unregularized profile can have no nonglobal differentiable critical point","plainExplanation":"The unregularized profile can have no nonglobal differentiable critical point.","primaryUrl":"https://arxiv.org/html/2510.04606","noteUrl":"/review?paper=review-neurips-2026-155421","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155425","title":"A Scalable Measure of Loss Landscape Curvature for Analyzing the Training Dynamics of LLMs","authors":["Dayal Singh Kalra","Jean-Christophe Gagnon-Audet","Andrey Gromov","Ishita Mediratta","Kelvin Niu","Alexander Miller","Michael Shvartsman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A doubling search can skip the first loss-increasing interval","plainExplanation":"Definition 2.1 takes the infimum of positive step sizes increasing loss, while Algorithm 1 claims that doubling/halving brackets that first threshold for a generic update direction. A nonconvex loss ray can contain an early positive interval between doubling probes, so the search can bracket a later crossing instead of the defined infimum.","primaryUrl":"https://arxiv.org/html/2601.16979","noteUrl":"/review?paper=review-neurips-2026-155425","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155426","title":"NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning","authors":["Ruiyu Yan","Bowen Chen","Shaowen Wan","Lin Zhao"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.38098","noteUrl":"/review?paper=review-neurips-2026-155426","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155427","title":"Dynamic Expert Sharing: Decoupling Memory from Parallelism in Mixture-of-Experts Diffusion LLMs","authors":["Hao Chen","Zhiwen Mo","Royson Lee","Qianzhou Wang","Da Li","Shell Xu Hu","Wayne Luk","Timothy Hospedales","Hongxiang Fan"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The preliminary union can contain fewer than the required experts","plainExplanation":"If all \\(32\\) tokens have the same strict ranking among \\(16\\) experts and \\(k=2\\), the union contains exactly two experts. The subsequent \\(K=8\\) selection is impossible. The union is guaranteed to have at least \\(k\\) elements, not \\(K\\), and softmax does not alter this order. This is a missing pseudocode boundary case, not a demonstrated failure of the implementation or benchmarks.","primaryUrl":"https://arxiv.org/html/2602.00879","noteUrl":"/review?paper=review-neurips-2026-155427","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155430","title":"Equivariant Spherical Transformer for Efficient Molecular Modeling","authors":["Junyi An","Xinyu Lu","Yun-Fei Shi","Peijia Lin","Qianwei Tang","Li-Cheng Xu","Chao Qu","Fenglei Cao","Yuan Qi"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"Scalar spherical operations omit parity-incompatible tensor-product paths","plainExplanation":"The final Fibonacci point has height greater than one. Scalar spherical operations omit parity-incompatible tensor-product paths.","primaryUrl":"https://arxiv.org/html/2505.23086","noteUrl":"/review?paper=review-neurips-2026-155430","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155431","title":"Provably Reliable Classifier Guidance via Cross-Entropy Control","authors":["Sharan Sahu","Arisina Banerjee","Yuchen Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The deterministic-label exception is false; the revised score factor is justified","plainExplanation":"A smooth classifier sequence disproves the deterministic-label impossibility remark. The sampler needs a local index correction. The revised sampling bound already includes label-probability dependence for score error; retaining exact normalization completes that transfer without a stronger assumption.","primaryUrl":"https://arxiv.org/html/2601.21200","noteUrl":"/review?paper=review-neurips-2026-155431","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155433","title":"Primal-Dual Flow Matching for Sample-Wise Constrained Generation","authors":["Zhengyan Huan","Peter Y. Lu","Shuchin Aeron"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Primal-Dual Flow Matching for Sample-Wise Constrained Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.12699","noteUrl":"/review?paper=review-neurips-2026-155433","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155434","title":"Mitigating Factual Hallucination in Large Reasoning Models via Mixed-Mode Advantage Regularization","authors":["Kaishen Wang","Tong Zheng","Xuehao Cui","Ruibo Chen","Tianyi Xiong","Heng Huang"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Mitigating Factual Hallucination in Large Reasoning Models via Mixed-Mode Advantage Regularization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.05861","noteUrl":"/review?paper=review-neurips-2026-155434","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155435","title":"One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer","authors":["Jucheng Shen","Barbara Su","Anastasios Kyrillidis"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"One Model, Two Roles: Emergent Specialization in a Shared Recurrent Transformer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17811","noteUrl":"/review?paper=review-neurips-2026-155435","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155437","title":"AsyncMesh: Fully Asynchronous Optimization for Data and Pipeline Parallelism","authors":["Thalaiyasingam Ajanthan","Sameera Ramasinghe","Gil Avraham","Hadi Mohaghegh Dolatabadi","Chamin Hewa Koneputugodage","Violetta Shevchenko","Yan Zuo","Alexander Long"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Exchangeability does not justify the EMA noise cancellation","plainExplanation":"Homogeneity gives equality of marginal expectations, not conditional mean zero relative to the previous EMA. With two iid workers, unit delay and period, quadratic gradient \\(w+z\\) and Rademacher noise, the first local update is \\(\\hat w_1=-0.1z\\), its EMA is \\(d_1=0.5\\hat w_1\\), and the next update is \\(\\hat w_2=0.95d_1-0.05z_2\\). The centered noise includes \\(-0.525\\hat w_1\\), so it correlates with the previous error.\n\nThe receipt also checks a bounded-below smooth cosine objective where the omitted cross term is positive, so it cannot generally be dropped as favorable. The empirical average has positive variance even when its expectation is zero. These are centering and proof-cancellation gaps. They do not refute eventual consensus or measured speed and accuracy.","primaryUrl":"https://arxiv.org/html/2601.22442","noteUrl":"/review?paper=review-neurips-2026-155437","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155438","title":"Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View","authors":["Yixian Xu","Yuanrui Zhang","Shengjie Luo","Kaiyuan Gao","Liwei Wang","Di He"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Group standardization can reverse the shared-parameter reward gradient","plainExplanation":"Section 3.1 claims that replacing bounded reward by its group-relative standardized advantage preserves the original policy gradient in the path-space estimator. Group standardization removes prompt-specific reward scale and reweights shared-parameter gradients, so it need not preserve the original reward gradient and can reverse it.","primaryUrl":"https://arxiv.org/html/2608.14430","noteUrl":"/review?paper=review-neurips-2026-155438","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155440","title":"UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction","authors":["Junno Yun","Yasar Utku Alcalar","Mehmet Akcakaya"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.01298","noteUrl":"/review?paper=review-neurips-2026-155440","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155441","title":"More Than Meets the Eye? Uncovering the Reasoning-Planning Disconnect in Training Vision-Language Driving Models","authors":["Xurui Song","Shuo Huai","Jingjing Jiang","Jiayi Kong","Jun Luo"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The displayed clipping objective mishandles negative advantages","plainExplanation":"The standard clipped surrogate takes the minimum after multiplication by the signed advantage. Factoring a negative advantage out of a minimum changes the minimum to a maximum, so the printed expression is not the same objective.","primaryUrl":"https://arxiv.org/html/2510.04532","noteUrl":"/review?paper=review-neurips-2026-155441","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155442","title":"Beyond Trajectory Matching: Reflow with Marginal Distribution Alignment","authors":["Chen Wang","Peiran Yun","Pan Xie","Ke Deng"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Beyond Trajectory Matching: Reflow with Marginal Distribution Alignment","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.29287","noteUrl":"/review?paper=review-neurips-2026-155442","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155443","title":"Targeted Review for AI-Assisted Biodiversity Surveys: Active Continuous-Score Occupancy Modeling","authors":["Timm Haucke","Lauren Harrell","Justin Kay","Mary K Clapp","Sara Beery"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Targeted Review for AI-Assisted Biodiversity Surveys: Active Continuous-Score Occupancy Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.25657","noteUrl":"/review?paper=review-neurips-2026-155443","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155444","title":"Online Data Selection for Instruction Tuning via Gaussian Processes","authors":["Jun Wang","Quoc Phong Nguyen","Julien Monteil","Vu Nguyen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A symmetric score mixture need not beat its more uniform component","plainExplanation":"The top-set likelihood is not a product of independent draw probabilities. A symmetric score mixture need not beat its more uniform component.","primaryUrl":"https://arxiv.org/html/2606.30077","noteUrl":"/review?paper=review-neurips-2026-155444","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155446","title":"Optimal In-context Adaptivity and Distributional Robustness of Transformers","authors":["Tianyi Ma","Tengyao Wang","Richard J Samworth"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Approximation proof needs finite encoders and bounded outputs","plainExplanation":"Signed or zero regression functions make the prescribed encoder undefined, although the posterior regression function itself is well-defined.","primaryUrl":"https://arxiv.org/html/2510.23254","noteUrl":"/review?paper=review-neurips-2026-155446","collection":"neurips-2026","severity":"repairable","issueTypes":["citation_hypothesis_mismatch","definition_error","false_statement","formula_error","mathematical_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155447","title":"Agent2World: Learning to Generate Symbolic World Models via Adaptive Multi-Agent Feedback","authors":["Mengkang Hu","Bowei Xia","Yuran Wu","Ailing Yu","Yude Zou","Qiguang Chen","Shijian Wang","Jiarui Jin","Kexin Li","Wenxiang Jiao","Yuan Lu","Ping Luo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Agent2World: Learning to Generate Symbolic World Models via Adaptive Multi-Agent Feedback","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2512.22336","noteUrl":"/review?paper=review-neurips-2026-155447","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155448","title":"Subject-Relative Micro-Motion and Sleep Dynamics for Near-Infrared Video Sleep Staging","authors":["Kunmin Jang","Yourim Choi","Hun Heo","Dongik Park","Suahn Bae","Heonjun Lee","Hyun-Woo Shin","Hyung-Sin Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SVD sign ambiguity changes the saliency map","plainExplanation":"An SVD allows simultaneous sign reversal of a singular vector and its paired factor without changing the activations. With rank-one positive activations whose second channel is twice the first, the leading channel vector can be \\(u=(1,2)/\\sqrt5\\) or \\(-u\\), even with a simple largest singular value. Summing two such vectors can give \\(2u,-2u\\), or zero. Applying \\(\\operatorname{ReLU}(A W)\\) then gives a positive or zero map for the same activations. The visualization is underdefined independently of a library convention. Classification scores and physiological conclusions are not directly refuted.","primaryUrl":"https://arxiv.org/html/2609.05550","noteUrl":"/review?paper=review-neurips-2026-155448","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155452","title":"Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring","authors":["Weixin Guan","Liang Li","Jiapeng Liu","Bing Li","Peng Fu","Chengyang Fang","Xiaoshuai Hao","Can Ma","Weiping Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The decoding loop has no stated budget or natural-end fallback","plainExplanation":"Take a history-independent next-token distribution over newline, EOS and think-end with probabilities \\(0.7,0.2,0.1\\), decoded greedily. Every entropy is the same positive value, so after a full window the local/global entropy ratio is \\(1\\). With threshold \\(2\\), the trigger never fires, the unique greedy token remains nonterminal, and the index exceeds the stated maximum budget. The later answer budget can then be negative. The pseudocode lacks termination guards. A real decoding engine may already enforce them; no named checkpoint is alleged to hang.","primaryUrl":"https://arxiv.org/html/2603.14251","noteUrl":"/review?paper=review-neurips-2026-155452","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155456","title":"On the Optimal Sample Complexity of Offline Multi-Armed Bandits with KL Regularization","authors":["Kaixuan Ji","Qiwei Di","Heyang Zhao","Qingyue Zhao","Quanquan Gu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"On the Optimal Sample Complexity of Offline Multi-Armed Bandits with KL Regularization","plainExplanation":"No issue was found in the checked claims. The empirical-best-arm tail calculation is corrected.","primaryUrl":"https://arxiv.org/html/2605.02141","noteUrl":"/review?paper=review-neurips-2026-155456","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155457","title":"Generative Modeling under Non-Monotone MAR Missingness via Approximate Wasserstein Gradient Flows","authors":["Gitte Kremling","Jeffrey Näf","Johannes Lederer"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Generative Modeling under Non-Monotone MAR Missingness via Approximate Wasserstein Gradient Flows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.04567","noteUrl":"/review?paper=review-neurips-2026-155457","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155460","title":"Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs","authors":["Youngwoo Shin","Yusung Ro","Minseo Kim","Junmo Kim"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.01595","noteUrl":"/review?paper=review-neurips-2026-155460","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155461","title":"LeanSearch v2: Global Premise Retrieval for Lean 4 Theorem Proving","authors":["Guoxiong Gao","Zeming Sun","Jiedong Jiang","Yutong Wang","jd xu","Peihao Wu","Bryan Dai","Bin Dong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LeanSearch v2: Global Premise Retrieval for Lean 4 Theorem Proving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13137","noteUrl":"/review?paper=review-neurips-2026-155461","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155462","title":"CRANE: Constrained Reasoning Injection for Code Agents via Nullspace Editing","authors":["Mingzhi Zhu","Michele Merler","Raju Pavuluri","Stacy Patterson"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"CRANE: Constrained Reasoning Injection for Code Agents via Nullspace Editing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14084","noteUrl":"/review?paper=review-neurips-2026-155462","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155464","title":"Tensorion: A Tensor-Aware Generalization of the Muon Optimizer","authors":["Vladimir Bogachev","Vladimir Aletov","Alexander Molozhavenko","Sergei Kudriashov","Maxim Rakhuba"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Tensorion: A Tensor-Aware Generalization of the Muon Optimizer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.25975","noteUrl":"/review?paper=review-neurips-2026-155464","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155465","title":"PixelDiT2: Representation-Grounded Pixel Diffusion Transformers","authors":["Yongsheng Yu","Wei Xiong","Yichen Sheng","Shiqiu Liu","Jiebo Luo"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"PixelDiT2: Representation-Grounded Pixel Diffusion Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.24919","noteUrl":"/review?paper=review-neurips-2026-155465","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155467","title":"Grasp-Then-Plan with Failure Attribution: A Closed Two-Stage Framework for Precise and Generalizable Robotic Manipulation","authors":["Jiahao Xu","Peiyuan Wang","Hanzhuo Zhang","Zihao Yu","Tianyu Fu","Hao Chen","Xuanhao Xiang","Zixuan Li","Jianbo Yu","Chenchen Fu","Wanyuan Wang"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Grasp-Then-Plan with Failure Attribution: A Closed Two-Stage Framework for Precise and Generalizable Robotic Manipulation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.03385","noteUrl":"/review?paper=review-neurips-2026-155467","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155468","title":"FLAME: Adaptive Mixture-of-Experts for Continual Multimodal Multi-Task Learning","authors":["Xing Han","Shravan Chaudhari","Tanvi Ranade","Rama Chellappa","Suchi Saria"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FLAME: Adaptive Mixture-of-Experts for Continual Multimodal Multi-Task Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09355","noteUrl":"/review?paper=review-neurips-2026-155468","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155470","title":"Flow Matching for Count Data","authors":["Ganchao Wei","John Pearson"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Flow Matching for Count Data","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07746","noteUrl":"/review?paper=review-neurips-2026-155470","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155471","title":"Rollout Pass-Rate Control: Steering Binary-Reward RL Toward Its Most Informative Regime","authors":["Tianshu Zhu","Wenyu Zhang","Xiaoying Zuo","Tianlun","Haotian Zhao","yucheng-zeng","Jingnan gu","Daxiang Dong","Jianmin Wu"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Rollout Pass-Rate Control: Steering Binary-Reward RL Toward Its Most Informative Regime","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05112","noteUrl":"/review?paper=review-neurips-2026-155471","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155473","title":"ST-DiffEye: Diffusion-based Continuous Gaze Generation via Joint Scanpath-Trajectory Modeling","authors":["Brian Nlong Zhao","Ozgur Kara","Junho Kim","James Rehg"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"An arbitrary sequence distance does not yield a proper energy score","plainExplanation":"An arbitrary sequence distance does not yield a proper energy score. Coincident boundary anchors give conflicting indices and zero denominators.","primaryUrl":"https://arxiv.org/html/2606.15486","noteUrl":"/review?paper=review-neurips-2026-155473","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-155474","title":"Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning","authors":["Hyeonjin Kim","Hangyeol Jung","Heechan Yun","Sungjun Yun","Dong-Jun Han"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12122","noteUrl":"/review?paper=review-neurips-2026-155474","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155476","title":"A Local Geometric Analysis of Maximal Coding Rate Reduction via Error Bounds","authors":["Peng Wang","po chen","Huikang Liu","Rujun Jiang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A degenerate endpoint breaks the local error bound","plainExplanation":"At the included endpoint, scalar roots coalesce and the gradient vanishes quadratically in distance from the explicit reference set, contradicting the claimed linear error bound. The intended strict-interior theorem is not refuted by that example. Separately, the squared singular-value derivation subtracts an error on the wrong scale, and the gradient-perturbation coefficient misses a singular-value factor. A smaller local neighborhood supplies a valid singular-value estimate without an unstated normalization.","primaryUrl":"https://arxiv.org/html/2610.05072","noteUrl":"/review?paper=review-neurips-2026-155476","collection":"neurips-2026","severity":"major","issueTypes":["mathematical_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155480","title":"Optimizing Computational-Statistical Runtime for Wasserstein Distance Estimation","authors":["Peter M Jacobs","Jeff M Phillips"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Optimizing Computational-Statistical Runtime for Wasserstein Distance Estimation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20122","noteUrl":"/review?paper=review-neurips-2026-155480","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155485","title":"PACEvolve++: Improving Test-time Learning for Evolutionary Search Agents","authors":["Minghao Yan","Bo Peng","Benjamin Coleman","Ziqi Chen","Zhouhang Xie","Shuo Chen","Zhankui He","Noveen Sachdeva","Weili Wang","Ed Chi","Shivaram Venkataraman","Wang-Cheng Kang","Derek Cheng","Beidou Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PACEvolve++: Improving Test-time Learning for Evolutionary Search Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07039","noteUrl":"/review?paper=review-neurips-2026-155485","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155486","title":"Nearly-Optimal Algorithm for Adversarial Kernelized Bandits","authors":["Shogo Iwazaki"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Nearly-Optimal Algorithm for Adversarial Kernelized Bandits","plainExplanation":"No issue was found in the checked claims. The displayed algorithm now agrees with the observed-prefix posterior identity used in its proof.","primaryUrl":"https://arxiv.org/html/2605.10299","noteUrl":"/review?paper=review-neurips-2026-155486","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155487","title":"Beyond the Dirac Delta: Mitigating Diversity Collapse in Reinforcement Fine-Tuning for Image Generation","authors":["Jinmei Liu","Haoru Li","Zhenhong Sun","Chaofeng Chen","Yatao Bian","Hongdong Li","Bo Wang","Daoyi Dong","Zhi Wang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Beyond the Dirac Delta: Mitigating Diversity Collapse in Reinforcement Fine-Tuning for Image Generation","plainExplanation":"A retained terminal diversity bonus can trade task reward for diversity even under exact optimization. It changes the objective rather than preserving the original task optimum.","primaryUrl":"https://arxiv.org/html/2601.12401v2","noteUrl":"/review?paper=review-neurips-2026-155487","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","missing_assumption"],"classifications":["critical"]},{"slug":"review-neurips-2026-155488","title":"On Concentration Inequalities for Sampling without Replacement","authors":["Shaojie Li","Bowei Zhu","Yong Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"On Concentration Inequalities for Sampling without Replacement","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/1309.4029","noteUrl":"/review?paper=review-neurips-2026-155488","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155491","title":"VCR: Learning Valid Contextual Representation for Incomplete Wearable Signals","authors":["Yuxuan Weng","Wenhan Luo","Qijia Shao"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"VCR: Learning Valid Contextual Representation for Incomplete Wearable Signals","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.18837","noteUrl":"/review?paper=review-neurips-2026-155491","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155493","title":"SkillOrchestra: Learning to Route Agents via Skill Transfer","authors":["Jiayu Wang","Yifei Ming","Zixuan Ke","Shafiq Joty","Aws Albarghouthi","Frederic Sala"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SkillOrchestra: Learning to Route Agents via Skill Transfer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.19672","noteUrl":"/review?paper=review-neurips-2026-155493","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155494","title":"HOPSE: Scalable Higher-Order Positional and Structural Encoder for Combinatorial Representations","authors":["Guillermo Bernárdez","Marco Montagna","Louis Van Langendonck","Martin Carrasco","Amirreza Akbari","Louisa Cornelis","Mathilde Papillon","Pere Barlet-Ros","Nina Miolane","Lev Telyatnikov"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"WL-equivalent features and injective maps do not force strict WL superiority","plainExplanation":"Theorem 4.1 uses injective feature maps and at least one WL-equivalent neighborhood encoder as sufficient conditions for strict superiority over CCWL. Injective maps preserve the distinctions supplied by a WL-equivalent encoder but do not force additional distinctions, so the stated conditions establish no strict superiority. The sufficient conditions do not force strict improvement. The flexible framework can be strictly stronger when supplied with genuinely stronger encoders.","primaryUrl":"https://arxiv.org/html/2505.15405","noteUrl":"/review?paper=review-neurips-2026-155494","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155499","title":"Augmented Lagrangian Method for Last-Iterate Convergence for Constrained MDPs","authors":["Michael Lu","Max Lin","Mo Chen","Sharan Vaswani"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The approximation-bias condition has the accuracy exponent reversed","plainExplanation":"For three-action probabilities, impose \\(p_1+0.499p_3\\geq0.4995\\) and \\(p_2+0.499p_3\\geq0.4995\\), and reward action three. Adding the constraints gives \\(p_3\\leq0.5\\), attained at \\((0.25,0.25,0.5)\\). Equality of its optimal action payoffs forces equal multipliers \\(\\lambda\\) satisfying \\(\\lambda=1+0.998\\lambda\\), hence \\(\\lambda=500\\). Each separately maximal slack is \\(0.5005\\), giving the claimed bound below \\(2\\). The joint strict policy instead has margin \\(0.0005\\). Strong duality and qualitative convergence may survive. The printed multiplier and downstream quantitative constants need correction.","primaryUrl":"https://arxiv.org/html/2605.11694","noteUrl":"/review?paper=review-neurips-2026-155499","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","missing_assumption","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155505","title":"Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG","authors":["Hanning Guo","Hanwen Bi","Farah Abdellatif","Andrei Galbenus","N. J Shah","Abigail Morrison","Jürgen Dammers"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Unit signed row sums do not bound attention amplification","plainExplanation":"Signed row sum controls constant signals, whereas stability against arbitrary bounded values depends on the absolute row sum.","primaryUrl":"https://arxiv.org/html/2602.23410","noteUrl":"/review?paper=review-neurips-2026-155505","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155506","title":"PORTool: Importance-Aware Policy Optimization with Rewarded Tree for Multi-Tool-Integrated Reasoning","authors":["Feijie Wu","Weiwu Zhu","Yuxiang Zhang","Soumya Chatterjee","Jiarong Zhu","Fan Mo","Rong Luo","Jing Gao"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Clipping a sum of opposite-sign advantages is not additive","plainExplanation":"The main objective clips a mixed trajectory/fork advantage once, while Theorem 3.2 identifies it with two separately clipped objectives. PPO clipping is not additive across opposite-sign advantages, so clipping their mixture once does not equal the appendix’s sum of separately clipped objectives.","primaryUrl":"https://arxiv.org/html/2510.26020","noteUrl":"/review?paper=review-neurips-2026-155506","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155507","title":"Test-Time Scaling with Diffusion Language Models via Reward-Guided Stitching","authors":["Roy Miles","Aysim Toker","Andreea-Maria Oncescu","Jiankang Deng","Ismail Elezi"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Test-Time Scaling with Diffusion Language Models via Reward-Guided Stitching","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.22871","noteUrl":"/review?paper=review-neurips-2026-155507","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155508","title":"Omni-Interactive Universal Embedder","authors":["Wei-Yao Wang","Kazuya Tateishi","Shuyang Cui","christian simon","Takashi Shibuya","Shusuke Takahashi","Yuki Mitsufuji"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Omni-Interactive Universal Embedder","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.27044","noteUrl":"/review?paper=review-neurips-2026-155508","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155511","title":"Pointwise Lipschitz Continuous Graph Algorithms","authors":["Quanquan C Liu","Grigoris Velegkas","Yuichi Yoshida","Felix Zhou"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Repeated insertions and deletions incur recourse despite zero net change","plainExplanation":"The proximal sensitivity bound is applied outside its controlled input set. Repeated insertions and deletions incur recourse despite zero net change.","primaryUrl":"https://arxiv.org/html/2405.08938","noteUrl":"/review?paper=review-neurips-2026-155511","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155513","title":"Controllable Generative Sandbox for Causal Inference","authors":["Qi Zhang","Harsh Parikh","Ashley I Naimi","Razieh Nabi","Christopher Kim","Timothy L Lash"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A constant conditional density ratio must be one","plainExplanation":"Two normalized conditional distributions cannot have a density ratio equal to \\(e\\) throughout the treated distribution's support. The specified population target is infeasible under the paper's definition.","primaryUrl":"https://arxiv.org/html/2603.03587","noteUrl":"/review?paper=review-neurips-2026-155513","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_error","false_statement","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155515","title":"The World is Not Mono: Enabling Spatial Understanding in Large Audio-Language Models","authors":["Yuhuan You","Lai Wei","Tianshu Qu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"The World is Not Mono: Enabling Spatial Understanding in Large Audio-Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.02954","noteUrl":"/review?paper=review-neurips-2026-155515","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155522","title":"Approximation Guarantees for Robust Aggregation in Federated Learning","authors":["Mélanie Cambus","Darya Melnyk","Tijana Milentijević","Stefan Schmid"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Box validity alone permits much larger normalized error","plainExplanation":"Box validity alone permits much larger normalized error. The FedAvg display omits the averaging factor.","primaryUrl":"https://arxiv.org/html/2506.15264v1","noteUrl":"/review?paper=review-neurips-2026-155522","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-155526","title":"Towards Convergence of PPO: An Approximate Descent Approach","authors":["Leif Döring","Daniel A Schmidt","Moritz Melcher","Sebastian Kassing","Benedikt Wille","Tilman Aach","Simon Weissmann"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The finite-sum formula can make a norm upper bound negative","plainExplanation":"The next-state value needs the next time index. The variance calculation drops the estimator’s horizon factor. The finite-sum formula can make a norm upper bound negative.","primaryUrl":"https://arxiv.org/html/2602.03386v1","noteUrl":"/review?paper=review-neurips-2026-155526","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155527","title":"On the Design Space of Discrete Diffusion Online Adaptation for Molecular Optimization","authors":["Trevor Chen","Ariel Dai","Jason Yang","Riccardo De Santi","Daniel Khalil","Wenda Chu","Nate Gruver","Pranav Murugan","Alexander Goldberg","Maruan Al-Shedivat","Yisong Yue"],"year":2026,"category":"Molecular models & equivariance","status":"finding","issueTitle":"A Monte Carlo ELBO converges to a bound rather than the log marginal","plainExplanation":"The path identity is \\(\\log p(x)=\\mathbb E_q\\log[p(x,\\mathrm{path})/q(\\mathrm{path}\\mid x)]+D_{\\mathrm{KL}}(q\\|p(\\mathrm{path}\\mid x))\\). A one-step half-masked model in the receipt has \\(p(x)=0.65\\), ELBO approximately \\(-0.45815\\), and \\(\\log p(x)\\approx-0.43078\\). More samples converge to the lower bound with its positive KL gap. Separately, exact reversal of data probabilities \\((0.8,0.2)\\) under half-masking gives the displayed forward-conditioned average \\((0.9,0.6)\\), not the marginal; no schedule-only scalar repairs both states. The likelihood justification fails. The deliberately approximate cheap modifier and empirical molecular gains are not disproved.","primaryUrl":"https://arxiv.org/html/2607.02834","noteUrl":"/review?paper=review-neurips-2026-155527","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155528","title":"LAPLEX: The FFT of Learnable Laplace Kernels","authors":["Łukasz Struski","Hanna Blazhko","Piotr Kubaty","Jacek Tabor"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The scan formulas omit the right exterior bucket","plainExplanation":"For one row anchor \\(a_1=0\\) and column anchor \\(b_1=1\\), every included bucket statistic is zero. The printed matvec and Gram are zero, but the actual kernel values are \\(e^{-1}\\) and \\(e^{-2}\\). The omitted support includes \\(b\\geq a_n\\), including equality.\n\nA tail anchored at \\(a_n\\) contributes \\(e^{a_i-a_n}\\sum_{b\\geq a_n}x_b e^{a_n-b}\\) to the matvec and \\(e^{a_i+a_j-2a_n}\\sum_{b\\geq a_n}D_b e^{2(a_n-b)}\\) to the Gram. The formulas need a right-tail correction; this can preserve their asymptotic complexity.","primaryUrl":"https://arxiv.org/html/2605.24584","noteUrl":"/review?paper=review-neurips-2026-155528","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155531","title":"Dandelions: A Spherical Flower for Neural Simulation of Planetary Dynamics","authors":["Till Muser","Giovanni Abati","Ivan Dokmanić"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Dandelions: A Spherical Flower for Neural Simulation of Planetary Dynamics","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.27521","noteUrl":"/review?paper=review-neurips-2026-155531","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155534","title":"Convergence Guarantees for Federated SARSA with Local Training and Heterogeneous Agents","authors":["Paul Mangold","Eloïse Berthier","Eric Moulines"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A shared initialization bias does not vanish by averaging agents","plainExplanation":"Pointwise vector bounds do not control the stated sum-of-suprema norms. A shared initialization bias does not vanish by averaging agents.","primaryUrl":"https://arxiv.org/html/2512.17688","noteUrl":"/review?paper=review-neurips-2026-155534","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155535","title":"Clearer Sight, Fewer Lies: Oriented Pickup Preference Optimization for Multimodal Hallucination Mitigation","authors":["Xin Zou","Haolin Deng","Yibo Yan","Shuliang Liu","Zhiwei Jin","Chen Chen","Haonan Lu","Xuming Hu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The rejected-path token term encourages its own KL drift","plainExplanation":"Equation (8) subtracts rejected-path sequential KL from a detached chosen-path KL, while describing this term as a penalty keeping the policy near its reference. The differentiated token term has negative KL sign, so its own gradient-descent contribution increases rejected-path KL rather than penalizing drift. This concerns the individual token term. Other losses may oppose it, so the argument does not imply that total training KL must increase.","primaryUrl":"https://arxiv.org/html/2606.29805","noteUrl":"/review?paper=review-neurips-2026-155535","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155536","title":"A Measure-Theoretic Analysis of Reasoning: Structural Generalization and Approximation Limits","authors":["Yuyang Zhang","Yifu Zhang","Xuehai Zhou","Xiaoyin Chen"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A Measure-Theoretic Analysis of Reasoning: Structural Generalization and Approximation Limits","plainExplanation":"No issue was found in the checked claims. Current Assumption 5 states \\(\\mathcal L_f(\\chi)\\ge C_{\\mathrm{depth}}\\) whenever \\(\\chi_1>\\alpha_{\\mathrm{circ}}L\\). Integrating over that region and using nonnegativity elsewhere gives \\(\\int\\mathcal L_f\\,d\\nu_T\\ge C_{\\mathrm{depth}}\\nu_T(\\mathcal X_{\\mathrm{severed}})\\). Assumption 1 then supplies the stated \\(-\\epsilon\\) slack. Thus the specific missing-premise objection is resolved in the conditional result.","primaryUrl":"https://arxiv.org/html/2605.19944","noteUrl":"/review?paper=review-neurips-2026-155536","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155539","title":"Self-Improving World Modelling with Latent Actions","authors":["Yifu QIU","Zheng Zhao","Waylon Li","Yftah Ziser","Anna Korhonen","Shay Cohen","Edoardo Ponti"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Self-Improving World Modelling with Latent Actions","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.06130","noteUrl":"/review?paper=review-neurips-2026-155539","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155540","title":"COVD: Continual Open-Vocabulary Object Detection with Novel Concept Injection","authors":["Yupeng Zhang","Ruize Han","Yuzhong Feng","Zixin Ren","Yuntong Tian","Liang Wan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"COVD: Continual Open-Vocabulary Object Detection with Novel Concept Injection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.27116","noteUrl":"/review?paper=review-neurips-2026-155540","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155544","title":"When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution","authors":["Yu Shi","Yuyao Zhang","Yu-Wing Tai"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.09133","noteUrl":"/review?paper=review-neurips-2026-155544","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155547","title":"COMPLLLM: Fine-tuning LLMs to Discover Complementary Signals for Decision-making","authors":["Ziyang Guo","Yifan Wu","Jason Hartline","Kenneth Holstein","Jessica Hullman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Conditional-label variance depends on the posterior rather than the prior","plainExplanation":"A supported but uninformative signal can trigger zero divided by zero. Conditional-label variance depends on the posterior rather than the prior.","primaryUrl":"https://arxiv.org/html/2602.19458","noteUrl":"/review?paper=review-neurips-2026-155547","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155550","title":"Beyond Parameter Aggregation: Semantic Consensus for Federated Fine-Tuning of LLMs","authors":["Amr Abourayya","Jens Kleesiek","Michael Kamp"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A finite sampled gradient gap is not the population bias","plainExplanation":"A finite sampled gradient gap is not the population bias.","primaryUrl":"https://arxiv.org/html/2605.11857","noteUrl":"/review?paper=review-neurips-2026-155550","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155551","title":"Coding with \"Enemy\": Can Human Developers Detect AI Agent Sabotage?","authors":["Jingheng Ye","Huiqi Zou","Simon Yu","Weiyan Shi"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Coding with \"Enemy\": Can Human Developers Detect AI Agent Sabotage?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.05647","noteUrl":"/review?paper=review-neurips-2026-155551","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155554","title":"Justitia: Fair and Efficient Scheduling of Task-parallel LLM Agents with Selective Pampering","authors":["Mingyan Yang","Guanjie Wang","Manqi Luo","Yifei Liu","Chen Chen","Han Zhao","Yu Feng","Quan Chen","Minyi Guo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Justitia: Fair and Efficient Scheduling of Task-parallel LLM Agents with Selective Pampering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.17015","noteUrl":"/review?paper=review-neurips-2026-155554","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155556","title":"The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining","authors":["Gavin Vy Nguyen","Ziqi Xu","Jeffrey Chan","Estrid He","Feng Xia","Renqiang Luo","Erik Cambria","Xiuzhen Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A gated component’s task derivative includes downstream Jacobians","plainExplanation":"Appendix D.1 identifies the actual joint-gating derivative with the ungated component projection times its sigmoid derivative, while holding other gate values fixed. Holding other gates fixed does not freeze their downstream activations; the omitted downstream Jacobian can reverse the direct-projection gradient sign.","primaryUrl":"https://arxiv.org/html/2609.32964","noteUrl":"/review?paper=review-neurips-2026-155556","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155557","title":"One-Shot Generative Flows: Existence and Obstructions","authors":["Panagiotis Tsimpos","Daniel Sharp","Youssef Marzouk"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Positive interval width does not control the speed of its boundaries","plainExplanation":"The bounded-velocity wording is stronger than its displayed formal condition. Positive interval width does not control the speed of its boundaries. Mode mass need not equal cumulative mass at its boundary.","primaryUrl":"https://arxiv.org/html/2604.15439","noteUrl":"/review?paper=review-neurips-2026-155557","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155558","title":"Free Decompression with Algebraic Spectral Curves","authors":["Siavash Ameli","Chris van der Heide","Liam Hodgkinson","Michael Mahoney"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A multiple leading root need not give an integer-power Laurent series","plainExplanation":"The active trajectory count can never exceed \\(2k_0\\): the loop only visits initial slots and no step inserts edges. A split from one bulk to three requires six edges, exceeding the initial two. The paper’s own benchmark describes that event; the independent polynomial critical-point and physical Stieltjes-root checks confirm that the extra intervals are genuine, rather than ghost branches. The printed tracker is incomplete for splitting. This does not invalidate the spectral-curve density method or the separate cusp equations.","primaryUrl":"https://arxiv.org/html/2605.03634","noteUrl":"/review?paper=review-neurips-2026-155558","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","interpretation_error","lemma_error","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155561","title":"The Best-Laid SCHEMEs: Coordinated Sabotage and Monitoring in Multi-Agent Systems","authors":["Nikolay Radev","Lennart J Haas","Benjamin Arnav","Pablo Bernabeu-Perez"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Deterministic easy and hard instances produce no any-versus-all gap","plainExplanation":"Deterministic easy and hard instances produce no any-versus-all gap.","primaryUrl":"https://arxiv.org/html/2605.29178","noteUrl":"/review?paper=review-neurips-2026-155561","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155563","title":"MedZERO: Self-Evolving Agents for Open-Ended Medical Reasoning Through Controlled Knowledge Accumulation","authors":["Xilin Dang","Weilin Ruan","Xue Yang","Jinghao Wang","Xiaowei Hu","Jinpeng Li","Pheng-Ann Heng"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"MedZERO: Self-Evolving Agents for Open-Ended Medical Reasoning Through Controlled Knowledge Accumulation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.08327","noteUrl":"/review?paper=review-neurips-2026-155563","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155564","title":"Beyond Steering Vector: Flow-based Activation Steering for Inference-Time Intervention","authors":["Zehao Jin","Ruixuan Deng","Junran Wang","Xinjie Shen","Chao Zhang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Beyond Steering Vector: Flow-based Activation Steering for Inference-Time Intervention","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.05892","noteUrl":"/review?paper=review-neurips-2026-155564","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155567","title":"Qrita: High-performance Top-k and Top-p using Pivot-based Truncation and Selection","authors":["Jongseok Park","Sunga Kim","Alvin Cheung","Ion Stoica"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The equal-logit branch skips the claimed tie selection","plainExplanation":"The equal-logit branch skips the claimed tie selection. Numeric interval search depends on logit gaps rather than candidate count.","primaryUrl":"https://arxiv.org/html/2602.01518","noteUrl":"/review?paper=review-neurips-2026-155567","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155569","title":"SEAD: Competence-Aware On-Policy Distillation via Entropy-Guided Supervision","authors":["Michael Lee","Zelei Cheng","Yu Wang","Renkun Ni","Sambit Sahu","Shi-Xiong Zhang","William Campbell"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Curriculum guarantees need stronger variance and competence conditions","plainExplanation":"Exact per-prompt gradients can increase the average stage loss despite zero conditional noise.","primaryUrl":"https://arxiv.org/html/2606.28562","noteUrl":"/review?paper=review-neurips-2026-155569","collection":"neurips-2026","severity":"major","issueTypes":["mathematical_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155572","title":"MLLMEraser: Achieving Test-Time Unlearning in Multimodal Large Language Models through Activation Steering","authors":["Chenlu Ding","Jiancan Wu","Leheng Sheng","Fan Zhang","Yancheng Yuan","Xiang Wang","Xiangnan He"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The closed form minimizes squared norms rather than the displayed norm objective","plainExplanation":"Equation (11) uses unsquared Frobenius norms, while Equation (12) gives a ridge least-squares closed form. The displayed objective uses unsquared norms, whereas the closed form minimizes squared ridge loss; the two optimization problems have different minimizers.","primaryUrl":"https://arxiv.org/html/2510.04217","noteUrl":"/review?paper=review-neurips-2026-155572","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155573","title":"Learning to Solve Generative ODEs Beyond the Linear Span","authors":["Sihyeon Kim","Seunghun Lee","Vikas Singh","Hyunwoo J. Kim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Learning to Solve Generative ODEs Beyond the Linear Span","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.08672","noteUrl":"/review?paper=review-neurips-2026-155573","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155576","title":"Denoise First, Orthogonalize Later: Understanding Momentum in Muon via Spectral Filtering","authors":["Xianliang Li","Zihan Zhang","Weiyang Liu","Han Bao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The finite-buffer proof drops an initialization term","plainExplanation":"The missing term is positive at finite time. Describing the transient as negligible does not justify an exact equality or the printed constant in the bound. Polar scale invariance offers a direct repair.","primaryUrl":"https://arxiv.org/html/2606.03899","noteUrl":"/review?paper=review-neurips-2026-155576","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155578","title":"V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models","authors":["Xinying Lin","Xuyang Liu","Yiyu Wang","Teng Ma","Jiasheng Li","Zichen Wen","Wenqi Ren"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.27650","noteUrl":"/review?paper=review-neurips-2026-155578","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155580","title":"On the Recall Scaling Laws in Mamba: A Theoretical and Mechanistic Study via Hashing","authors":["Yuval Koren","Assaf Ben-Kish","Raja Giryes","Lior Wolf","Itamar Zimerman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Conditioning on the target score does not make empty scores independent","plainExplanation":"Equation (6) adds the successor token for every occurrence of the queried key. Use context pairs \\(A,a,B,b\\), query \\(A\\), then allowed padding \\(B,a,B,a,a,a\\), then query \\(B\\). The predecessors \\(B\\) contribute \\(b,a,a\\), so the circuit selects \\(a\\) rather than the required \\(b\\). Context pairs and queries remain distinct; value-only output masking does not remove \\(a\\). The displayed circuit’s unrestricted perfect-recall proof fails. Another gated or positional circuit is not ruled out.","primaryUrl":"https://arxiv.org/html/2609.07681","noteUrl":"/review?paper=review-neurips-2026-155580","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155581","title":"Direct Estimation of Schrödinger Bridge Time-Series Drifts: Finite-Sample, Asymptotic, and Adaptive Guarantees","authors":["Othmane Mazhar","Huyen PHAM"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Direct Estimation of Schrödinger Bridge Time-Series Drifts: Finite-Sample, Asymptotic, and Adaptive Guarantees","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05432","noteUrl":"/review?paper=review-neurips-2026-155581","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155584","title":"Neural networks are more modular than single neurons suggest","authors":["Shreya Saha","Benjamin Bergen","Meenakshi Khosla"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Neural networks are more modular than single neurons suggest","plainExplanation":"No issue was found in the checked claims. The contextual guarantee is valid for the named selection proxy. Extending it to the separate full DDI formula or held-out performance would be stronger, but that extension is not required by the stated checkpoint guarantee.","primaryUrl":"https://arxiv.org/html/2610.12842","noteUrl":"/review?paper=review-neurips-2026-155584","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155588","title":"InfoSFT: Learn More and Forget Less with Information-Aware Token Weighting","authors":["Mahdi Sabbaghi","George J. Pappas","Adel Javanmard","Hamed Hassani"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Equal movement budgets can make the two weighting schemes identical","plainExplanation":"The substitution introduces a spurious probability factor. Equal movement budgets can make the two weighting schemes identical.","primaryUrl":"https://arxiv.org/html/2605.14967","noteUrl":"/review?paper=review-neurips-2026-155588","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155589","title":"Branching Flows: Discrete, Continuous, and Manifold Flow Matching with Splits and Deletions","authors":["Lukas Billera","Hedwig Nora Nordlinder","Jack C Ryder","Anton Oresten","Aron Stålmarck","Theodor M Björk","Ben Murrell"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"An unshifted Poisson draw cannot guarantee enough endpoint leaves","plainExplanation":"Appendix A.2 samples deletion padding from an unshifted Poisson law and claims that each augmented endpoint group contains at least as many elements as its initial roots. An unshifted Poisson padding count can fall below the deterministic root–leaf deficit, leaving too few leaves for the required forest.","primaryUrl":"https://arxiv.org/html/2511.09465","noteUrl":"/review?paper=review-neurips-2026-155589","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155590","title":"Conveyance: A Versatile Framework for Learning in Structured Class Spaces","authors":["Yasser Taha","Grégoire Montavon","Nils Körber"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Conveyance: A Versatile Framework for Learning in Structured Class Spaces","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.28420","noteUrl":"/review?paper=review-neurips-2026-155590","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155593","title":"LEIA: Learned Environment for Interactive Architected Materials","authors":["Haiqian Yang","Yuan Cao","Markus Buehler"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Linear tensor relaxation does not preserve unit determinant","plainExplanation":"Linear tensor relaxation does not preserve unit determinant.","primaryUrl":"https://arxiv.org/html/2605.28368","noteUrl":"/review?paper=review-neurips-2026-155593","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155594","title":"High-dimensional Limit of SGD for Diagonal Linear Networks","authors":["Begoña García Malaxechebarría","Courtney Paquette","Maryam Fazel","Dmitriy Drusvyatskiy"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Gaussian projection has the wrong scale in the gradient martingale proof","plainExplanation":"The gradient proof assigns the wrong scale to a Gaussian projection, and the Hessian proof drops a nonzero trace contribution. Both claimed martingale rates fail on a permitted linear model. Additional stability, boundary, loss-normalization and remainder steps need explicit repairs.","primaryUrl":"https://arxiv.org/html/2605.17177","noteUrl":"/review?paper=review-neurips-2026-155594","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","missing_assumption","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155596","title":"Neural Proposals, Symbolic Guarantees: Neuro-Symbolic Graph Generative Modeling","authors":["Chuqin Geng","Li Zhang","Mark Zhang","Zhaoyue(Rebecca) Wang","Haolin Ye","Xujie Si"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The reconstruction proof must allow more than two shared-edge copies","plainExplanation":"A minimum cycle basis need not have pairwise edge overlap. The two-copy assertion is too restrictive, although the later canonical-class construction already handles higher multiplicity.","primaryUrl":"https://arxiv.org/html/2602.16954","noteUrl":"/review?paper=review-neurips-2026-155596","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["proof_gap"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-155597","title":"Neural-Behavioral Representation of Natural Whole-body Movement in Monkeys","authors":["Jieshi He","Puzhe Li","Yanan Sui","Mu-ming Poo"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Neural-Behavioral Representation of Natural Whole-body Movement in Monkeys","plainExplanation":"The declared block dimensions add to \\(151\\). They cannot also describe a concatenated vector of dimension \\(152\\).","primaryUrl":"https://arxiv.org/html/2605.29355","noteUrl":"/review?paper=review-neurips-2026-155597","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["algorithm_specification","notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-155600","title":"Free Heavy-Tailed Lunch for Muon: A Theoretical Justification of Empirical Success","authors":["Florian Hübler","Thomas Pethick","Suvrit Sra"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Symmetric smoothness does not give the stated martingale coefficient","plainExplanation":"Symmetric smoothness does not give the stated martingale coefficient. The completed-square constant is missing a squared norm. The polar-factor restatement has the wrong sign and lacks its constraint.","primaryUrl":"https://arxiv.org/html/2606.14560","noteUrl":"/review?paper=review-neurips-2026-155600","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-155601","title":"Bayes-Sufficient Compression Is Not Enough: How Communication Helps in Multi-Agent Systems?","authors":["Yi Xie","Zhanke Zhou","Yi Fan","Yong Ge","Bo Han","Bo Liu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Bayes-Sufficient Compression Is Not Enough: How Communication Helps in Multi-Agent Systems?","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03769","noteUrl":"/review?paper=review-neurips-2026-155601","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155602","title":"Analytical Correction for Subsampling Bias in Drifting Models","authors":["Jiaru Zhang","Zeyun Deng","Juanwu Lu","Ziran Wang","Ruqi Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Analytical Correction for Subsampling Bias in Drifting Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.27239","noteUrl":"/review?paper=review-neurips-2026-155602","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155603","title":"Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation","authors":["Amin Karimi Monsefi","Dominic Culver","Nikhil Bhendawade","Manuel R Ciosici","Yizhe Zhang","Irina Belousova"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The chained target estimator is not the classical fourth-order Runge–Kutta construction","plainExplanation":"Classical RK4 stages restart from the initial state. Reusing the preceding midpoint and repeatedly advancing by $h/2$ gives a different construction. The weights $1,2,2,1$ alone do not establish fourth-order integration.","primaryUrl":"https://arxiv.org/html/2605.07924","noteUrl":"/review?paper=review-neurips-2026-155603","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","unsupported_claim"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155604","title":"Performative Prediction with Selective Labels","authors":["Giovani Valdrighi","Isabel Valera","Marcos M. Raimundo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Performative Prediction with Selective Labels","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.08272","noteUrl":"/review?paper=review-neurips-2026-155604","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155605","title":"Hallucination in World Models is Predictable and Preventable","authors":["Nick Hansen","Xiaolong Wang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Hallucination in World Models is Predictable and Preventable","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.27326","noteUrl":"/review?paper=review-neurips-2026-155605","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155606","title":"Why Cross-Skeleton Retargeting Is Non-Identifiable: Structural Limits of Generative Motion Models","authors":["Zhiyuan Li","Wenyan Yang","Pekka Marttinen","Joni Pajarinen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Why Cross-Skeleton Retargeting Is Non-Identifiable: Structural Limits of Generative Motion Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.37297","noteUrl":"/review?paper=review-neurips-2026-155606","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155610","title":"Sparsely Supervised Diffusion","authors":["Wenshuai Zhao","Zhiyuan Li","Yi Zhao","Mohammad Vali","Martin Trapp","Joni Pajarinen","Juho Kannala","Arno Solin"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Sparsely Supervised Diffusion","plainExplanation":"The covariance calculation is correct for masked inputs. It does not establish a change in the population objective or its deterministic learning dynamics for the loss-only masking specified in Equation (8).","primaryUrl":"https://arxiv.org/html/2602.02699","noteUrl":"/review?paper=review-neurips-2026-155610","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155612","title":"Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE","authors":["Haozhan Tang","Zerui Wang","Yuxian Gu","Song Han","Han Cai"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.07740","noteUrl":"/review?paper=review-neurips-2026-155612","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155613","title":"Compositional Reasoning in Language Models under Reinforcement Learning Post-Training","authors":["Yu He","Yingxi Li","Yifei Wang","Ellen Vitercik"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Compositional Reasoning in Language Models under Reinforcement Learning Post-Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.19465","noteUrl":"/review?paper=review-neurips-2026-155613","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155615","title":"Learning Sparse Compositional Functions with Norm-Constrained Neural Networks","authors":["shuo HUANG","Lorenzo Fiorito","Lorenzo Rosasco","Tomaso Poggio"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Sparse Compositional Functions with Norm-Constrained Neural Networks","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25608","noteUrl":"/review?paper=review-neurips-2026-155615","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155616","title":"Efficient Dataset Distillation for Pre-Trained Self-Supervised Models via Statistical Flow Matching","authors":["Qianxin Xia","JIAWEI DU","Yuhan Zhang","xin zhang","Xuewan He","Wenbo Jiang","Jielei Wang","Tao Luo","Guoming Lu"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Independent class weights need not be identically distributed","plainExplanation":"Theorem 1 permits class-specific Gaussian variances but claims exchangeable softmax probabilities with every class mean equal to \\(1/C\\). Independence with class-specific variances does not give permutation symmetry, so softmax probabilities need not be exchangeable or have equal class means.","primaryUrl":"https://arxiv.org/html/2602.05391","noteUrl":"/review?paper=review-neurips-2026-155616","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155617","title":"Learning-Augmented Online Scheduling with Parsimonious Preemption","authors":["Mugen Blue","Sungjin Im","Alexander Lindermayr"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning-Augmented Online Scheduling with Parsimonious Preemption","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.23255","noteUrl":"/review?paper=review-neurips-2026-155617","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155619","title":"A Memory Efficient Unified Algorithm for Online Learning of Linear Dynamical Systems","authors":["Yuval Ran-Milo","Angelos Assos","Elad Hazan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The regret bound omits its regularization dependence","plainExplanation":"An arbitrary fixed regularizer cannot satisfy the printed comparator regret bound without a regularization penalty or explicit norm-dependent tuning. A feature-rescaling example violates that lemma while meeting its displayed assumptions. The final rate also becomes zero at the explicitly allowed minimum complexity because it uses log one. Restoring a positive logarithmic convention preserves the intended asymptotic rate; neither finding proves that linear dynamical systems are unlearnable.","primaryUrl":"https://arxiv.org/html/2607.02050","noteUrl":"/review?paper=review-neurips-2026-155619","collection":"neurips-2026","severity":"major","issueTypes":["formula_error","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-155620","title":"Generative Scenario Rollouts for End-to-End Autonomous Driving","authors":["Rajeev Yasarla","Deepti Hegde","Shizhong Han","Hsin-Pai Cheng","Yunxiao Shi","MeysamSadeghi","Shweta Mahajan","Apratim Bhattacharyya","Litian Liu","Risheek Garrepalli","Thomas Svantesson","Mohammad Ghavamzadeh","Fatih Porikli","Herbert Cai"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"KL between continuous token collections needs a probability model","plainExplanation":"KL requires two probability measures on a shared measurable space. A token collection can induce empirical point masses, a Gaussian approximation, a categorical distribution or other models, each yielding a different divergence. Empirical Dirac measures commonly give infinite KL for noncoincident tokens. Neither a distributional parameterization nor the matching of tokens is specified by the equation. The loss is underdefined as written. The presence of a VAE elsewhere does not specify a Gaussian law for every VLA token, and no actual implementation is judged.","primaryUrl":"https://arxiv.org/html/2601.11475","noteUrl":"/review?paper=review-neurips-2026-155620","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155622","title":"PiCA: Pivot-Based Credit Assignment For Search Agentic Reinforcement Learning","authors":["Dongyi Liu","Yifan Niu","Qinwen Wang","Han Xiao","Jia Li"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The pivot-gain loss is undefined on allowed predictions and unbounded below","plainExplanation":"Equation (8) minimizes minus the logarithm of relative success gain \\(g=(f_t-f_{t-1})/f_{t-1}\\), with success probabilities otherwise unconstrained. Unconstrained success predictions can give nonpositive relative gain, making its logarithm undefined; even positive gains allow an unbounded-below loss through a vanishing prior prediction. The printed reward-model objective is not well defined on its stated probability domain; a barrier interpretation would require additional feasibility conditions.","primaryUrl":"https://arxiv.org/html/2605.09287","noteUrl":"/review?paper=review-neurips-2026-155622","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155624","title":"G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models","authors":["Junxian Li","Kai Liu","Zizhong Ding","Zhixin Wang","Zhikai Chen","Renjing Pei","Yulun Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12309","noteUrl":"/review?paper=review-neurips-2026-155624","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155629","title":"Demystifying Numerical Errors in LLM Inference: Achieving Reproducible Inference for Mission-Critical Tasks with HEAL","authors":["Zhenting Zhu","Lucas Thai","Shan Yu","Yicheng Liu","Yifan Qiao","Chenxi Wang","Harry Xu","Junyi Shu"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Machine epsilon does not bound absolute dot-product accumulation error","plainExplanation":"The FP16 spacing is off by a factor of two and the comparison reverses direction. Machine epsilon does not bound absolute dot-product accumulation error.","primaryUrl":"https://arxiv.org/html/2606.21023v1","noteUrl":"/review?paper=review-neurips-2026-155629","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155631","title":"Self-Programmed Execution for Language-Model Agents","authors":["Luke J O&amp;#x27;Connor"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Self-Programmed Execution for Language-Model Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.06898","noteUrl":"/review?paper=review-neurips-2026-155631","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155634","title":"TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking","authors":["Jisu Nam","Jahyeok Koo","Soowon Son","Jaewoo Jung","Honggyu An","Junhwa Hur","Seungryong Kim"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12587","noteUrl":"/review?paper=review-neurips-2026-155634","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155635","title":"Hardware-Friendly Token-Group Activation Quantization for Low-Bit Mamba Super-Resolution","authors":["Jinwoo Chung","Jangho Kim"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Hardware-Friendly Token-Group Activation Quantization for Low-Bit Mamba Super-Resolution","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16261","noteUrl":"/review?paper=review-neurips-2026-155635","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155638","title":"Can We Model the Artifacts Explicitly? Disentangle Artifacts via Pairwise Edit Relations for Image Manipulation Localization","authors":["Xuekang Zhu","Kaiwen Feng","Ruifeng Wang","Xiwen Wang","Xiaochen Ma","Bo Du","Changjiang Jiang","Chenfan Qu","Songyu Ye","Jian liu","Ji-Zhe, Knight Zhou"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Equal marginals do not remove semantic information in pair relations","plainExplanation":"Equal marginals do not remove semantic information in pair relations. The negative mixture does not enforce the claimed image marginals.","primaryUrl":"https://arxiv.org/html/2610.07916","noteUrl":"/review?paper=review-neurips-2026-155638","collection":"neurips-2026","severity":"major","issueTypes":["proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155640","title":"EvoOptiGraph: Weakness-Driven Coevolution via Graph-Based Structural Generation for Optimization Modeling","authors":["Qingcan Kang","Mingyang LIU","Xiaojin Fu","Shixiong Kai","Tao Zhong","Mingxuan Yuan"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"EvoOptiGraph: Weakness-Driven Coevolution via Graph-Based Structural Generation for Optimization Modeling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.26578","noteUrl":"/review?paper=review-neurips-2026-155640","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155641","title":"MemCoRe: Recovering Evidence from Progressively Compressed Factual Knowledge for Agent Memory","authors":["Zhenyuan Zhang","Jia Xianzhang","Zhiqin Yang","Zhenbo Song","Wei Xue","Sirui Han","Yike Guo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Raw-retaining memory already contains the target’s Shannon information","plainExplanation":"Section 2.1 interprets unorganized but raw-retaining memory as having low \\(I(M,Q;T)\\), and organization as increasing that Shannon relevance information. Raw-retaining memory already determines query-selected evidence, so retrieval difficulty cannot be identified with low Shannon relevance information in that full memory.","primaryUrl":"https://arxiv.org/html/2602.07885","noteUrl":"/review?paper=review-neurips-2026-155641","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155642","title":"What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion","authors":["Zhengrong Yue","Taihang Hu","Mengting Chen","Haiyu Zhang","Zihao Pan","Tao Liu","Zikang Wang","Jinsong Lan","Xiaoyong Zhu","Bo Zheng","Yali Wang"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.07915","noteUrl":"/review?paper=review-neurips-2026-155642","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155643","title":"Budgeted Multi-Source Counterfactual Annotation for Off-Policy Evaluation","authors":["Biao Xiang","Ali Eshragh","Yuexing Li","Kai Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Budgeted Multi-Source Counterfactual Annotation for Off-Policy Evaluation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10974","noteUrl":"/review?paper=review-neurips-2026-155643","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155645","title":"Smoothed Score Queries and the Complexity of Sampling","authors":["Jingbo Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Smoothed Score Queries and the Complexity of Sampling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27769","noteUrl":"/review?paper=review-neurips-2026-155645","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155646","title":"B-CALM: Bias-Limited Bayesian Borrowing for RCT-Anchored Treatment Effects under Covariate Mismatch","authors":["Amir Asiaee","Samhita Pal"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"B-CALM: Bias-Limited Bayesian Borrowing for RCT-Anchored Treatment Effects under Covariate Mismatch","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.04036","noteUrl":"/review?paper=review-neurips-2026-155646","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155647","title":"Pause and Reflect: Conformal Aggregation for Chain-of-Thought Reasoning","authors":["Yu Gu","Zijun Yu","Vahid Partovi Nia","Masoud Asgharian"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Strict and inclusive survival functions differ at voting atoms","plainExplanation":"Let half the prompts always be correct with confidence \\(1\\); on the other half, two iid paths choose two wrong labels fairly. Wrong confidence is \\(1\\) with probability \\(1/2\\), otherwise \\(1/2\\). At threshold \\(1/2\\), strict survival has \\(H=3/4\\) and correct mass \\(G=1/2\\), whereas inclusive empirical \\(H\\) tends to \\(1\\). The accuracy-ratio limit changes. Products over support points strictly below the threshold likewise retain its atom. The endpoint mismatch affects the ratio and convergence statement. The separate strict-loss conformal-risk theorem is not challenged.","primaryUrl":"https://arxiv.org/html/2605.14098","noteUrl":"/review?paper=review-neurips-2026-155647","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","interpretation_error","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155649","title":"Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction","authors":["Zhuofeng Li","Haoxiang Zhang","Cong Wei","Pan Lu","Ping Nie","Yuyang Bai","Shangbin Feng","Hangxiao Zhu","Ming Zhong","Yuyu Zhang","Jianwen Xie","Yejin Choi","James Zou","Jiawei Han","Wenhu Chen","Jimmy Lin","Dongfu Jiang","Yu Zhang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.05242","noteUrl":"/review?paper=review-neurips-2026-155649","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155651","title":"Computationally sufficient statistics for Ising models","authors":["Abhijith Jayakumar","Shreya Shukla","Marc Vuffray","Andrey Lokhov","Sidhant Misra"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The allowed error lets edges and nonedges meet at the same threshold","plainExplanation":"The Hoeffding exponent uses the wrong variable range. The allowed error lets edges and nonedges meet at the same threshold. Neighbor-coupling changes can offset a field displacement in the loss.","primaryUrl":"https://arxiv.org/html/2602.12449","noteUrl":"/review?paper=review-neurips-2026-155651","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155652","title":"Combating Data Laundering in LLM Training","authors":["Muxing Li","Zesheng Ye","Sharon Li","Feng Liu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Combating Data Laundering in LLM Training","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.01904","noteUrl":"/review?paper=review-neurips-2026-155652","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155655","title":"From Retrieval to Reasoning: Agentic Mechanism Prediction from Cell Painting Profiles","authors":["Jiayuan Chen","Botao Yu","Tianyu Liu","Thai-Hoang Pham","Meng Wu","Ping Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"From Retrieval to Reasoning: Agentic Mechanism Prediction from Cell Painting Profiles","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.36406","noteUrl":"/review?paper=review-neurips-2026-155655","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155656","title":"Aligning Forest and Trees in Images & Long Captions for Visually Grounded Understanding","authors":["Byeongju Woo","Zilin Wang","Byeonghyun Pak","Sangwoo Mo","Stella X. Yu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Aligning Forest and Trees in Images & Long Captions for Visually Grounded Understanding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.02977","noteUrl":"/review?paper=review-neurips-2026-155656","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155658","title":"Local Guidance, Global Impact: Gaussian-Reshaped Trust Region Unlocks Behavior Transitions","authors":["Bingxu Liu","Jiashun Liu","Johan Obando Ceron","Hao Wang","Runze Liu","Pablo Samuel Castro","Aaron Courville","Ling Pan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Gaussian damping does not control stale-reference importance weights","plainExplanation":"Section 4 attributes inherently bounded variance to Gaussian tails in the mixture regularizer sampled under behavior policy \\(\\mu\\). The Gaussian damps current/reference ratios but leaves reference/behavior importance weights uncontrolled, allowing divergent sampled second moments despite bounded population regularization.","primaryUrl":"https://arxiv.org/html/2606.03382","noteUrl":"/review?paper=review-neurips-2026-155658","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155660","title":"The Interplay of Data Structure and Imbalance in the Learning Dynamics of Diffusion Models","authors":["Flavio Nicoletti","Chenxiao Ma","Enrico Ventura","Luca Saglietti","Stefano Sarao Mannelli"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Class weights alone do not determine separate speciation times","plainExplanation":"Opposite extensive centroids produce one shared spectral direction, so class weights alone do not imply the separate transition hierarchy of Result 1. Its stated scaling is consistent with the orthogonal geometry used in Figure 1. Additional corrections concern the reverse SDE and score-error KL factor, smoothing and centering in the Gaussian-equivalent model, and the normalization of the self-consistency, train-error and spectral-edge expressions. The paper already supplies several of the needed identities; those local inconsistencies are distinguished from the unsupported general claims.","primaryUrl":"https://arxiv.org/html/2605.06367","noteUrl":"/review?paper=review-neurips-2026-155660","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","formula_error","incorrect_bound","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-155662","title":"Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents","authors":["Yeonjun In","Wonjoong Kim","Sangwu Park","Kanghoon Yoon","Chanyoung Park"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25535","noteUrl":"/review?paper=review-neurips-2026-155662","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155664","title":"DinoComplete: 3D Shape Completion with Distilled Semantic Priors and State Space Models","authors":["Furkan Mert Algan","Eckehard Steinbach"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DinoComplete: 3D Shape Completion with Distilled Semantic Priors and State Space Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.26949","noteUrl":"/review?paper=review-neurips-2026-155664","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155665","title":"UniFunc3D: Unified Active Spatial-Temporal Grounding for 3D Affordance Segmentation","authors":["Jiaying Lin","Dan Xu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"UniFunc3D: Unified Active Spatial-Temporal Grounding for 3D Affordance Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.23478","noteUrl":"/review?paper=review-neurips-2026-155665","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155666","title":"Diffusion-warm sampling of the XY model enables fast thermalization at scale","authors":["Sehmimul Hoque","Roger Melko","Pooya Ronagh"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The printed energy sign disagrees with the ferromagnetic sampler","plainExplanation":"On an even periodic \\(4\\times4\\) lattice, alignment gives energy \\(32\\) and a checkerboard gives \\(-32\\) under the printed sign, so the target favors antialignment. A single-spin reflection from alignment has printed energy change \\(-8\\), but the ferromagnetic cluster boundary probabilities give forward/reverse ratio \\(e^{-8\\beta}\\), opposite the required \\(e^{8\\beta}\\). This is a manuscript target/kernel sign inconsistency, not a verified simulation-code defect.","primaryUrl":"https://arxiv.org/html/2606.30773","noteUrl":"/review?paper=review-neurips-2026-155666","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155667","title":"CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction","authors":["Bing He","Jingnan Gao","Yunuo Chen","Zhengxue Cheng","Rong Xie","Li Song","Wenjun Zhang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Flat image intensity can hide a genuine depth discontinuity","plainExplanation":"Flat image intensity can hide a genuine depth discontinuity.","primaryUrl":"https://arxiv.org/html/2610.13177","noteUrl":"/review?paper=review-neurips-2026-155667","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155669","title":"Distillation of Foundation Models for Time-dependent PDEs","authors":["Daniel Musekamp","Boshra Ariguib","Andrei Manolache","Mathias Niepert"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Distillation of Foundation Models for Time-dependent PDEs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.11937","noteUrl":"/review?paper=review-neurips-2026-155669","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155671","title":"Training Language Models via Neural Cellular Automata","authors":["Dan Lee","Seungwook Han","Akarsh Kumar","Pulkit Agrawal"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Training Language Models via Neural Cellular Automata","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.10055","noteUrl":"/review?paper=review-neurips-2026-155671","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155672","title":"Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models","authors":["Huoren Yang","Jianchao Zhao","Yusong Hu","Qiguan Ou","Yuyang Gao","Wei Ke","Yuhang He","SongLin Dong","Zhiheng Ma","Yihong Gong"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11809","noteUrl":"/review?paper=review-neurips-2026-155672","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155674","title":"SAM 3D Animal: Promptable Animal 3D Reconstruction from Images in the Wild","authors":["Xuyi Hu","Jin Lyu","Jiuming Liu","Yebin Liu","Silvia Zuffi","Liang An","Stefan Goetz"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"SAM 3D Animal: Promptable Animal 3D Reconstruction from Images in the Wild","plainExplanation":"No issue was found in the checked claims. The erroneous softmax expression is absent from the current manuscript. Equation (2) now names a decoder block containing self-attention, image cross-attention, and a feed-forward network.","primaryUrl":"https://arxiv.org/abs/2605.07604v2","noteUrl":"/review?paper=review-neurips-2026-155674","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155676","title":"Hydra-X: Native Unified Multimodal Models with Holistic Visual Tokenizers","authors":["Guozhen Zhang","Xuerui Qiu","Yutao Cui","Tianhui Song","Changlin Li","Junzhe Li","Tao Huang","Xiao Zhang","Yang Li","Jianbing Wu","Miles Yang","Zhao Zhong","Liefeng Bo","Limin Wang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Hydra-X: Native Unified Multimodal Models with Holistic Visual Tokenizers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.13289","noteUrl":"/review?paper=review-neurips-2026-155676","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155678","title":"Delightful Distributed Policy Gradient","authors":["Ian Osband"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Delightful Distributed Policy Gradient","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.20521","noteUrl":"/review?paper=review-neurips-2026-155678","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155684","title":"Perfect Parallelization in Mini-Batch SGD with Classical Momentum Acceleration","authors":["Sachin Garg","Michal Derezinski"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The step-size factor in the condition number is inverted","plainExplanation":"The step-size factor in the condition number is inverted. Successive transition products share a random matrix. An admissible point lies below the asserted infimum bound.","primaryUrl":"https://arxiv.org/html/2605.18609","noteUrl":"/review?paper=review-neurips-2026-155684","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155686","title":"SOLAR: AI-Powered Speed-of-Light Performance Analysis","authors":["Qijing Huang","Sana Damani","Zhifan Ye","Athinagoras Skiadopoulos","Siva Kumar Sastry Hari","Jason Clemons","Sahil Modi","Jingquan Wang","Aditya Kane","Edward C Lin","Humphrey Shi","Christos Kozyrakis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Iteration shape alone does not determine exact arithmetic cost for extended einsums","plainExplanation":"Section 3.2.1 assigns one MAC per iteration-space point and calls the shape-only formula exact for extended einsums, including non-sum reductions and general scalar operations. The same iteration shape can contain different scalar operations or reduction types, so subscript dimensions alone cannot determine exact general MAC cost.","primaryUrl":"https://arxiv.org/html/2606.26383v2","noteUrl":"/review?paper=review-neurips-2026-155686","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155687","title":"A simple model of co-emergence of grid and place fields","authors":["Zhaoze Wang","Genela Morris","Dori Derdikman","Pratik Chaudhari","Vijay Balasubramanian"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Post-root clipping cannot make a negative variance real","plainExplanation":"Softplus is applied inside the recurrent sum, not to the newly updated membrane potential. Sufficiently inhibitory row sums therefore make that new potential negative from zero initialization. The receipt constructs a full \\(2048\\)-unit Dale-sign matrix with the stated excitatory/inhibitory proportions and spectral radius one for which \\(\\sigma_n h+\\varepsilon<0\\) at the allowed parameters. Clipping after the square root cannot define this real-valued update. The printed update has a real-domain gap on an open set of allowed parameters. No actual training NaN was reproduced.","primaryUrl":"https://arxiv.org/html/2605.21356","noteUrl":"/review?paper=review-neurips-2026-155687","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155689","title":"On the Nonlinearity of Learning Rate Scaling for LLM Training","authors":["ZAIWEN YANG","Huaqing Zhang","Jing Xu","Jingzhao Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"On the Nonlinearity of Learning Rate Scaling for LLM Training","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.29158","noteUrl":"/review?paper=review-neurips-2026-155689","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155690","title":"Making Open-Source Text LLM Watermarks Durable Against Merging","authors":["Luisa Scharff","Thibaud Gloaguen","Robin Staab","Martin Vechev"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The SLERP comparison parameter is not an exact linear weight fraction","plainExplanation":"SLERP uses sine coefficients rather than linear convex weights. The exact direct-base coefficient therefore does not generally equal the product of the nominal interpolation parameters.","primaryUrl":"https://arxiv.org/html/2607.20435","noteUrl":"/review?paper=review-neurips-2026-155690","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-155691","title":"How are linear representations learned? Exact solutions to the dynamics of abstraction","authors":["William W Yang","Peter E Latham","Andrew Saxe"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"How are linear representations learned? Exact solutions to the dynamics of abstraction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.08843","noteUrl":"/review?paper=review-neurips-2026-155691","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155694","title":"Sparse Reward Subsystem in Large Language Models","authors":["Guowei Xu","Mert Yuksekgonul","James Zou"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Sparse Reward Subsystem in Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.00986","noteUrl":"/review?paper=review-neurips-2026-155694","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155696","title":"Learning Better Certified Models from Empirically-Robust Teachers","authors":["Alessandro De Palma"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The intermediate interval recurrence halves the negative-weight contribution","plainExplanation":"For weight \\(W=-1\\), bias \\(3\\), and an input ReLU range \\([1,2]\\), the exact output range is \\([1,2]\\). The printed recurrence returns \\([2,2.5]\\). A final linear margin \\(h-1.5\\) would then receive lower bound \\(0.5\\), although attainable margin is \\(-0.5\\). Negative-weight multiplication must exchange endpoints without this halving. This refutes the displayed bound, not the library-based certificates used by an uninspected implementation.","primaryUrl":"https://arxiv.org/html/2602.02626","noteUrl":"/review?paper=review-neurips-2026-155696","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155697","title":"GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization","authors":["Pengyue Jia","Derong Xu","Yingyi Zhang","Xiaopeng Li","Wenlin Zhang","Yi Wen","Liu lan","Yuanshao Zhu","Xuetao Wei","Wanyu Wang","Xiangyu Zhao"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.24376","noteUrl":"/review?paper=review-neurips-2026-155697","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155699","title":"Supervised Distributional Reduction via Optimal Transport and Dependence Maximization","authors":["Sai-Aakash Ramesh","Archit Sood","Andrew Corbett","Tim Dodwell"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The stated kernel penalty gives a different normal equation","plainExplanation":"The gradient of half the residual norm plus the stated penalty is \\(K^\\top(KL-TZ)+\\lambda KL\\). For symmetric positive-definite \\(K\\), canceling \\(K\\) gives \\(L=(K+\\lambda I)^{-1}TZ\\). The printed equation instead contains the \\(\\lambda L\\) contribution of a different penalty. Symmetry does not make \\(K\\) the identity. This is a manuscript objective/solution mismatch; no implementation or empirical invalidity is inferred.","primaryUrl":"https://arxiv.org/html/2605.27619","noteUrl":"/review?paper=review-neurips-2026-155699","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155701","title":"Explaining and Preventing Alignment Collapse in Iterative RLHF","authors":["Etienne Gauthier","Francis Bach","Michael Jordan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Compact constrained optima need not satisfy the unconstrained sensitivity formula","plainExplanation":"Take a single prompt, \\(Y\\in\\{0,1\\}\\), \\(P_\\theta(Y=1)=\\theta\\in[1/4,3/4]\\), reward \\(r_\\phi=\\phi\\), parameter domain \\([0,1]\\), utility \\(2+Y\\), and squared loss \\((\\phi-(2+Y))^2/2\\). The loss is smooth and one-strongly convex; its unconstrained minimizer is \\(2+\\theta\\), so its constrained minimizer is always \\(1\\), with derivative zero. The printed inverse-Hessian formula instead gives derivative one, using a stationarity equation that fails at the boundary. At the symmetric reference policy, the KL derivative is zero, so the total-gradient claim is affected too. Compactness and strong convexity give uniqueness, not interiority. The boundary issue affects the cited lemma and related proposition/theorem uses, not every interior-instance derivation.","primaryUrl":"https://arxiv.org/html/2605.04266","noteUrl":"/review?paper=review-neurips-2026-155701","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155702","title":"Expanding Flow Maps","authors":["Sophia Tang","Pranam Chatterjee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Marginal flow identities do not establish posterior recovery","plainExplanation":"Different couplings can have the same endpoint marginal. Disintegration uniqueness requires the same joint law, not just the same marginal.","primaryUrl":"https://arxiv.org/html/2607.21585","noteUrl":"/review?paper=review-neurips-2026-155702","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155705","title":"LeAct: Learning to Reason from Expert Actions","authors":["Ziran Yang","Chengshuai Shi","Raj Ghugare","Benjamin Eysenbach","Karthik Narasimhan","Chi Jin"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Hard positive selection does not ensure bound improvement","plainExplanation":"Optimizing the selected trace objective can lower expert-action likelihood even with exact scores and an identically constant proposal ratio.","primaryUrl":"https://arxiv.org/html/2607.21856","noteUrl":"/review?paper=review-neurips-2026-155705","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","mathematical_error"],"classifications":["major"]},{"slug":"review-neurips-2026-155706","title":"Interactive 4D Volumetric Liquid Forecasting under Moving-Solid Interaction","authors":["Jongwon Ryu","Sunjae Yoon","Junyeong Kim"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Interactive 4D Volumetric Liquid Forecasting under Moving-Solid Interaction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.16167","noteUrl":"/review?paper=review-neurips-2026-155706","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155709","title":"Distribution Shift in Missing Data Imputation: A Risk-Based Perspective and Importance-Weighted Correction under MAR","authors":["Luke Shannon","Song Liu","Katarzyna Reluga"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Observed-entry preservation makes the proposed training error identically zero","plainExplanation":"Appendix A.2 replaces missing-target and observed-target response laws using ordinary patternwise MAR, after viewing non-target observations as common predictors. Ordinary patternwise MAR does not imply equality of target response laws under a common non-target predictor representation, leaving the risk-transfer step unjustified. The inference from generic patternwise MAR is insufficient even after target masking repairs the separate representation problem. Missingness driven only by fully observed covariates is a valid narrower setting.","primaryUrl":"https://arxiv.org/html/2602.06713","noteUrl":"/review?paper=review-neurips-2026-155709","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155713","title":"Learning Augmented Exact Exponential Algorithms","authors":["Tatiana Belova","Yuriy Dementiev","Danil Sagunov"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning Augmented Exact Exponential Algorithms","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.18807","noteUrl":"/review?paper=review-neurips-2026-155713","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155714","title":"PDE-PFN: Prior-Data Fitted Neural PDE Solver","authors":["Jaehyeon Park","Mingu Kang","Dongseok Lee","Woojin Cho","Kookjin Lee","Anthony Gruber","Youngjoon Hong","Noseong Park"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"PDE-PFN: Prior-Data Fitted Neural PDE Solver","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.13334","noteUrl":"/review?paper=review-neurips-2026-155714","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155719","title":"MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents","authors":["Haozhen Zhang","Quanyu Long","Jianzhu Bao","Tao Feng","Weizhi Zhang","Haodong Yue","Wenya Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.02474","noteUrl":"/review?paper=review-neurips-2026-155719","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155720","title":"Diff3R: Feed-forward 3D Gaussian Splatting with Uncertainty Aware Differentiable Optimization","authors":["Yueh-Cheng Liu","Jozef Hladký","Matthias Niessner","Angela Dai"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Photometric null directions retain initialization at every positive proximal weight","plainExplanation":"Photometric null directions retain initialization at every positive proximal weight.","primaryUrl":"https://arxiv.org/html/2604.01030","noteUrl":"/review?paper=review-neurips-2026-155720","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155722","title":"Graph Distance Based on Cause-Effect Estimands with Latents","authors":["Zhufeng Li","Niki Kilbertus"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The fixing criterion must exclude proper descendants in the district","plainExplanation":"The verifier has an unhandled empty-denominator branch. The fixing criterion must exclude proper descendants in the district.","primaryUrl":"https://arxiv.org/html/2510.25037","noteUrl":"/review?paper=review-neurips-2026-155722","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155723","title":"Disentangling Continuous-Time Latent Dynamics: Identifiability of Latent SDEs via Diffusion Shifts","authors":["Yuanyuan Wang","Wenjie Wang","Haoxuan Li","Mingming Gong","Kun Zhang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Disentangling Continuous-Time Latent Dynamics: Identifiability of Latent SDEs via Diffusion Shifts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.28228","noteUrl":"/review?paper=review-neurips-2026-155723","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155726","title":"Learning Unbiased Permutations via Flow Matching","authors":["Yimeng Min","Carla Gomes"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Learning Unbiased Permutations via Flow Matching","plainExplanation":"No issue was found in the checked claims. In the current formula, \\(J\\) is added after projecting the noise. The projected term has zero row and column sums and \\(J\\) has unit sums, so \\(\\widetilde X_0\\in\\mathcal B_a\\). A tangent velocity preserves those sums, and \\(P^*-\\widetilde X_0\\) is itself tangent. The earlier offset error and its associated loss floor are removed.","primaryUrl":"https://arxiv.org/html/2605.16755","noteUrl":"/review?paper=review-neurips-2026-155726","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155729","title":"Prompting Diffusion Models for Zero-Shot Instance Segmentation","authors":["İrem Z Alagöz","Nils Morbitzer","Andrea Ramazzina","Nassir Navab","Federico Tombari","Stefano Gasperini"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Prompting Diffusion Models for Zero-Shot Instance Segmentation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.22660","noteUrl":"/review?paper=review-neurips-2026-155729","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155730","title":"Online Learning via Learned Latent Bayesian Tracking","authors":["Guy Gerson","Tomer Raviv","Nir Shlezinger","Tirza Routtenberg","Osvaldo Simeone"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Online Learning via Learned Latent Bayesian Tracking","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31559","noteUrl":"/review?paper=review-neurips-2026-155730","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155733","title":"When Trackers Fail: VLM-Guided Verification and Recovery for Robust Video Object Segmentation","authors":["Valay Mahesh Bundele","Susmit Agrawal","Mehran Hosseinzadeh","The Nam Nguyen","Hendrik PA Lensch"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"When Trackers Fail: VLM-Guided Verification and Recovery for Robust Video Object Segmentation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.15929","noteUrl":"/review?paper=review-neurips-2026-155733","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155734","title":"Approximation of Maximally Monotone Operators : A Graph Convergence Perspective","authors":["Takashi Furuya","Yury Korolev","Takaharu Yaguchi"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Approximation of Maximally Monotone Operators : A Graph Convergence Perspective","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12301v1","noteUrl":"/review?paper=review-neurips-2026-155734","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155736","title":"Point4D: Long-range 4D Motion Reconstruction","authors":["Minsik Jeon","Jay Karhade","Deva Ramanan","Shubham Tulsiani"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Point4D: Long-range 4D Motion Reconstruction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.09145","noteUrl":"/review?paper=review-neurips-2026-155736","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155737","title":"Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models","authors":["Meng Zheng","Samhita Marri","Anwesa Choudhuri","Benjamin Planche","Zhongpai Gao","Van N Nguyen","Terrence Chen","Girish Chowdhary","Ziyan Wu"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The repulsive sign is corrected; the guidance equivalence still drops a factor","plainExplanation":"Changing only the failure coefficient cannot reproduce the missing factor on the success output.","primaryUrl":"https://arxiv.org/html/2605.08434","noteUrl":"/review?paper=review-neurips-2026-155737","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155738","title":"Disentanglement as Identifiable Pushforward Factorisation","authors":["Carl Allen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Disentanglement as Identifiable Pushforward Factorisation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2410.22559","noteUrl":"/review?paper=review-neurips-2026-155738","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155741","title":"Generative OOD-regularized Model-based Policy Optimization","authors":["Aysin Tumay","Jiahe Huang","Elise Jortberg","Rose Yu"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Generative OOD-regularized Model-based Policy Optimization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.24405","noteUrl":"/review?paper=review-neurips-2026-155741","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155742","title":"Structural Causal Bottleneck Models","authors":["Simon Bing","Jonas Wahl","Jakob Runge"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"A direct bottleneck does not determine the total intervention effect","plainExplanation":"Equality of the direct bottleneck can coexist with a changed mediated path. The factored model therefore does not justify the stated equality of total interventional distributions; controlled direct effects or conditions on every relevant outgoing path supply valid alternatives.","primaryUrl":"https://arxiv.org/html/2603.08682","noteUrl":"/review?paper=review-neurips-2026-155742","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155743","title":"Steering Visual Generation in Unified Multimodal Models with Understanding Supervision","authors":["Zeyu Liu","Zanlin Ni","Yang Yue","Cheng Da","Huan Yang","Di ZHANG","Kun Gai","Gao Huang"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Steering Visual Generation in Unified Multimodal Models with Understanding Supervision","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05781","noteUrl":"/review?paper=review-neurips-2026-155743","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155747","title":"Aligning Few-Step Generative Model via Amortizing Sample-Based Variational Inference","authors":["Jaewoo Lee","Hyeongyu Kang","Dohyun Kim","Kyuil Sim","Woocheol Shin","Minsu Kim","Taeyoung Yun","Jeongjae Lee","Sanghyeok Choi","Tabitha Edith Lee","Jong Chul Ye","Jinkyoo Park"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Aligning Few-Step Generative Model via Amortizing Sample-Based Variational Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26552","noteUrl":"/review?paper=review-neurips-2026-155747","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155749","title":"Visual Text Compression as Measure Transport","authors":["Lv Tang","Tianyi Zheng","Yang Liu","Bo Li","Xingyu Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Visual Text Compression as Measure Transport","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06708","noteUrl":"/review?paper=review-neurips-2026-155749","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155756","title":"ARK: A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge","authors":["Yijie Lin","Guofeng Ding","Haochen Zhou","Haobin Li","Mouxing Yang","Xi Peng"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"ARK: A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.09839","noteUrl":"/review?paper=review-neurips-2026-155756","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155757","title":"Deep Gaussian Processes on Directed Acyclic Graphs","authors":["Federico L Perlino","Oliver Hamelijnck","Adam Johansen","Theodoros Damoulas"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The dense-sampler complexity omits a conditioning term","plainExplanation":"With \\(M=1\\) and \\(K=J^2\\), the summary is of order \\(J^4\\), but the stated conditioning step is of order \\(J^5\\). Thus the extra term is not uniformly absorbed when all three parameters may vary. The complexity summary is incomplete for the described dense sampler; this is not a lower bound for every possible implementation.","primaryUrl":"https://arxiv.org/html/2607.09645","noteUrl":"/review?paper=review-neurips-2026-155757","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155759","title":"UFO: A Unified Flow-Oriented Framework for Robust Continual Graph Learning","authors":["Danhui Zhang","Zhe Wang","Qing Qing","Jiarui Liu","Wentao Gao","Ziqi Xu","Mingliang Hou","Xikun Zhang","Renqiang Luo"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"UFO: A Unified Flow-Oriented Framework for Robust Continual Graph Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.09862","noteUrl":"/review?paper=review-neurips-2026-155759","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155764","title":"Viverra: Text-to-Code with Guarantees","authors":["Haoze Wu","Rocky Klopfenstein","Keith Farkas","Nina Narodytska"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Facts checked on a bound-reduced program need a transfer proof","plainExplanation":"Let the original accept \\(0\\leq n\\leq9\\), run \\(n\\) loop iterations only in mode one, then test candidate fact \\(n\\leq5\\). Replacing the cap by \\(5\\) makes that fact true on every accepted reduced run. Yet original input \\(n=6\\) in mode zero reaches the false fact with zero iterations. Restricting the original guarantee to bounded executions does not exclude it. Reducing an input/loop cap can change admissible traces even when the cap is not copied into data. The missing transfer concerns original-program annotations in public v1. The fixed-program dependency theorems remain sound; no released annotation or attributed implementation was run.","primaryUrl":"https://arxiv.org/html/2605.14972","noteUrl":"/review?paper=review-neurips-2026-155764","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155765","title":"FlowHOI: Flow-based Semantics-Grounded Generation of Hand-Object Interactions for Dexterous Robot Manipulation","authors":["Huajian Zeng","Lingyun Chen","Jiaqi Yang","Yuantai Zhang","Fan Shi","Peidong Liu","Xingxing Zuo"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The contact-feature dimensions and fusion complexity need consistent definitions","plainExplanation":"Full cross-attention on two length-\\(N_s\\) sequences cannot implement the asserted linear fusion cost.","primaryUrl":"https://arxiv.org/html/2602.13444","noteUrl":"/review?paper=review-neurips-2026-155765","collection":"neurips-2026","severity":"repairable","issueTypes":["definition_ambiguity","invalid_inference"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155767","title":"Multi-Environment POMDPs with Finite-Horizon Objectives","authors":["Léonard Brice","Filip Cano","Krishnendu Chatterjee","Thomas Henzinger","Stefanie Muroya"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Conditional continuation values need a different rational grid","plainExplanation":"The decision algorithm uses a rational continuation-value grid that is not preserved by conditioning on observations. Its payoff range also counts one fewer reward than the trajectory definition and recursion.","primaryUrl":"https://arxiv.org/html/2605.07537","noteUrl":"/review?paper=review-neurips-2026-155767","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155768","title":"Geometric Factual Recall in Transformers","authors":["Shauli Ravfogel","Gilad Yehudai","Joan Bruna","Alberto Bietti"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The stated dimension regime does not eliminate the local bound","plainExplanation":"The exact dimension regime following the lower bound does not follow from its formula. The proof also uses an unsupported tree extraction and leaf count, and an overly strong random-sign tail exponent.","primaryUrl":"https://arxiv.org/html/2605.12426","noteUrl":"/review?paper=review-neurips-2026-155768","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155769","title":"Actor-Accelerated Policy Dual Averaging for Reinforcement Learning in Continuous Action Spaces","authors":["Ji Gao","Caleb Ju","Guanghui Lan","Zhaohui Tong"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Euclidean curvature does not give the claimed Bregman gap bound","plainExplanation":"The general-prox convergence argument uses a Bregman gap estimate that Euclidean strong convexity does not provide. Its initial summand also divides by a curvature parameter defined to be zero.","primaryUrl":"https://arxiv.org/html/2603.10199","noteUrl":"/review?paper=review-neurips-2026-155769","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155770","title":"$\\delta$-Mem: : Efficient Online Memory For Large Language Models","authors":["Jingdi Lei","Di Zhang","Junxian Li","Weida Wang","Kaixuan Fan","Xiang Liu","Qihan Liu","Xiaoteng Ma","Baian Chen","Soujanya Poria"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"$\\delta$-Mem: : Efficient Online Memory For Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.12357","noteUrl":"/review?paper=review-neurips-2026-155770","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155772","title":"Next-Latent Prediction Transformers Learn Compact World Models","authors":["Jayden Teoh","Manan Tomar","Kwangjun Ahn","Edward Hu","Tim Pearce","Pratyusha Sharma","Akshay Krishnamurthy","Riashat Islam","Alex Lamb","John Langford"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Next-Latent Prediction Transformers Learn Compact World Models","plainExplanation":"No mathematical or statistical issue was identified in the reviewed text.","primaryUrl":"https://arxiv.org/html/2511.05963","noteUrl":"/review?paper=review-neurips-2026-155772","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155774","title":"Black-Box Followers, White-Box Leaders: Partial Zeroth-Order Methods for MPECs","authors":["Miriam Fischer","Dario Paccagnan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Black-Box Followers, White-Box Leaders: Partial Zeroth-Order Methods for MPECs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/pdf/2605.16685v1","noteUrl":"/review?paper=review-neurips-2026-155774","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155775","title":"Inference-time Alignment via Sparse Junction Steering","authors":["Runyi Hu","Jie Zhang","Shiqian Zhao","Jiale Meng","Jiwei Li","Jason Zeng","Ming Wu","Michael Heinrich","Yonggang Wen","Tianwei Zhang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The skipped-step bound uses the wrong KL direction","plainExplanation":"The skipped-steering guarantee reverses the KL divergence in its local objective gap. The value-noise proposition also omits the dependence of its leading coefficient on action probabilities and the noise covariance.","primaryUrl":"https://arxiv.org/html/2602.21215","noteUrl":"/review?paper=review-neurips-2026-155775","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155776","title":"Drift Flow Matching","authors":["Chenrui Ma","Xi Xiao","Lin Zhao","Tianyang Wang","Ferdinando Fioretto","Yanning Shen"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Drift Flow Matching","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.17244","noteUrl":"/review?paper=review-neurips-2026-155776","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155777","title":"The Laplacian Keyboard: Beyond the Linear Span","authors":["Siddarth Chandrasekar","Marlos C. Machado"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Arithmetic symmetrization of a Markov kernel need not be stochastic","plainExplanation":"Appendix B.1.2 claims that \\(W=(P+P^{\\top})/2\\) is doubly stochastic for any MDP transition kernel and therefore \\(I-W\\) is a positive-semidefinite Laplacian with constant zero eigenvector. Arithmetic symmetrization preserves symmetry but not row sums of a general Markov kernel, so subtracting it from identity need not give a positive-semidefinite Laplacian. The arbitrary-MDP Laplacian extension fails. The spectral approximation argument under a valid positive-semidefinite graph Laplacian remains separate.","primaryUrl":"https://arxiv.org/html/2602.07730","noteUrl":"/review?paper=review-neurips-2026-155777","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155779","title":"TabDLM: Free-Form Tabular Data Generation via Joint Numerical–Language Diffusion","authors":["Donghong Cai","Jiarui Feng","Yanbo Wang","Da Zheng","Yixin Chen","Muhan Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"TabDLM: Free-Form Tabular Data Generation via Joint Numerical–Language Diffusion","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.22586","noteUrl":"/review?paper=review-neurips-2026-155779","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155781","title":"DiRotQ: Rotation-Aware Quantization for 4-bit Diffusion Transformers","authors":["Sayeh Sharify","Mahsa Salmani","Hesham Mostafa"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DiRotQ: Rotation-Aware Quantization for 4-bit Diffusion Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16732","noteUrl":"/review?paper=review-neurips-2026-155781","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155786","title":"WorldVLA: A Unified Vision-Language-Action and World Model","authors":["Jun CEN","Siteng Huang","Yuqian Yuan","Kehan Li","Hangjie Yuan","Chaohui Yu","Bohan Hou","Yuming Jiang","Jiayan Guo","Xin Li","Hao Luo","Fan Wang","Deli Zhao","Hao Chen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"WorldVLA: A Unified Vision-Language-Action and World Model","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2511.17502","noteUrl":"/review?paper=review-neurips-2026-155786","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155787","title":"From Noise to Diversity: Random Embedding Injection in LLM Reasoning","authors":["Heejun Kim","Seungpil Lee","Jewon Yeom","Jaewon Sok","Seonghyeon Park","Jeongjae Park","Taesup Kim","Sundong Kim"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"From Noise to Diversity: Random Embedding Injection in LLM Reasoning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11936","noteUrl":"/review?paper=review-neurips-2026-155787","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155789","title":"Conditioning Gaussian Processes on Almost Anything","authors":["Henry Moss","Lachlan Astfalck","Tom Cowperthwaite","Colin Doumont","Samuel Willis","Christopher Nemeth","Philipp Hennig","Andrew Zammit-Mangion"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Histogram widths belong inside the convolution sum","plainExplanation":"The convolved density is \\(q(x)=\\sum_k(p_k/w_k)[\\Phi((u_k-x)/\\nu)-\\Phi((\\ell_k-x)/\\nu)]\\), where \\(w_k=u_k-\\ell_k\\). Taking the logarithm leaves each width inside this sum. For a single width-two bin, the printed outer factor multiplies the score derivative by \\(1/2\\); correct normalization only subtracts the constant \\(\\log2\\). This is a local histogram log-density formula error; unit-width bins hide it. No implementation is judged.","primaryUrl":"https://arxiv.org/html/2605.21041","noteUrl":"/review?paper=review-neurips-2026-155789","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155790","title":"Root Cause Analysis of Measurement and Mechanistic Anomalies","authors":["Hendrik Suhr","David Kaltenpoth","Jilles Vreeken"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Bernoulli code-length approximation has the wrong sign","plainExplanation":"Appendix A.4.3 calls a negative log-binomial expression a positive Bernoulli sequence coding cost and subtracts its change from optimistic gains. The positive Bernoulli coding cost has leading term positive log-binomial, so the printed negative sign reverses its sparsity penalty in the optimistic gains.","primaryUrl":"https://arxiv.org/html/2601.23026","noteUrl":"/review?paper=review-neurips-2026-155790","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155793","title":"Nautilus: From One Prompt to Plug-and-Play Robot Learning","authors":["Yufeng Jin","Jianfei Guo","Xiaogang Jia","Yu Deng","Steven Li","Han Liu","Weiran Liao","Vignesh Prasad","Mathias Franzius","Gerhard Neumann","Georgia Chalvatzaki"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Nautilus: From One Prompt to Plug-and-Play Robot Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.11665","noteUrl":"/review?paper=review-neurips-2026-155793","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155796","title":"MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models","authors":["Yingzi Ma","Zhengyue Zhao","Xiaogeng Liu","Minhui Xue","Yue Zhao","Chaowei Xiao"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.04027","noteUrl":"/review?paper=review-neurips-2026-155796","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155798","title":"DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising","authors":["Paolo Giaretta","Zeyang Li","Navid Azizan"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising","plainExplanation":"No mathematical or statistical issue was identified in the reviewed text.","primaryUrl":"https://arxiv.org/html/2606.15359","noteUrl":"/review?paper=review-neurips-2026-155798","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155799","title":"Optimal Convergence Analysis of DDPM for General Distributions","authors":["Yuchen Jiao","Yuchen Zhou","Gen Li"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The score-error assumption omits the first reverse step","plainExplanation":"The DDPM theorem measures score errors only through $T-1$, while its first reverse update uses $s_T$. An arbitrary bias in that unmeasured score shifts the Gaussian output and makes its KL divergence unbounded. Including the terminal score in the weighted error average supplies the missing estimate without changing the sampler.","primaryUrl":"https://arxiv.org/html/2510.27562","noteUrl":"/review?paper=review-neurips-2026-155799","collection":"neurips-2026","severity":"repairable","issueTypes":["incorrect_bound","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155800","title":"Filtered Conformal Ellipsoids for Graph-Native Time Series","authors":["Yannick Limmer"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"A local density bound does not control threshold covariance at the boundary","plainExplanation":"The local-density covariance theorem drops the buffer needed to control thresholds at the boundary of its density interval. A small contracting perturbation can move an uncontrolled external density spike across that boundary and violate the stated envelope.","primaryUrl":"https://arxiv.org/html/2606.17014","noteUrl":"/review?paper=review-neurips-2026-155800","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-155802","title":"TED: Text-Axis Evidence Decomposition for Prompted Anomaly Localization","authors":["JINYOUNG KIM","Geonho Kim","GiJeong Park","Geonu Lee","YoungJoon Yoo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TED: Text-Axis Evidence Decomposition for Prompted Anomaly Localization","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.39033","noteUrl":"/review?paper=review-neurips-2026-155802","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155803","title":"When and How Much to Imagine: Adaptive Test-Time Scaling with World Models for Visual Spatial Reasoning","authors":["Shoubin Yu","Yue Zhang","Zun Wang","Jaehong Yoon","Huaxiu Yao","Mingyu Ding","Mohit Bansal"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"When and How Much to Imagine: Adaptive Test-Time Scaling with World Models for Visual Spatial Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.08236","noteUrl":"/review?paper=review-neurips-2026-155803","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155806","title":"LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing","authors":["Jiawei Hao","Zhiwei Hao","Jianyuan Guo","Li Shen","Yong Luo","Han Hu","Dan Zeng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.12645","noteUrl":"/review?paper=review-neurips-2026-155806","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155807","title":"Model Capacity Determines Grokking through Competing Memorisation and Generalisation Speeds","authors":["Yiding Song","Hanming Ye"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The memorized-bit statistic has no universal zero lower bound","plainExplanation":"The defined memorized-bit statistic can be negative and unbounded below, contrary to its stated range. Its upper bound remains valid.","primaryUrl":"https://arxiv.org/html/2605.09724","noteUrl":"/review?paper=review-neurips-2026-155807","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155808","title":"AI-Mediated Communication Can Steer Collective Opinion","authors":["Stratis Tsirtsis","Kai Rawal","Chris Russell","Brent Mittelstadt","Sandra Wachter"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"AI-Mediated Communication Can Steer Collective Opinion","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16245","noteUrl":"/review?paper=review-neurips-2026-155808","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155810","title":"Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR","authors":["Utkarsh Tyagi","Xingang Guo","MohammadHossein Rezaei","Daniel George","Anas Mahmoud","Jackson Lee","Bing Liu","Yunzhong He"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.20164","noteUrl":"/review?paper=review-neurips-2026-155810","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155811","title":"Z0-Inf: Zeroth Order Approximation for Data Influence","authors":["Narine Kokhlikyan","Diego Garcia-Olano","Kamalika Chaudhuri","Saeed Mahloujifar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Directional secants estimate projected gradients","plainExplanation":"Checkpoint secants recover directional projections of the loss gradient; shrinking the displacements does not restore missing directions, and even isotropic directions need calibration. The variance bound must retain its checkpoint-distance and normalization factor. Smaller corrections concern the appendix’s derivative convention, dependence on other samples through training, a copied polynomial exercise, and displayed notation already partly corrected in Algorithm 1.","primaryUrl":"https://arxiv.org/html/2510.11832","noteUrl":"/review?paper=review-neurips-2026-155811","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","incorrect_bound","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155813","title":"Towards Understanding the Power and Limits of the Muon Optimizer: A River-Valley Perspective","authors":["Tianqi Shen","Jinji Yang","Runze SHI","Jianhao Ma","Jiaye Teng","Ziye Ma"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Gradient-momentum sign agreement can be negative","plainExplanation":"The reduced dynamics omit a rank-dependent time factor, and the sign-agreement lemma incorrectly guarantees positivity. The proposed residual Lyapunov function can increase in a realizable model. An illustrative covariance also misses the theorem's block-curvature assumptions, while the stated stepsize endpoint gives nonincrease rather than strict descent.","primaryUrl":"https://arxiv.org/html/2606.21514","noteUrl":"/review?paper=review-neurips-2026-155813","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155814","title":"Retain-Neutral Surrogates for Min-Max Unlearning","authors":["Junhao Cai","Dohun Kim","Dowon Kim","Sung Il Choi","Chengjun Jin","Juhyun Park","Changhee Joo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Retain-Neutral Surrogates for Min-Max Unlearning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05871","noteUrl":"/review?paper=review-neurips-2026-155814","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155815","title":"PAMod: Modeling Cyclical Shifts via Phase-Amplitude Modulation for Non-stationary Time Series Forecasting","authors":["Yingbo Zhou","Yutong Ye","Shuhao Li","Rui Qian","Qiang Huang","Lemao Liu","Li Sun","Dejing Dou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The target-risk bound omits sampling and loss assumptions","plainExplanation":"The domain-adaptation bound substitutes empirical source risk for population risk without an estimation term and uses squared loss as though it obeyed a metric triangle inequality.","primaryUrl":"https://arxiv.org/html/2605.00466","noteUrl":"/review?paper=review-neurips-2026-155815","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155816","title":"Latent Process Generator Matching","authors":["Lukas Billera","Hedwig Nora Nordlinder","Ben Murrell"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Pointwise integrability does not ensure a finite time-integrated loss","plainExplanation":"A singleton zero-generator process admits target \\((1-t)^{-1/2}\\), scalar predictor \\(\\theta\\) and squared divergence. At \\(\\theta=0\\), every loss for \\(t<1\\) is finite, but the integrated loss is \\(\\int_0^1dt/(1-t)=\\infty\\), so its classical gradient is undefined. Even finite integrated losses would still need differentiability and justified interchange. The finite-loss matching identity is not refuted. The stated hypotheses alone do not justify its classical-gradient version.","primaryUrl":"https://arxiv.org/html/2605.20547","noteUrl":"/review?paper=review-neurips-2026-155816","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155817","title":"GMOS: Grounding Moving Object Segmentation in 3D Space and Time","authors":["Junyu Xie","Tengda Han","Weidi Xie","Andrew Zisserman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The centered proposer window requires future frames","plainExplanation":"Two frames in that centered window occur after the timestamp being predicted, with maximum lookahead \\(0.25\\) seconds. Cross-frame geometric attention can use them. Forward propagation after calling this proposer therefore does not make the target-time output causal. The strict target-time causality description conflicts with the input specification. A buffered streaming implementation remains possible; unseen online code is not judged.","primaryUrl":"https://arxiv.org/html/2605.30352","noteUrl":"/review?paper=review-neurips-2026-155817","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155818","title":"RICE-PO: Turning Retrieval Interactions into Credit Signals for Reasoning Agents","authors":["mingchen li","Hansi Zeng","Zhuo Qian","Jiatan Huang","Sunjae Kwon","Hamed Zamani","Hong Yu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"RICE-PO: Turning Retrieval Interactions into Credit Signals for Reasoning Agents","plainExplanation":"No issue was found in the checked claims. Under this functional reading, the conditional-expectation calculation is correct. An example with fresh summary noise given the same reasoning action is outside that reading; it does not establish a false proposition as stated.","primaryUrl":"https://arxiv.org/html/2605.26352","noteUrl":"/review?paper=review-neurips-2026-155818","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155819","title":"FiLoRA: Focus-and-Ignore LoRA for Controllable Feature Reliance","authors":["Hyunsuk Chung","Caren Han","Seungyeon Ji","Jinwoo Kim","Eun-Jung Holden","Kyungreem Han"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FiLoRA: Focus-and-Ignore LoRA for Controllable Feature Reliance","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.02060","noteUrl":"/review?paper=review-neurips-2026-155819","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155820","title":"ReflectMT: Internalizing Reflection for Efficient and High-Quality Machine Translation","authors":["Kunquan Li","Yingxue Zhang","Zhibin Lan","Fandong Meng","Jinsong Su"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ReflectMT: Internalizing Reflection for Efficient and High-Quality Machine Translation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2604.19144","noteUrl":"/review?paper=review-neurips-2026-155820","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155821","title":"From Topology to Retrieval: Decoding Embedding Spaces with Unified Signatures","authors":["Florian Rottach","William Rudman","Bastian Rieck","Harrisen Scells","Carsten Eickhoff"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"The displayed asymptotic magnitude dimension is zero for a finite distinct-point cloud","plainExplanation":"Appendix A.9 applies an infinite-scale magnitude-growth limit to the finite point cloud defined in Appendix A.8, while interpreting nonzero computed values as that descriptor. Finite-cloud magnitude approaches the point count at infinite scale, forcing the displayed logarithmic growth dimension to zero rather than the reported positive finite-scale values. The displayed finite-cloud limit cannot yield the reported positive descriptor values. This does not determine whether the library’s different computed statistic is useful.","primaryUrl":"https://arxiv.org/html/2511.22150","noteUrl":"/review?paper=review-neurips-2026-155821","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155822","title":"Can Circuit Alignment Predict OOD Generalization?","authors":["Ayan Banerjee","ABHRA CHAUDHURI","Josep Llados","Umapada Pal","Anjan Dutta"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The random-walk kernel does not separate all circuit topologies","plainExplanation":"The random-walk kernel can assign similarity one to nonisomorphic, functionally distinct circuits. The robustness gap does not imply maximum CAS, and the proposed extraction-noise correction assumes a universal positive bias that kernel curvature does not provide.","primaryUrl":"https://arxiv.org/html/2609.31996","noteUrl":"/review?paper=review-neurips-2026-155822","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-155824","title":"How Diffusion Models Memorize","authors":["Juyeop Kim","Songkuk Kim","Jong-Seok Lee"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"DDPM reverse innovations prevent reuse of one terminal residual","plainExplanation":"Perfect clean-target prediction uses the current residual \\(\\epsilon_t=(x_t-\\sqrt{\\bar\\alpha_t}x)/\\sqrt{1-\\bar\\alpha_t}\\). Substitution gives \\(x_{t-1}=A_tx_t+B_tx+\\sigma_tz_t\\), where \\(A_t=\\sqrt{\\alpha_t}(1-\\bar\\alpha_{t-1})/(1-\\bar\\alpha_t)\\). Hence the initial-state coefficient at step \\(s\\) is \\(\\sqrt{\\bar\\alpha_T/\\bar\\alpha_s}(1-\\bar\\alpha_s)/(1-\\bar\\alpha_T)\\), and innovation variance accumulates as \\(V_{t-1}=A_t^2V_t+\\sigma_t^2\\). Equation (27) instead freezes terminal noise and retains only the current innovation. The Appendix E multistep DDPM decomposition is wrong. The first step, the current-residual one-step identity, DDIM and empirical correlations are not disproved.","primaryUrl":"https://arxiv.org/html/2509.25705","noteUrl":"/review?paper=review-neurips-2026-155824","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155827","title":"Predictive Geometry of Hidden Trajectories in Transformers","authors":["Timur Mudarisov","Mikhail Burtsev","Tatiana Petrova","Radu State"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Predictive Geometry of Hidden Trajectories in Transformers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.37717","noteUrl":"/review?paper=review-neurips-2026-155827","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155829","title":"Learning POMDP World Models from Observations with Language-Model Priors","authors":["Valentin Six","Frederik Panse","Mathis Fajeau","Lancelot Da Costa","Mridul Sharma","Alfonso Amayuelas","Tim Xiao","David Hyland","Philipp Hennig","Bernhard Schölkopf"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Learning POMDP World Models from Observations with Language-Model Priors","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13740","noteUrl":"/review?paper=review-neurips-2026-155829","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155833","title":"Explicit Geometric Chain-of-Thought for Vision-Language-Action in Autonomous Driving","authors":["Xingtai Gui","Yucheng Zhou","Dongqian Guo","jiahao gong","Feiyang Tan","Jianbing Shen"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Explicit Geometric Chain-of-Thought for Vision-Language-Action in Autonomous Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.10390","noteUrl":"/review?paper=review-neurips-2026-155833","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155839","title":"Learn where to Click from Yourself: On-Policy Self-Distillation for GUI Grounding","authors":["Yan Zhang","Daiqing Wu","Huawen Shen","Can Ma","Yu Zhou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A nearly one-hot teacher does not turn reverse KL into hard-label cross-entropy","plainExplanation":"The introduction identifies the displayed student-to-teacher reverse-KL objective with hard-label SFT when the teacher becomes nearly one-hot. Reverse KL penalizes student mass on vanishing teacher probabilities and can diverge, whereas hard-label cross-entropy remains finite; shared minimizers do not make these objectives equivalent. The explanatory equivalence is false. Shared limiting minimizers and observed near-SFT behavior do not establish equivalence of objectives or updates.","primaryUrl":"https://arxiv.org/html/2605.00642","noteUrl":"/review?paper=review-neurips-2026-155839","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155843","title":"When Policy Entropy Constraint Fails: Preserving Diversity in Flow-based RLHF via Perceptual Entropy","authors":["Xiaofeng Tan","Jun Liu","Bin-Bin Gao","Yuanting Fan","Xi Jiang","Chengjie Wang","Hongsong Wang","Feng Zheng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"When Policy Entropy Constraint Fails: Preserving Diversity in Flow-based RLHF via Perceptual Entropy","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12112","noteUrl":"/review?paper=review-neurips-2026-155843","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155844","title":"RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO","authors":["Yanzuo Lu","Ronglai Zuo","Jiankang Deng"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The printed transition range needs its terminal member removed","plainExplanation":"The range includes \\(k_i=K-1\\), whose destination is \\(\\tau_K=0\\). It therefore fails to implement the exclusion stated in the adjacent paragraph.","primaryUrl":"https://arxiv.org/html/2605.15190","noteUrl":"/review?paper=review-neurips-2026-155844","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-155847","title":"MoSE3: Learning World-Space SE(3) at Every Pixel","authors":["Jiahuan(Joanna) Cheng","Zhiyi Li","Tian Xia","Ruojin Cai","Yilun Du","Qianqian Wang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MoSE3: Learning World-Space SE(3) at Every Pixel","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.03716","noteUrl":"/review?paper=review-neurips-2026-155847","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155848","title":"LoRaQ: Optimized Low Rank Approximation for 4-bit Quantization","authors":["Yann Bouquet","Alireza Khodamoradi","Sophie Y Shen","Kristof Denolf","Mathieu Salzmann"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The optimized correction and inference branch use opposite signs","plainExplanation":"Zero optimized residual means \\(Q_1(W+D)-W-D=0\\), hence \\(W=Q_1(W+D)-D\\). The figure instead reconstructs \\(Q_1(W-D)+D\\) with the same factors. Choose \\(Q_1\\) as nearest-integer rounding on the sixteen levels from \\(-8\\) through \\(7\\), and \\(Q_2\\) as nearest-tenth rounding on the sixteen levels from \\(-0.8\\) through \\(0.7\\). An example with \\(W=0.3,L=-0.6,R=0.5\\), so \\(D=-0.3\\), has zero objective but figure reconstruction \\(0.7\\), not \\(0.3\\). Both conventions can work with a conversion. This is a printed optimizer/inference mismatch, not a verified code defect.","primaryUrl":"https://arxiv.org/html/2604.18117","noteUrl":"/review?paper=review-neurips-2026-155848","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155851","title":"Re-evaluating Confidence Remasking in Masked Diffusion Language Models","authors":["Stipe Frković","Metod Jazbec","Dan Zhang","Christian Andersson Naesseth","Ilija Bogunovic","Eric Nalisnick"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Re-evaluating Confidence Remasking in Masked Diffusion Language Models","plainExplanation":"Full shadow attention retains an extra-key normalization discrepancy even in one layer, and the stated ELBO needs its masking-schedule qualification. The empirical pass-at-k interpretation is outside the mathematical review.","primaryUrl":"https://arxiv.org/html/2606.12232","noteUrl":"/review?paper=review-neurips-2026-155851","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-155852","title":"XDecomposer: Learning Prior-Free Set Decomposition for Multiphase X-ray Diffraction","authors":["Hanyu Gao","Bin Cao","YUNYUE SU","Tong-yi Zhang","Qiang Liu"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"XDecomposer: Learning Prior-Free Set Decomposition for Multiphase X-ray Diffraction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05866","noteUrl":"/review?paper=review-neurips-2026-155852","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155853","title":"A Reduction from Delayed to Immediate Feedback for Online Convex Optimization with Improved Guarantees","authors":["Alexander Ryabchenko","Idan Attias","Dan Roy"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The arbitrary-norm squared prox violates the stated regret bound","plainExplanation":"The arbitrary-norm squared-prox regret claim fails for the l1 norm. The printed FTRL update also differs from the regularization analyzed in the appendix, and the round-dependent bandit smoothing schedule starts at an undefined zero radius.","primaryUrl":"https://arxiv.org/html/2602.02634","noteUrl":"/review?paper=review-neurips-2026-155853","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155854","title":"PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference","authors":["Chenglei Yu","Chuanrui Wang","Bangyan Liao","Tailin Wu"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.18587","noteUrl":"/review?paper=review-neurips-2026-155854","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155855","title":"CI4A: Semantic Component Interfaces for Agents Empowering Web Automation","authors":["Zhi Qiu","Jiazheng Sun","Chenxiao Xia","Jun Zheng","Xin Peng"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"CI4A: Semantic Component Interfaces for Agents Empowering Web Automation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.14790","noteUrl":"/review?paper=review-neurips-2026-155855","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155856","title":"CoreQ: Learning-Free Mismatch Correction and Successive Rounding for Quantization","authors":["Seohyeon Cha","Huancheng Chen","Dongjun Kim","Haoran Zhang","Kevin Chan","Gustavo De Veciana","Haris Vikalo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Damped reachable and residual components need not be orthogonal","plainExplanation":"Algorithm 1 adds covariance damping but retains the orthogonal projection geometry used to justify its closed-form correction coefficient. Damping destroys the orthogonality of reachable and residual components, so the undamped norm-ratio correction need not minimize the stated displacement. Damping invalidates this exact geometric minimizer argument. The separately assumed population projection theorem is not contradicted.","primaryUrl":"https://arxiv.org/html/2602.05902","noteUrl":"/review?paper=review-neurips-2026-155856","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155860","title":"Tools as Continuous Flow for Evolving Agentic Reasoning","authors":["Tairan Huang","Siyu Shang","Qiang Chen","Xiu Su","Yi Chen"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"An integrated velocity residual needs the maximum Gronwall weight","plainExplanation":"Equation (37) applies \\((e^K-1)/K\\) to an integrated residual, although that constant averages the Gronwall kernel over time. An integrated residual can concentrate where the Gronwall kernel is largest, so the kernel’s average weight is too small to bound that error. The printed constant is too small for its integrated-error premise. Replacing it can preserve the convergence order while changing the constants.","primaryUrl":"https://arxiv.org/html/2605.07339","noteUrl":"/review?paper=review-neurips-2026-155860","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155863","title":"Revealing Epistemic Uncertainty in MLLMs via Causal-Invariant Masking","authors":["Haoyang Luo","Linwei Tao","Jie Gui","Xinghao Chen","Chang Xu","Jianyuan Guo","Minjing Dong"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Cancelling the explicit causal energy does not remove its effect on KL","plainExplanation":"The cancellation identity is correct, but it does not establish the additional invariance claim. The KL divergence averages with respect to \\(Q\\), whose probabilities depend on the causal energy.","primaryUrl":"https://arxiv.org/html/2610.02887","noteUrl":"/review?paper=review-neurips-2026-155863","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","invalid_inference","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155864","title":"No Pose, No Problem in 4D: Feed-Forward Dynamic Gaussians from Unposed Multi-View Videos","authors":["Matteo Balice","Yanik Künzi","Chenyangguang Zhang","Matteo Matteucci","Marc Pollefeys","Sunghwan Hong"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"No Pose, No Problem in 4D: Feed-Forward Dynamic Gaussians from Unposed Multi-View Videos","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.22190","noteUrl":"/review?paper=review-neurips-2026-155864","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155865","title":"The Mechanism of Weak-to-Strong Generalization: Feature Elicitation from Latent Knowledge","authors":["Ryoya Awano","Taiji Suzuki"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Evenness does not remove the transformed teacher’s mean term","plainExplanation":"For standard Gaussian \\(X\\), unit directions \\(\\theta,w\\), \\(U=\\theta^\\top X\\), \\(V=w^\\top X\\) and \\(\\rho=\\theta^\\top w\\), integration by parts gives \\(\\mathbb E[F(U)VX]=\\rho\\mathbb E[F(U)(U^2-1)]\\theta+\\mathbb E[F(U)]w\\). Evenness kills the first Hermite coefficient, not the mean \\(\\mathbb E[F]\\). The transformed teacher generally has a nonzero mean, omitted by the display. The receipt checks a compatible activation with all required Hermite coefficients nonzero. The displayed decomposition is false. A projected or rate-level repair remains a separate proof obligation; no full theorem counterexample is asserted.","primaryUrl":"https://arxiv.org/html/2605.12908","noteUrl":"/review?paper=review-neurips-2026-155865","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155866","title":"AdaST: Adaptive Coupling for Spatial-Temporal Forecasting","authors":["Zhenyu Lei","Chenghao Liu","Yushun Dong","Ryan Qi Wang","Jundong Li"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The positive AR coefficient sampler is usually unstable","plainExplanation":"The proposed AR(12) coefficient sampler does not ensure stability. The adaptive gate also applies a fractional power to unrestricted cosine means, leaving it undefined on allowed negative inputs.","primaryUrl":"https://arxiv.org/html/2609.36119","noteUrl":"/review?paper=review-neurips-2026-155866","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155871","title":"Learning to Learn from Multimodal Experience","authors":["Xingyu Sui","Weixiang Zhao","Yongxin Tang","Yanyan Zhao","Yang Wu","Dandan Tu","Bing Qin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Learning to Learn from Multimodal Experience","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16857","noteUrl":"/review?paper=review-neurips-2026-155871","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155873","title":"Measuring Black-Box Confidence via Reasoning Trajectories: Geometry, Coverage, and Verbalization","authors":["Marc Boubnovski Martell","Josefa Stoisser","Kaspar Märtens","Jialin Yu","Robert Kitchen","Philip Torr","Jesper Ferkinghoff-Borg"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"A top-versus-rest posterior has a competitor-shape term","plainExplanation":"If option probabilities are proportional to \\(e^{-bD_j}\\), the nearest-option logit is \\(b(D_2-D_1)-\\log\\sum_{j>1}e^{-b(D_j-D_2)}\\). The second term varies with the competitors and is not one common intercept. Feasible distance vectors with relative probabilities \\((4,2,1,1)\\) and \\((8,2,2,1)\\) have corrections \\(\\log2\\) and \\(\\log2.5\\). The receipt constructs exact gold counts and unit-vector geometry yielding all-option slope \\(-10\\) but binary raw slope about \\(6.78072\\), inconsistent with the asserted algebraic equality after standardization. The exact slope identity fails. Pairwise odds, empirical AUCs, approximate binary scoring and observed interval overlap are not refuted.","primaryUrl":"https://arxiv.org/html/2605.06308","noteUrl":"/review?paper=review-neurips-2026-155873","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155875","title":"Controllable Road Marking Generation","authors":["Zhiyu Joey Cai","Yufan Zhang","Ruichen Tan","Zengxiang Lei","Satish Ukkusuri"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Controllable Road Marking Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2610.05771","noteUrl":"/review?paper=review-neurips-2026-155875","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155877","title":"The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans","authors":["Odysseas S Chlapanis","Orfeas Menis Mastromichalakis","Christos H Papadimitriou"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Panel reliability and a noiseless model ceiling use different scales","plainExplanation":"Under \\(Y_A=T+e_A\\), \\(Y_B=T+e_B\\), with independent errors of variance \\(v\\) and signal variance \\(s\\), panel correlation is \\(R=s/(s+v)\\). A noiseless predictor \\(T\\) instead correlates with a panel by \\(\\sqrt R\\), strictly larger for \\(0<R<1\\). At \\(R=0.5\\), these values are \\(0.5\\) and about \\(0.7071\\). Averaging square roots also differs from taking the square root of an averaged reliability. The scale and near-ceiling interpretation need correction. Exact empirical corrected ceilings require the underlying per-dimension values; no replacement empirical number is invented.","primaryUrl":"https://arxiv.org/html/2605.08837","noteUrl":"/review?paper=review-neurips-2026-155877","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155879","title":"Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet","authors":["Zeyuan Hu","Noah Brenowitz","Akshay Subramaniam","Jaideep Pathak","Tao Ge","Mohammad S Abbas","Suman Ravuri","Karthik Kashinath","Noel Keen","Naser Mahfouz","Peter Caldwell","Mike Pritchard"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Global centering does not imply orthogonality to patchwise constants","plainExplanation":"For a simple dominant eigenpair, the coefficient is \\((w^\\top x)/(w^\\top v)\\), where \\(w\\) is a left eigenvector. Take \\(A=\\begin{pmatrix}3&2\\\\0&1\\end{pmatrix}\\), \\(x=(1,2)\\). The unit right eigenvector is \\(e_1\\), but the left eigenvector is \\((1,1)\\), so the coefficients are \\(1\\) versus \\(3\\). Directly, \\(A^jx=(3^{j+1}-2,2)\\), whose normalized first-coordinate limit is \\(3\\). The coefficient and eigenspace explanation are oversimplified. Qualitative growth when the spectral radius exceeds one remains possible.","primaryUrl":"https://arxiv.org/html/2606.31248","noteUrl":"/review?paper=review-neurips-2026-155879","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"consistency-test-degenerate-null","title":"Consistency as a Testable Property: Statistical Methods to Evaluate AI Agent Reliability","authors":["Harsh Raj","Niranjan Orkat","Suvrorup Mukherjee","Aritra Guha","Cheryl Flynn","Subhabrata Majumdar"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Perfect consistency makes the Gaussian statistic undefined","plainExplanation":"The perfect-consistency null makes the proposed studentized statistic undefined. Separately, the general variance formula omits the first Hoeffding projection, and one minus the normalized alignment kernel does not upper-bound edit distance. These corrections preserve the similarity measurements and the valid unnormalized soft-min calculation.","primaryUrl":"https://arxiv.org/abs/2605.10516v1","collection":"neurips-2026","severity":"major","issueTypes":["calibration_error","false_statement","interpretation_error"],"classifications":["major"]},{"slug":"review-neurips-2026-155881","title":"End-to-End Training for Unified Tokenization and Latent Denoising","authors":["Shivam Duggal","Xingjian Bai","Zongze Wu","Richard Zhang","Eli Shechtman","Antonio Torralba","Phillip Isola","Bill Freeman"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"End-to-End Training for Unified Tokenization and Latent Denoising","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.22283","noteUrl":"/review?paper=review-neurips-2026-155881","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155883","title":"Revisiting the Generic Transformer: Deconstructing a Strong Baseline for Time Series Foundation Models","authors":["Yunshi Wen","Wesley M Gifford","Chandra Reddy","Lam Nguyen","Jayant Kalagnanam","Anak Agung Julius"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Revisiting the Generic Transformer: Deconstructing a Strong Baseline for Time Series Foundation Models","plainExplanation":"Without either time reduction, the expression is \\(|x_t-\\mu|\\) at each visible time and \\(0/0\\) at a masked time. It is not the scalar standard deviation described by the surrounding text.","primaryUrl":"https://arxiv.org/html/2602.06909","noteUrl":"/review?paper=review-neurips-2026-155883","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["algorithm_specification","formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-155885","title":"TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents","authors":["Jingyu Sun","Yuyang Xue","Mingyang Li","Zhengtao Yao","Jiachen Li","Yang Cui","Wenhao Cai","Haozhe Liu","Fangying Wang","Magdalene K Montgomery","Syed M Baker","Hongpeng Zhou"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.00967","noteUrl":"/review?paper=review-neurips-2026-155885","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155886","title":"DiscoLoop: Looping Discrete Embeddings and Continuous Hidden States for Multi-hop Reasoning","authors":["Hengyu Fu","Tianyu Guo","Zixuan Wang","Hanlin Zhu","Jason Lee","Jiantao Jiao","Stuart J Russell","Song Mei"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"DiscoLoop: Looping Discrete Embeddings and Continuous Hidden States for Multi-hop Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2607.00341","noteUrl":"/review?paper=review-neurips-2026-155886","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155887","title":"Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents","authors":["Ziyan Liu","Zhezheng Hao","Yeqiu Chen","Hong Wang","JingRen Hou","Ruiyi Ding","Yongkang Yang","Wence Ji","Wei Xia","Feng Liu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The specified response sampler has zero conditional mutual information","plainExplanation":"The stated factorization is \\(P(S,M)\\pi_{\\mathrm{LLM}}(Y\\mid M,q)\\). It makes \\(S\\) and \\(Y\\) conditionally independent given \\(M,q\\), hence \\(I(Y;S\\mid M,q)=0\\). Unconditional response relevance can still hold through noisy memory, so that assumption does not undo the Markov relation. Response entropy need not track changes in \\(H(S\\mid M)\\). The entropy identity itself is algebraically valid, but its residual-state interpretation is vacuous under this sampler. Empirical entropy–performance correlation is not refuted.","primaryUrl":"https://arxiv.org/html/2605.30159","noteUrl":"/review?paper=review-neurips-2026-155887","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155890","title":"Sketch Then Paint: Hierarchical Reinforcement Learning for Diffusion Multi-Modal Large Language Models","authors":["Siqi Luo","Jianghan Shen","Yi Xin","Huayu Zheng","Haoxing Chen","Yan Tai","Yue Li","Junjun He","Yihao Liu","Guangtao Zhai","Yuewen Cao","Xiaohong Liu"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Future contamination is only one way the remasked context can differ","plainExplanation":"Condition on the target being masked. With one past and one future token independently remasked at probability \\(1/2\\), full equality requires the past token visible and the future token masked, with probability \\(1/4\\). Mismatch is therefore \\(3/4\\), whereas the printed expression gives \\(1/2\\) and counts only visible future tokens. The event identity is wrong. The main conclusion that future contamination has positive probability survives.","primaryUrl":"https://arxiv.org/html/2605.16842","noteUrl":"/review?paper=review-neurips-2026-155890","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155891","title":"In-context learning to predict critical transitions in dynamical systems","authors":["Yunus Sevinchan","Juan Nathaniel","Kai Ueltzhöffer","Carla Roesch","Tobias Weber","Vaios Laschos","Hang Fan","Gregor Ramien","Johannes Haux","Pierre Gentine","Benjamin Herdeanu"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Pooling reused lead times does not give the claimed variance bound","plainExplanation":"The pooled-lead AUC formula is not a conservative uncertainty bound when lead times reuse samples. The appendix also misstates an invasion threshold, gives an amplitude normal form without the claimed stable high branch, and reverses the branch-drift sign in tracking-error dynamics.","primaryUrl":"https://arxiv.org/html/2605.12308","noteUrl":"/review?paper=review-neurips-2026-155891","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155893","title":"Accelerated last-iterate convergence of Extragradient via power-law stepsizes","authors":["Yue Wu","Weiqiang Zheng","Yang Cai","Haipeng Luo"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Correlated Pareto steps can violate the vector rate","plainExplanation":"Expectation of a modal log product does not commute with the maximum or sum over different modes.","primaryUrl":"https://arxiv.org/pdf/2606.01764v1","noteUrl":"/review?paper=review-neurips-2026-155893","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155894","title":"Fréchet Regression on the Bures-Wasserstein Manifold","authors":["Duc Toan Nguyen","Cesar Uribe"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Fréchet Regression on the Bures-Wasserstein Manifold","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.03566","noteUrl":"/review?paper=review-neurips-2026-155894","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155895","title":"ART for Diffusion Sampling: A Reinforcement Learning Approach to Timestep Scheduling","authors":["Yilie Huang","Wenpin Tang","XUNYU ZHOU"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The unconstrained HJB solution need not satisfy the terminal time constraint","plainExplanation":"Theorem 3.2 derives endpoint-constrained ART optimality from a classical solution of the displayed fixed-multiplier HJB equation. A classical fixed-multiplier HJB solution need not generate a policy satisfying the terminal time constraint, so it does not establish constrained optimality without admissibility and multiplier selection.","primaryUrl":"https://arxiv.org/html/2601.18681v3","noteUrl":"/review?paper=review-neurips-2026-155895","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155897","title":"Self-Compacting Language Model Agents","authors":["Tianjian Li","Jingyu (Jack) Zhang","Xi Wang","William Jurayj","Chuanyang Jin","Mehrdad Farajtabar","Eric Nalisnick","Daniel Khashabi"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Self-Compacting Language Model Agents","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.23525","noteUrl":"/review?paper=review-neurips-2026-155897","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155899","title":"TILT: Target-induced loss tilting under covariate shift","authors":["Kakei Yamamoto","Martin Wainwright"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Support and auxiliary-range assumptions need correction","plainExplanation":"On target-only support, this weight is zero for every positive regularization value rather than tending to one.","primaryUrl":"https://arxiv.org/pdf/2605.14280","noteUrl":"/review?paper=review-neurips-2026-155899","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155900","title":"MeshFIM: Local Low-Poly Mesh Editing via Fill-in-the-Middle Autoregressive Generation","authors":["Dingdong Yang","Jian Liu","Biwen Lei","Haohan Weng","Zhuo Chen","Song Guo","Hao Zhang","Ali Mahdavi Amiri","Chunchao Guo"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"MeshFIM: Local Low-Poly Mesh Editing via Fill-in-the-Middle Autoregressive Generation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08744","noteUrl":"/review?paper=review-neurips-2026-155900","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155901","title":"Strategic PAC Learnability via Geometric Definability","authors":["Yuval Filmus","Shay Moran","Elizaveta Nesterova","Nir Rosenfeld","Alexander Shlimovich"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Strategic PAC Learnability via Geometric Definability","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.13426","noteUrl":"/review?paper=review-neurips-2026-155901","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155905","title":"More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding","authors":["Noam Elata","Itay Lamprecht","Mikey Shechter","Daniel Ohayon","Itay Hubara","Daniel Soudry"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.04753","noteUrl":"/review?paper=review-neurips-2026-155905","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155906","title":"Squeezing Capacity from Multimodal Large Language Models for Subject-driven Generation","authors":["Shuhong Zheng","Aashish K Misraa","Kevin Li","Yu-Jhe Li","Igor Gilitschenski"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Squeezing Capacity from Multimodal Large Language Models for Subject-driven Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26111","noteUrl":"/review?paper=review-neurips-2026-155906","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155907","title":"GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation","authors":["Sixiang Chen","Zhaohu Xing","Tian Ye","Xinyu Geng","Yunlong Lin","Jianyu Lai","Xuanhua He","Fuxiang Zhai","Jialin Gao","Lei Zhu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.21605","noteUrl":"/review?paper=review-neurips-2026-155907","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155913","title":"No Model Required: Text Entropy Rate Filtering Prevents Iterative Fine-Tuning Collapse","authors":["Lewis Mitchell"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The arithmetic match-length estimator is not universally ergodic-consistent","plainExplanation":"The arithmetic match-length estimator is not consistent for every stationary ergodic source: a finite-mean regenerative process can force subsequences toward zero despite positive entropy. The rank-based explanations also overreach: compressed score scale need not destroy ranking, and correlation 0.997 does not ensure identical selected sets. A suffix-length description is off by one.","primaryUrl":"https://arxiv.org/html/2610.01493","noteUrl":"/review?paper=review-neurips-2026-155913","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155916","title":"Meta-Learning Preferences for Multilingual LLM Alignment","authors":["Jiaying Lin","Seongho Son","Nam P Tran","Long Tran-Thanh","Ilija Bogunovic","Debmalya Mandal"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Meta-Learning Preferences for Multilingual LLM Alignment","plainExplanation":"Driving an upper error bound below tolerance gives a sufficient budget, not a necessary hitting time. Subtracting lower bounds would not prove a bound on the difference either.","primaryUrl":"https://arxiv.org/html/2607.13315","noteUrl":"/review?paper=review-neurips-2026-155916","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155918","title":"RPC-GS: Gaussian Splatting with native RPC Rendering for Satellite Imagery","authors":["Valentin Wagner","Sebastian Bullinger","Christoph Bodensteiner","Michael Arens"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An inverse RPC pixel locus need not be a straight ENU ray","plainExplanation":"A degree-two RPC with unit denominators can use \\(r=\\lambda/10^{-6}-(h/100)^2\\), \\(c=\\phi\\). Pixel \\((0,0)\\) then has longitude \\(\\lambda=10^{-10}h^2\\), a curved height locus. Converting heights \\(0,50,100\\) to ENU at the equator leaves the middle point about \\(1.5913\\) meters off the endpoint chord after orthogonal projection. An orthogonal residual cannot be removed by any different scalar depth. The exact on-ray representation is false for admissible RPCs. Approximate depth ordering remains possible, and no implementation or reconstruction-quality defect is established.","primaryUrl":"https://arxiv.org/html/2606.06690","noteUrl":"/review?paper=review-neurips-2026-155918","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155919","title":"OmniGF: A Dual-Branch Vision-Language Framework for Unified Gaze Following","authors":["Qiaomu Miao","Haoyu Wu","Jingyi Xu","Minh Hoai","Dimitris Samaras"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"OmniGF: A Dual-Branch Vision-Language Framework for Unified Gaze Following","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.26399","noteUrl":"/review?paper=review-neurips-2026-155919","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155921","title":"Semiparametrically Efficient Inference for Kernel Measures of Noise Heterogeneity","authors":["Jakub Wornbard","Zikai Shen","Dimitri Meunier","Arthur Gretton"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"Semiparametrically Efficient Inference for Kernel Measures of Noise Heterogeneity","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.27526","noteUrl":"/review?paper=review-neurips-2026-155921","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155922","title":"Dynamic Delayed Tree Expansion For Improved Multi-Path Speculative Decoding","authors":["Rahul K Thomas","Teo Kitanovski","Micah Goldblum","Arka Pal"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Dynamic Delayed Tree Expansion For Improved Multi-Path Speculative Decoding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.16994","noteUrl":"/review?paper=review-neurips-2026-155922","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155923","title":"What Does the AI Doctor Value? Auditing Pluralism in the Clinical Ethics of Language Models","authors":["Payal Chandak","Victoria Alkin","David Wu","Maya Dagan","Taposh D Roy","Maria Clara Saad Menezes","Ayush Noori","Nirali Somia","John S Brownstein","Ran Balicer","Rebecca W Brendel","Noa Dagan","Isaac S Kohane","Gabriel A Brat"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"What Does the AI Doctor Value? Auditing Pluralism in the Clinical Ethics of Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.18738","noteUrl":"/review?paper=review-neurips-2026-155923","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155924","title":"Decision-Aware Proximal Bridge Learning for Optimal Treatment Selection","authors":["Tomas Garriga","Alejandro Almodóvar","Axel Brando","Gerard Sanz","Eduard Serrahima de Cambra","Juan Parras"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Decision-Aware Proximal Bridge Learning for Optimal Treatment Selection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.16989","noteUrl":"/review?paper=review-neurips-2026-155924","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155925","title":"Assessing Per-Sample Membership Inference Vulnerability without Retraining","authors":["Valentin Dorseuil","Jamal Atif","Olivier Cappé"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"A noncentral squared-normal likelihood ratio need not be monotone","plainExplanation":"The frozen-model loss likelihood ratio is not monotone for arbitrary noncentral targets, so universal lower-tail optimality fails. The appendices also contain two local Gaussian scaling errors, a deletion Newton step using the wrong objective, and an ordinary full-softmax Hessian inverse without a gauge convention.","primaryUrl":"https://arxiv.org/html/2602.15919","noteUrl":"/review?paper=review-neurips-2026-155925","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155927","title":"Reward Score Matching: Unifying Reward-based Fine-tuning for Flow and Diffusion Models","authors":["Jeongjae Lee","Jinho Chang","Jeongsol Kim","Jong Chul Ye"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Mean adjustment cannot realize a general tilted transition kernel","plainExplanation":"Gaussian mean shifts do not in general match the full soft-RL tilted kernel. The soft-value derivative also requires a reference-mean Jacobian that mean matching does not supply, and the native-flow weight contains an extra squared factor.","primaryUrl":"https://arxiv.org/html/2604.17415","noteUrl":"/review?paper=review-neurips-2026-155927","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155928","title":"Revisiting On-policy Adversarial Black-Box Distillation: Calibrating Groupwise Reward Geometry for Effective Advantage Construction","authors":["Xiao Cui","Mo Zhu","Yulei Qin","Yuze Wu","Wengang Zhou","Houqiang Li"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Revisiting On-policy Adversarial Black-Box Distillation: Calibrating Groupwise Reward Geometry for Effective Advantage Construction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.39757","noteUrl":"/review?paper=review-neurips-2026-155928","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155935","title":"The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility","authors":["David Pape","Jonathan Evertz","Lea Schönherr"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.19537","noteUrl":"/review?paper=review-neurips-2026-155935","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155936","title":"SWE-Protégé: Learning to Selectively Collaborate With an Expert Unlocks Small Language Models as Software Engineering Agents","authors":["Patrick Tser Jern Kon","Archana Pradeep","Ang Chen","Alex Ellis","Warren Hunt","Zijian Wang","John Yang","Samuel Thompson"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The formula applies the repeat threshold before any first activation","plainExplanation":"The operational rule requires \\(k_1=15\\) until an actual activation occurs, then \\(k_2=8\\). Streaks of lengths \\((1,8)\\) never reach \\(15\\), yet formula (4) charges the second streak at threshold \\(8\\). The first temporal streak is not necessarily a first triggering streak. This is a trigger-state formula/prose discrepancy; actual code was not inspected.","primaryUrl":"https://arxiv.org/html/2602.22124","noteUrl":"/review?paper=review-neurips-2026-155936","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155937","title":"Detecting Hidden ML Training With Zero-Overhead Telemetry","authors":["Robi Rahman","Sabiha Tajdari"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Detecting Hidden ML Training With Zero-Overhead Telemetry","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.19262","noteUrl":"/review?paper=review-neurips-2026-155937","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155938","title":"Vision-Language Grounding as Bidirectional Concept Correspondence","authors":["Jieyu Zhang","Ziqi Gao","Luke Zettlemoyer","Ranjay Krishna"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Vision-Language Grounding as Bidirectional Concept Correspondence","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2608.07886","noteUrl":"/review?paper=review-neurips-2026-155938","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155939","title":"Balancing Image Compression and Generation with Bootstrapped Tokenization","authors":["Haozhe Chi","Jinghan Li","Hao Jiang","Wu Sheng","Yi Ma","Jing Wang","Yadong Mu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Balancing Image Compression and Generation with Bootstrapped Tokenization","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.05552","noteUrl":"/review?paper=review-neurips-2026-155939","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155943","title":"Exact Gaussian Moment Matching for Residual Networks: a Second-Order Method","authors":["Simon Kuang","Xinfan Lin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Exact Gaussian Moment Matching for Residual Networks: a Second-Order Method","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.22307","noteUrl":"/review?paper=review-neurips-2026-155943","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155946","title":"Inexact Bregman Sparse Newton Method for Efficient Optimal Transport","authors":["Jianting Pan","Jian Li","Sirong Dai","Ming Yan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Thresholding can erase a row before its normalization","plainExplanation":"Algorithm 1 permits every nonnegative threshold, and Theorem 3.2 asserts a real positive-semidefinite Hessian with the stated kernel for all such thresholds. An allowed threshold can erase an entire nonexempt row, making normalization undefined before any Hessian kernel or spectral statement can apply.","primaryUrl":"https://arxiv.org/html/2603.07156","noteUrl":"/review?paper=review-neurips-2026-155946","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155947","title":"A Bitter Lesson for Data Filtering","authors":["Christopher Mohri","John Duchi","Tatsunori Hashimoto"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"A Bitter Lesson for Data Filtering","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.19407","noteUrl":"/review?paper=review-neurips-2026-155947","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155949","title":"UniVer: A Unified Perspective for Multi-step and Multi-draft Speculative Decoding","authors":["Yepeng Weng","Qiao Hu","Takehisa Yairi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The completion identity omits extensions of a short query","plainExplanation":"In a depth-one example, draft token \\(0\\) is accepted with probability \\(0.5/0.8\\), while draft token \\(1\\) is always accepted; accepted leaves append another target token. Thus outputs starting with \\(0\\) have length two and total mass \\(1/2\\). For the length-one query \\(0\\), the definition sums exact completed outputs that are prefixes of the query, yielding zero and missing all accepted extensions. The short-query quantifier and completion identity fail. The intended speculative sampler can still be target-exact.","primaryUrl":"https://arxiv.org/html/2605.04543","noteUrl":"/review?paper=review-neurips-2026-155949","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155952","title":"MindLoom: Composing Thought Modes for Frontier-Level Reasoning Data Synthesis","authors":["Haiyang Shen","Taian Guo","Xuanzhong Chen","Mugeng Liu","Sixiong Xie","Zhuofan Shi","Chongyang Pan","Siqi Zhong","Guoqing Wang","Ming Zhang","Yun Ma"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"MindLoom: Composing Thought Modes for Frontier-Level Reasoning Data Synthesis","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.21630","noteUrl":"/review?paper=review-neurips-2026-155952","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155954","title":"RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer","authors":["Shuo Lyu","Jinhao You","Zhuohang Lyu","Tanxuan Li","Zibo Zhao","Jiaxiang Hu","Kai Tang","Yichen Guo"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.18439","noteUrl":"/review?paper=review-neurips-2026-155954","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-155955","title":"Halt Fast! Early Stopping for Certified Robustness","authors":["Andrew Cullen","Paul Montague","Benjamin Rubinstein"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Halt Fast! Early Stopping for Certified Robustness","plainExplanation":"No mathematical or statistical issue was identified in the reviewed text.","primaryUrl":"https://arxiv.org/html/2606.27694","noteUrl":"/review?paper=review-neurips-2026-155955","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155956","title":"Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models","authors":["Senqiao Yang","Chengyao Wang","Yuxin Chen","Zixuan WANG","Longxiang Tang","Haokun GUI","Jinhui Ye","Changsheng Lu","Xiaoyang Wu","Mingkang Zhu","Pengguang Chen","Shu Liu","Zhuotao Tian","Hengshuang Zhao","Bei Yu","Jiaya Jia"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.27550","noteUrl":"/review?paper=review-neurips-2026-155956","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155958","title":"InTAct: Interval-based Task Activation Consolidation for Continual Learning","authors":["Patryk Krukowski","Jan Miksa","Piotr Helm","Jacek Tabor","Paweł Wawrzyński","Przemysław Spurek"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Current-input feature matching does not ensure historical feature-map invariance","plainExplanation":"The method claims that gated feature distillation preserves the historical input-to-feature map and hence historical predictions, while Equation (9) is evaluated on sampled current-task inputs. Feature matching on current-task samples does not control historical inputs, so it cannot supply the historical feature stability required for end-to-end prediction invariance.","primaryUrl":"https://arxiv.org/html/2511.17439","noteUrl":"/review?paper=review-neurips-2026-155958","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155960","title":"Conservative neural posterior estimation via distributionally robust training","authors":["William Laplante","Yuga Hikida","Charita Dellaporta","Francois-Xavier Briol","Ayush Bharti"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Marginal PIT uniformity does not imply conditional calibration","plainExplanation":"The calibration interpretation upgrades marginal PIT uniformity to almost-everywhere conditional equality. Observation-dependent deviations can cancel exactly in the marginal, so that implication needs a stronger definition.","primaryUrl":"https://arxiv.org/html/2605.28516","noteUrl":"/review?paper=review-neurips-2026-155960","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-155961","title":"Optimal and Efficient Contextual Combinatorial Semi-bandits with General Function Approximation","authors":["Hao Qin","Chicheng Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The log-barrier Frank–Wolfe loop leaves its domain","plainExplanation":"For action set \\(\\{(1,0),(0,1)\\}\\), both arms are covered, but each prescribed initial vertex has a zero coordinate, making the reciprocal gradient undefined. Even replacing initialization by \\((1/2,1/2)\\) does not suffice: the first step size \\(1\\) moves to a vertex and makes the next call undefined. The printed efficient-solver claim lacks a valid domain-preserving procedure. No implementation or empirical failure is inferred.","primaryUrl":"https://arxiv.org/html/2607.13686","noteUrl":"/review?paper=review-neurips-2026-155961","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155962","title":"Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for   RL-based LLM Recommenders","authors":["Wentao Shi","Qifan Wang","Chen Chen","Fei Liu","Dongfang Liu","Xu Liu","Wanli Ma","Junfeng Pan","Linhong Zhu","Fuli Feng"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for   RL-based LLM Recommenders","plainExplanation":"The report separates the failed objective equivalence from valid combinatorial recall bounds and supplies a qualified interval comparison and a corrected reweighting operator.","primaryUrl":"https://arxiv.org/html/2604.22504","noteUrl":"/review?paper=review-neurips-2026-155962","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","incorrect_bound","invalid_inference","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-155963","title":"Bentkus-type asymptotic e-values","authors":["Diego Martinez Taboada","Ben Chugg","Aaditya Ramdas"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Bentkus-type asymptotic e-values","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.06332","noteUrl":"/review?paper=review-neurips-2026-155963","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155968","title":"$h$-control: Training-Free Camera Control via Block-Conditional Gibbs Refinement","authors":["Yuzhu Wang","Xi Ye","Duo Su","Yangyang Xu","Jun Zhu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Optimal deterministic denoising does not preserve the data law","plainExplanation":"Optimal deterministic denoising is not posterior sampling and does not establish the stated stationary data law. The velocity-score identity mismatches the linear interpolation, and the explicit guidance formula diverges in a case where its claimed limit is hard replacement.","primaryUrl":"https://arxiv.org/html/2605.11871","noteUrl":"/review?paper=review-neurips-2026-155968","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155973","title":"Block-Wise Differentiable Sinkhorn Attention: Tail-Refinement Gradients with a Gap-Aware Dustbin Bridge","authors":["Dylan Forde"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Block-Wise Differentiable Sinkhorn Attention: Tail-Refinement Gradients with a Gap-Aware Dustbin Bridge","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.08123","noteUrl":"/review?paper=review-neurips-2026-155973","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155978","title":"TUBE: Tangent Upper Bound on Evidence for Discrete Diffusion Language Models","authors":["Arseny Ivanov","Sergei Kholkin","Vladislav Gromadskii","Grigoriy Ksenofontov","Ivan Oseledets","Aleksandr Korotin"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"TUBE: Tangent Upper Bound on Evidence for Discrete Diffusion Language Models","plainExplanation":"No issue was found in the checked claims. The source promises population bounds and expectation-preserving estimation. It does not assert that every sampled pair contains the true log likelihood.","primaryUrl":"https://arxiv.org/html/2605.24292","noteUrl":"/review?paper=review-neurips-2026-155978","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155979","title":"VicEdit: Learning to Edit Videos from Visual In-Context Examples","authors":["Yuji Wang","Teng Hu","Yuheng Chen","Ran Yi","Han Feng","Weijian Cao","Chengjie Wang","Lizhuang Ma","Jiangning Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The stated shift depends on modality rather than exemplar content","plainExplanation":"Equation (2) constructs \\(Q_{\\mathrm{vis}}\\) solely from learned base and modality embeddings, and equation (4) applies a fixed MLP to it. At fixed modality and parameters, \\(\\partial\\Delta_{\\mathrm{inst}}/\\partial\\mathrm{reference}=0\\). The content-dependent \\(T_{\\mathrm{vis}}\\) exists elsewhere but is not the shift’s input. The false functional-dependence claim is local to the specified shift. Joint MLLM states and later visual injection can still make final outputs exemplar-dependent.","primaryUrl":"https://arxiv.org/html/2608.16745","noteUrl":"/review?paper=review-neurips-2026-155979","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155982","title":"Extracting Governing Equations from Latent Dynamics via Multi-View Contrastive Learning","authors":["Paolo Muratore","Mackenzie Mathis"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Extracting Governing Equations from Latent Dynamics via Multi-View Contrastive Learning","plainExplanation":"The representation can identify the true state while a free potential hides an incorrect dynamics offset. Contrastive optimality alone does not establish the stated dynamics identification.","primaryUrl":"https://arxiv.org/html/2606.13260","noteUrl":"/review?paper=review-neurips-2026-155982","collection":"neurips-2026","severity":"critical","issueTypes":["false_statement","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-155983","title":"Cylindrical Geodesic Flow Matching for Quasiperiodic Physiological Signal Transformation","authors":["Onur Selim Kilic","Afra Nawar","Cem O Yaldiz","Michael J Cho","Ahmet R Emirdagi","Demet Tangolar","Amirali Aghazadeh","Amit J Shah","Omer T Inan"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Cylindrical Geodesic Flow Matching for Quasiperiodic Physiological Signal Transformation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.08510","noteUrl":"/review?paper=review-neurips-2026-155983","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155985","title":"The Query Complexity of Local Search in Rounds on General Graphs","authors":["Simina Branzei","Ioannis Panageas","Dimitris Paparas"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"The Query Complexity of Local Search in Rounds on General Graphs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.13266","noteUrl":"/review?paper=review-neurips-2026-155985","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155986","title":"Bernoulli Flow Models: Self-Consistent Generative Modeling for Binary Data","authors":["Hao Mo","Liying Yang","Shumin Yao","Xinxing Yu","Ajian Liu","Xudong Mao","Yanyan Liang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Bernoulli Flow Models: Self-Consistent Generative Modeling for Binary Data","plainExplanation":"No issue was found in the checked claims. The current qualified claim does not assert exact grid consistency for an imperfect learned predictor. The retained discrepancy between oracle and learned kernels is now explicitly within the source's stated scope.","primaryUrl":"https://arxiv.org/html/2610.11362","noteUrl":"/review?paper=review-neurips-2026-155986","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-155987","title":"MaxSketch: Robust Distinct Counting in Streams via Random Projections","authors":["Nikos Tsikouras","Constantine Caramanis","Christos Tzamos"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The threshold estimator can overshoot its multiplicative interval","plainExplanation":"The explicit geometric-threshold estimator can return a count outside its claimed multiplicative interval. The threshold-learning proof also simplifies an inverse-confidence sample count into an unsupported polylogarithmic dependence.","primaryUrl":"https://arxiv.org/html/2605.15571","noteUrl":"/review?paper=review-neurips-2026-155987","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-155995","title":"Efficient, Accurate and Stable Gradients for Neural ODEs","authors":["Sam McCallum","James Foster"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The backward reconstruction uses the opposite step sign","plainExplanation":"Algorithm 1 claims exact state reconstruction and discrete gradients but adds the forward-step increment when reconstructing the auxiliary state. The backward reconstruction adds a positive-step increment where the forward equation requires the negative-step increment, changing the recovered state and its discrete gradient.","primaryUrl":"https://arxiv.org/html/2410.11648","noteUrl":"/review?paper=review-neurips-2026-155995","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155998","title":"The Future of Facts: Tracing the Factual Generation-Verification Gap","authors":["Tim R Davidson","Anja Surina","Caglar Gulcehre"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The self-generation utility formula loses query-dependent weighting","plainExplanation":"Generation correctness and verification accuracy can vary together across queries. Replacing the query-dependent generation probability by its marginal average loses this dependence, so the displayed functional is not general self-generated verification accuracy.","primaryUrl":"https://arxiv.org/html/2605.27564","noteUrl":"/review?paper=review-neurips-2026-155998","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-155999","title":"TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting","authors":["Vladyslav Moroshan","Julien Siems","Arber Zela","Timur Carstensen","Frank Hutter"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting","plainExplanation":"Near-orthogonal parameterization or nonexpansiveness does not imply norm equality.","primaryUrl":"https://arxiv.org/html/2510.25502","noteUrl":"/review?paper=review-neurips-2026-155999","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156000","title":"Mecha-nudges for Machines","authors":["Giulio Frey","Kawin Ethayarajh"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Mecha-nudges for Machines","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.23433","noteUrl":"/review?paper=review-neurips-2026-156000","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156001","title":"FASTER: Rethinking Real-Time Flow VLAs","authors":["Yuxiang Lu","Zhe Liu","Xianzhe Fan","Zhenya YANG","Jinghua Hou","Junyi Li","kaixin Ding","Hengshuang Zhao"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"FASTER: Rethinking Real-Time Flow VLAs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.19199","noteUrl":"/review?paper=review-neurips-2026-156001","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156003","title":"MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse","authors":["Dong Hwan Kim","Hyunsoo Yoon"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.14413","noteUrl":"/review?paper=review-neurips-2026-156003","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156004","title":"Robust Multi-view Clustering against Imperfect Information","authors":["Zhichao Huang","Haochen Zhou","Hao Wang","Xi Peng","Mouxing Yang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Robust Multi-view Clustering against Imperfect Information","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.04343","noteUrl":"/review?paper=review-neurips-2026-156004","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156005","title":"The Panel Complexity of Sortition: Is 12 Angry Men Enough?","authors":["Johannes Brustle","Simone Fioravanti","Tomasz Ponitka","Jeremy Vollen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Panel Complexity of Sortition: Is 12 Angry Men Enough?","plainExplanation":"One wrong sign already contributes two to the \\(\\ell_1\\) error. The allocation moment bound controls a fraction of wrong coordinates, not the dimension-free event printed in the proof.","primaryUrl":"https://arxiv.org/html/2504.20508","noteUrl":"/review?paper=review-neurips-2026-156005","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156006","title":"Learning Dynamic Evidence Routes for Vision Transformer Probing","authors":["Zice Wang","Zhenyu Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Learning Dynamic Evidence Routes for Vision Transformer Probing","plainExplanation":"No issue was found in the checked claims. The missing output projection in the historical formula is restored.","primaryUrl":"https://arxiv.org/html/2605.00915","noteUrl":"/review?paper=review-neurips-2026-156006","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156008","title":"SimReg: Achieving Higher Performance in the Pretraining via Embedding Similarity Regularization","authors":["Yan Sun","Guoxia Wang","Jinle Zeng","Jiabin Yang","Shuai Li","Li Shen","Dacheng Tao","Dianhai Yu","Haifeng Wang"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"SimReg: Achieving Higher Performance in the Pretraining via Embedding Similarity Regularization","plainExplanation":"Contracting an upper envelope does not establish the same contraction of the quantity below it.","primaryUrl":"https://arxiv.org/html/2605.08809","noteUrl":"/review?paper=review-neurips-2026-156008","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","mathematical_error","proof_gap","reporting_inconsistency"],"classifications":["major"]},{"slug":"review-neurips-2026-156010","title":"M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling","authors":["Mayank Mishra","Shawn Tan","Ion Stoica","Joseph Gonzalez","Tri Dao"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The declared recurrent gradient needs future-state adjoints","plainExplanation":"The missing future adjoint breaks the chain rule across timesteps, independently of the separate cache implementation omission.","primaryUrl":"https://arxiv.org/html/2603.14360","noteUrl":"/review?paper=review-neurips-2026-156010","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156011","title":"SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning","authors":["Peng Xia","Jianwen Chen","Hanyang Wang","Jiaqi Liu","Kaide Zeng","Yu Wang","Siwei Han","Yiyang Zhou","Xujiang Zhao","Haifeng Chen","Zeyu Zheng","Cihang Xie","Huaxiu Yao"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.08234","noteUrl":"/review?paper=review-neurips-2026-156011","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156014","title":"When Noise Meets Long-Tail: Feature-Threshold Dual Calibration for Robust Pseudo-Labeling","authors":["Ping Guo","Zhiqi Huang","Xinran Li"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"When Noise Meets Long-Tail: Feature-Threshold Dual Calibration for Robust Pseudo-Labeling","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.33668","noteUrl":"/review?paper=review-neurips-2026-156014","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156015","title":"CoWorld-VLA: Thinking in a Multi-Expert World Model for Autonomous Driving","authors":["Jingqi Wang","minqing huang","Zihan Liang","Yujiao Xiang","JiaJie Huang","Zhi Xu","Feiyang Tan","Hangning Zhou","Mu Yang","Gong Chen"],"year":2026,"category":"Robotics & action models","status":"no-error","issueTitle":"CoWorld-VLA: Thinking in a Multi-Expert World Model for Autonomous Driving","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.10426","noteUrl":"/review?paper=review-neurips-2026-156015","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156016","title":"The Quantization Benefits of Residual-Free Transformers","authors":["Yiping Ji","Mahalakshmi Sabanayagam","Peyman Moghadam","Hemanth Saratchandran","Simon Lucey"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Kurtosis mixing averages the post-activation cumulant","plainExplanation":"The kurtosis lemma mixes input and post-activation cumulants. The isometry estimates omit initial error and underestimate coherent sign-gradient effects, while the attention calculation infers zero bilinear mean from centering without controlling token dependence.","primaryUrl":"https://arxiv.org/html/2605.25880","noteUrl":"/review?paper=review-neurips-2026-156016","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156018","title":"Language-Based Agent Control","authors":["Timothy Zhou","Loris D'Antoni","Nadia Polikarpova"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Language-Based Agent Control","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12863","noteUrl":"/review?paper=review-neurips-2026-156018","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156019","title":"Characterizing Universal Object Representations Across Vision Models","authors":["Florian P Mahner","Johannes Roth","Ka Chun Lam","Mick Bonner","Francisco Pereira","Martin N Hebart"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Symmetric NMF has ambiguities beyond column permutation","plainExplanation":"Nonnegativity alone does not reduce all Gram-factor ambiguities to column permutations. Strictly positive factors admit small non-permutation orthogonal rotations that remain positive.","primaryUrl":"https://arxiv.org/html/2605.13675","noteUrl":"/review?paper=review-neurips-2026-156019","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["missing_assumption"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-156020","title":"A Steerable Deep Network for Model-Free Diffusion MRI Registration","authors":["Gianfranco Cortés","Xiaoda Qu","Baba C Vemuri"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"An equivariant field predictor need not produce an invariant fixed field","plainExplanation":"The deformation argument confuses equivariance of the image-to-field predictor with invariance of a fixed field. The correct relation compares flows for transformed inputs by conjugacy. A local pullback formula also prints position where the direction variable is required; changing that one letter restores the action.","primaryUrl":"https://arxiv.org/pdf/2501.04794v2","noteUrl":"/review?paper=review-neurips-2026-156020","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156022","title":"Nonconvex Decentralized Stochastic Bilevel Optimization under Heavy-Tailed Noise","authors":["Xinwen Zhang","Yihan Zhang","Heng Liang","Hongchang Gao"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Expected-norm smoothness does not imply a mean-square increment bound","plainExplanation":"The stochastic increment proof upgrades expected-norm smoothness to a second-moment bound without support. A separate y-estimator increment lemma also uses the x-estimator residual instead of its own error.","primaryUrl":"https://arxiv.org/html/2509.15543","noteUrl":"/review?paper=review-neurips-2026-156022","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156023","title":"Multiscale Supervised Unbalanced Optimal Transport Flow Matching","authors":["Qiangwei Peng","Lezhi Chen","Peijie Zhou"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Cutoff rejection in independent lifting changes endpoint marginals","plainExplanation":"Arbitrary masks can break the claimed WFR formulation equivalence by isolating positive endpoint mass. The printed relative-mass loss has an extra pair-weight division already absent from the pinned code. Independent-lifting cutoff rejection can also change endpoint marginals, beyond the acknowledged loss of microscopic optimality.","primaryUrl":"https://arxiv.org/html/2605.16529","noteUrl":"/review?paper=review-neurips-2026-156023","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","definition_error","formula_error","implementation_error","missing_assumption"],"classifications":["major"]},{"slug":"review-neurips-2026-156024","title":"Optimizer-Induced Mode Connectivity: From AdamW to Muon","authors":["Fangzhao Zhang","Sungyoon Kim","Erica Zhang","Yiqi Jiang","Mert Pilanci"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Optimizer-Induced Mode Connectivity: From AdamW to Muon","plainExplanation":"At equality the runner-up minimum is attained and belongs to the closed ball.","primaryUrl":"https://arxiv.org/html/2605.09991","noteUrl":"/review?paper=review-neurips-2026-156024","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156028","title":"Mini-batch kernel $k$-means","authors":["Ben Jourdan","Gregory Schwartzman"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mini-batch kernel $k$-means","plainExplanation":"A sample mean introduces variance even when the current center is already population optimal.","primaryUrl":"https://arxiv.org/html/2410.05902","noteUrl":"/review?paper=review-neurips-2026-156028","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156033","title":"Toward Multimodal Sheet Music Recognition and Understanding","authors":["Guang Yang","Brian Zheng","Victoria Ebert","Noah Smith"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Toward Multimodal Sheet Music Recognition and Understanding","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.05769","noteUrl":"/review?paper=review-neurips-2026-156033","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156034","title":"3D Molecule Generation from Rigid Motifs via $\\mathrm{SE}(3)$ Flows","authors":["Roman Poletukhin","Marcel Kollovieh","Eike S. Eberhard","Stephan Günnemann"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"3D Molecule Generation from Rigid Motifs via $\\mathrm{SE}(3)$ Flows","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.16955","noteUrl":"/review?paper=review-neurips-2026-156034","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156035","title":"MARS: Multi-resolution Adaptive Routing for Sequential Recommendation","authors":["Ming Yin","Sixun Dong","Yudong Liu","Wenyun Yang","Yunjiang Jiang","Yiran Chen"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"MARS: Multi-resolution Adaptive Routing for Sequential Recommendation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.07505","noteUrl":"/review?paper=review-neurips-2026-156035","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156037","title":"Commutator Memory: Sparse, Path-Local Reading and Steering in Language Models","authors":["John Sweeney"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Commutator Memory: Sparse, Path-Local Reading and Steering in Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.34348","noteUrl":"/review?paper=review-neurips-2026-156037","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156039","title":"MobTA: Bus-Conditioned Zero-Shot Trajectory Generation via Task Arithmetic","authors":["SHUAI LIU","Ning Cao","Yue Jiang","Gao Cong"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"MobTA: Bus-Conditioned Zero-Shot Trajectory Generation via Task Arithmetic","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2602.13071","noteUrl":"/review?paper=review-neurips-2026-156039","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156041","title":"Nonparametric In-Context Learning under Growing Geometric Complexity: Minimax Optimality and Local Geometry-Adaptivity of Transformers","authors":["Jaehee Seo","Jisu Kim"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Nonparametric In-Context Learning under Growing Geometric Complexity: Minimax Optimality and Local Geometry-Adaptivity of Transformers","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31458","noteUrl":"/review?paper=review-neurips-2026-156041","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156042","title":"ASTRA: ADMM-Accelerated Topology Reconfiguration for Dynamic Satellite Constellations","authors":["João Norberto","Ricardo N. Ferreira","Claudia Soares"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"ASTRA: ADMM-Accelerated Topology Reconfiguration for Dynamic Satellite Constellations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.33993","noteUrl":"/review?paper=review-neurips-2026-156042","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156043","title":"Imagine Before You Draw: Visual Prompt Engineering for Image Generation","authors":["Liyu Jia","Fengda Zhang","Jiachun Pan","Kesen Zhao","Saining Zhang","Wang Lin","Weijia Wu","Yue Liao","Aojun Zhou","Hanwang Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Strict region-information loss needs a condition on the actual decoder features","plainExplanation":"The non-strict data-processing bound is valid for the complete decoder condition. Strictness requires positive conditional entropy of the particular preserved region, not merely loss of some image information.","primaryUrl":"https://arxiv.org/html/2606.04457","noteUrl":"/review?paper=review-neurips-2026-156043","collection":"neurips-2026","severity":"repairable","issueTypes":["missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156045","title":"Model-based Bootstrap of Controlled Markov Chains","authors":["Ziwei Su","Imon Banerjee","Diego Klabjan"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Model-based Bootstrap of Controlled Markov Chains","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12410","noteUrl":"/review?paper=review-neurips-2026-156045","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156046","title":"C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation","authors":["Cameron Dufault","Scott Xu","Alan Moses"],"year":2026,"category":"Molecular models & equivariance","status":"no-error","issueTitle":"C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.25242","noteUrl":"/review?paper=review-neurips-2026-156046","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156048","title":"SecureClaw: Clawing Back Control of LLM Agents","authors":["Yuhan Ma","Stefan Schmid"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"SecureClaw: Clawing Back Control of LLM Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2606.09549","noteUrl":"/review?paper=review-neurips-2026-156048","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156051","title":"Fast and Stable Gradient Approximation for Bilinear Forms of Hermitian Matrix Functions","authors":["Navjot Singh","Kipton Barros","Sherry Li"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Fast and Stable Gradient Approximation for Bilinear Forms of Hermitian Matrix Functions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.12801","noteUrl":"/review?paper=review-neurips-2026-156051","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156052","title":"Attention Transfer Is Not Universally Effective for Vision Transformers","authors":["Huaiyuan Qin","Muli Yang","Gabriel James Goenawan","Peng Hu","Chen Gong","Xi Peng","Hongyuan Zhu"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Attention Transfer Is Not Universally Effective for Vision Transformers","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.07191","noteUrl":"/review?paper=review-neurips-2026-156052","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156055","title":"Physics Unrolled Neural Operator for Wireless Field Modeling","authors":["Rafid Umayer Murshed","Saif U Rahman","Mingyue Tang","Elahe Soltanaghai"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Mean-zero noise does not preserve the Huber optimum","plainExplanation":"Let noise be \\(-1\\) with probability \\(0.9\\) and \\(9\\) with probability \\(0.1\\), giving mean zero. For unit-threshold Huber loss, the expected score at clean prediction zero is \\(0.8\\), not zero. Near the optimum, the gradient is \\(0.9(a+1)-0.1\\), which vanishes at \\(a=-8/9\\), not the clean target. Fixed spatial weights do not remove this bias, and a fixed spatial pattern extends it to a Sobel coordinate. The squared-loss expansion remains valid. Its applicability to Huber training is not established by conditional mean-zero noise alone.","primaryUrl":"https://arxiv.org/html/2608.18495","noteUrl":"/review?paper=review-neurips-2026-156055","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156057","title":"Drift-Aware Multimodal User Representation Learning via Multi-Scale Temporal Modeling and Sparse Mixture-of-Experts","authors":["Ziqing Qian","Haohang Chen","Shengqi Dang","Yuhan Xiong","Canyu Shen","Jiaying Lei","Nan Cao"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"Drift-Aware Multimodal User Representation Learning via Multi-Scale Temporal Modeling and Sparse Mixture-of-Experts","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.25773","noteUrl":"/review?paper=review-neurips-2026-156057","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156058","title":"Scalable Derivative Gaussian Processes via Exact Gradient Reduction","authors":["Hyunseok Seung","Matthias Katzfuss"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Projected derivative covariances can be singular with distinct references","plainExplanation":"The reduced derivative is \\(q=(I\\otimes D^\\top)g\\). Distinct target-centered references can be collinear, making \\(D\\) rank one even when the number of references equals the ambient dimension. Projected components are then deterministically proportional, so their covariance and every conditional covariance share a nonzero null vector. Gradient noise added before projection cannot remove that deterministic relation. The ordinary-inverse evaluation formulas lack a condition. Conditional-law sufficiency itself is not refuted; a support-based formulation can handle degeneracy.","primaryUrl":"https://arxiv.org/html/2606.02909","noteUrl":"/review?paper=review-neurips-2026-156058","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156060","title":"MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving","authors":["Rajeev Yasarla","Deepti Hegde","Hsin-Pai Cheng","Shizhong Han","Yunxiao Shi","MeysamSadeghi","Hanno Ackermann","Litian Liu","Pranav Desai","Fatih Porikli","Mohammad Ghavamzadeh","Herbert Cai"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The positive TTC shortfall rewards lower time to collision","plainExplanation":"The illustrative TTC term is a positive shortfall added to a reward that the policy maximizes, while described as penalizing low TTC. Adding a positive TTC shortfall to maximized reward makes lower time to collision more rewarding, the reverse of its stated penalty interpretation. This is the printed illustrative term’s sign error. Other collision terms may dominate particular trajectories without repairing this local preference.","primaryUrl":"https://arxiv.org/html/2605.14201","noteUrl":"/review?paper=review-neurips-2026-156060","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156061","title":"DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation","authors":["Haozhe Xie","Beichen Wen","Jiarui Zheng","Zhaoxi Chen","Fangzhou Hong","Haiwen Diao","Ziwei Liu"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"The interpolation covariance is the square of its noise coefficient","plainExplanation":"Section 3.2 identifies the conditional law of \\(A^\\tau=\\tau A+(1-\\tau)\\varepsilon\\), with standard normal noise, as a Gaussian with covariance \\((1-\\tau)I\\). Scaling standard normal noise multiplies covariance by the coefficient’s square, so the stated interpolation and conditional Gaussian use inconsistent variances.","primaryUrl":"https://arxiv.org/html/2601.22153","noteUrl":"/review?paper=review-neurips-2026-156061","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156062","title":"DyRA: Dynamic Residual Approximation for Efficient Matrix Multiplication in DNNs","authors":["Daewon Chae","Hyunwon Chung","Changwoo Lee","Hun-Seok Kim"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Input covariance must weight the left side of the matrix error","plainExplanation":"The input-aware factorization puts the input-covariance factor on the wrong side. The sequence-shared ALS correction also requires a causal fitting protocol before it can supply autoregressive conditional likelihoods; whole-sequence fitting need not preserve prefixes.","primaryUrl":"https://arxiv.org/html/2610.02882","noteUrl":"/review?paper=review-neurips-2026-156062","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156063","title":"HEART: Hyperspherical Embedding Alignment via Kent-Representation Traversal in Diffusion Models","authors":["Arani Roy","Shristi Das Biswas","Kaushik Roy"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The attribute vector is tangent at the anchor rather than the current token","plainExplanation":"The projection ensures \\(d^\\top\\mu_-=0\\), but the domain of \\(\\operatorname{Exp}_h\\) requires \\(d^\\top h=0\\). These are different conditions unless the base points coincide or an additional relation holds. Inserting an off-tangent vector into the cosine–sine formula does not generally produce unit norm, even when \\(h\\) is near the anchor. The tangent-base mismatch is local. Final rescaling does not make the original off-domain expression the intended exponential map; no code or benchmark failure is established.","primaryUrl":"https://arxiv.org/html/2605.07973","noteUrl":"/review?paper=review-neurips-2026-156063","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156064","title":"Model-Aware Tokenizer Transfer","authors":["Mykola Haltiuk","Aleksander Smywiński-Pohl"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Model-Aware Tokenizer Transfer","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.21954","noteUrl":"/review?paper=review-neurips-2026-156064","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156067","title":"UNICOM: Unified Multimodal Modeling via Compressed Continuous Semantic Representations","authors":["Yaqi Zhao","Wang Lin","Zijian Zhang","Miles Yang","Zhao Zhong","Liefeng Bo","Jingyuan Chen","Wentao Zhang"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"UNICOM: Unified Multimodal Modeling via Compressed Continuous Semantic Representations","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.10702","noteUrl":"/review?paper=review-neurips-2026-156067","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156068","title":"Inference-Time Vulnerability Beyond Shallow Safety: Alignment Along Generation Trajectories","authors":["Kyungmin Park","Taesup Kim"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Inference-Time Vulnerability Beyond Shallow Safety: Alignment Along Generation Trajectories","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.04778","noteUrl":"/review?paper=review-neurips-2026-156068","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156072","title":"Bridging Modalities, Spanning Time: Structured Memory for Ultra-Long Agentic Video Reasoning","authors":["Jiazheng Li","Chi-Hao Wu","Yunze Liu","Kaize Ding","Jundong Li","Chuxu Zhang"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Bridging Modalities, Spanning Time: Structured Memory for Ultra-Long Agentic Video Reasoning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.08271","noteUrl":"/review?paper=review-neurips-2026-156072","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156073","title":"Control Charts for Multi-Agent Systems","authors":["Hayden Helm","Carey E Priebe","Brandon Duderstadt"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Control Charts for Multi-Agent Systems","plainExplanation":"The sample variance includes dispersion between the two means, which can grow quadratically with the shift.","primaryUrl":"https://arxiv.org/html/2605.11135","noteUrl":"/review?paper=review-neurips-2026-156073","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","mathematical_error","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156075","title":"Asymmetric Flow Models","authors":["Hansheng Chen","Jan Ackermann","Minseo Kim","Gordon Wetzstein","Leonidas Guibas"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"A sample-dependent control-variate coefficient can change the conditional target","plainExplanation":"Appendix C.3 extends the zero-conditional-mean identity to any coefficient, while Equation (18) chooses that coefficient using the same sampled full and low-rank targets. Conditional mean-zero residuals need not remain mean-zero after multiplication by a coefficient fitted from the same sampled targets, so the adaptive correction can bias the target. The adaptive coefficient need not preserve the target even before the frozen-model approximation. Constant coefficients restore the zero-mean identity.","primaryUrl":"https://arxiv.org/html/2605.12964","noteUrl":"/review?paper=review-neurips-2026-156075","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156077","title":"The Alignment Illusion in Multimodal Large Language Models","authors":["Hong-Han Wang","Yuntao Wang","Hu Ding"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"The Alignment Illusion in Multimodal Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2609.30210","noteUrl":"/review?paper=review-neurips-2026-156077","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156078","title":"BSO: Safety Alignment Is Density Ratio Matching","authors":["Tien-Phat Nguyen","Truong Nguyen","Thin Nguyen","Duy M. H. Nguyen","Ngoc-Thanh Dinh","Trung Le"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Strict convexity does not imply strictly positive curvature everywhere","plainExplanation":"The generator \\(h(t)=(t-1)^4\\) is strictly convex because its derivative is strictly increasing, but \\(h^{\\prime\\prime}(1)=0\\). The displayed weight therefore vanishes at ratio one. The receipt also supplies an admitted preference setup with model ratio one and positive true ratio-matching divergence, so this need not be a solved-instance zero weight. The universal positivity claim fails for the stated class. The named positive-curvature generators and strict-convex Bregman argmin identification survive.","primaryUrl":"https://arxiv.org/html/2605.12339","noteUrl":"/review?paper=review-neurips-2026-156078","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156084","title":"HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily","authors":["Xinyi Li","Ming Li","Lu Bai","Lixin Cui","Feilong Cao","Ke Lv","Yunliang Jiang","Pietro Lió"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Gate regularity cannot order every pair of model risks","plainExplanation":"The risk proposition quantifies over arbitrary full and semantic model pairs without controlling classifier quality. The expressivity proof also claims class separation from a construction whose normalized branch inputs remain identical.","primaryUrl":"https://arxiv.org/html/2605.04594","noteUrl":"/review?paper=review-neurips-2026-156084","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156085","title":"CD-RCM: Generalizable Continuous-Depth Novel View Synthesis for Reflectance Confocal Microscopy","authors":["Tooba Imtiaz","Milind Rajadhyaksha","Kivanc Kose","Jennifer Dy"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Neural interpolation is not inherently restricted to fixed query depths","plainExplanation":"The claimed intrinsic fixed-depth limitation of neural interpolation is false: a network conditioned on continuous query depth can approximate interpolation uniformly. The performance of the compared fixed-output baselines is a narrower empirical question.","primaryUrl":"https://arxiv.org/html/2606.12635","noteUrl":"/review?paper=review-neurips-2026-156085","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-156086","title":"PAIR: Prefix-Aware Internal Reward Model for Multi-Turn Agent Optimization","authors":["Wonjoong Kim","Yeonjun In","Sangwu Park","Dongha Lee","Chanyoung Park"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A frozen first-stage probe does not structurally preserve second-stage clean predictions","plainExplanation":"Equation (4) feeds the frozen base score into an unconstrained learned logistic correction head, while claiming structural preservation of clean-prefix performance. The correction head is unconstrained on clean inputs, so freezing its base probe does not prevent it from changing or degrading clean predictions. The universal structural-preservation assertion fails. Good measured clean-prefix performance remains compatible with the architecture.","primaryUrl":"https://arxiv.org/html/2605.17877","noteUrl":"/review?paper=review-neurips-2026-156086","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156089","title":"Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery","authors":["Jowaria Khan","Anindya Sarkar","Yevgeniy Vorobeychik","Elizabeth Bondi-Kelly"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The displayed ELBO must be maximized","plainExplanation":"Appendix F removes the squared mean-gradient contribution from the second moment without changing the distance-contraction coefficient. The convergence inequality drops the squared mean-gradient contribution without changing its contraction coefficient, and later steps conflate squared, norm, averaged and final-iterate bounds. The stated convergence/rate justification is incomplete. A corrected strongly convex stochastic-gradient theorem may still be available under properly established assumptions.","primaryUrl":"https://arxiv.org/html/2602.17605","noteUrl":"/review?paper=review-neurips-2026-156089","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156090","title":"Predict-Project-Renoise: Sampling Diffusion Models under Hard Constraints","authors":["Omer Rochman Sharabi","Gilles Louppe"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Predict-Project-Renoise: Sampling Diffusion Models under Hard Constraints","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2601.21033","noteUrl":"/review?paper=review-neurips-2026-156090","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156093","title":"Frank-Wolfe Beyond 1/t Convergence","authors":["Sebastian Pokutta"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Frank-Wolfe Beyond 1/t Convergence","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.28006","noteUrl":"/review?paper=review-neurips-2026-156093","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156096","title":"Neural Continuous-Time Markov Chain: Discrete Diffusion via Decoupled Jump Timing and Direction","authors":["Jingyuan Li","Xiaoyi Jiang","Fukang Wen","WEI LIU","Renqian Luo","Yi Zhu","Zuoqiang Shi","Pipi Hu"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Neural Continuous-Time Markov Chain: Discrete Diffusion via Decoupled Jump Timing and Direction","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.15694","noteUrl":"/review?paper=review-neurips-2026-156096","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156097","title":"TriAxialKV: Toward Extreme Low-Precision KV-Cache Quantization for Agentic Inference Tasks","authors":["Hanzhang Shen","Haoran Wu","Yiren Zhao","Robert Mullins"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"A budget-fitting precision upgrade need not reduce distortion","plainExplanation":"The scalable allocator's printed return path still performs exhaustive enumeration. Separately, its budget-only greedy upgrades can worsen attention-output distortion because the compared precision levels use different quantization groupings.","primaryUrl":"https://arxiv.org/html/2605.17170","noteUrl":"/review?paper=review-neurips-2026-156097","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156098","title":"Video-Mirai: Autoregressive Video Diffusion Models Need Foresight","authors":["Yonghao Yu","Lang Huang","Runyi Li","Zerun Wang","Toshihiko Yamasaki"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A positive mean bootstrap interval does not establish broad prompt coverage","plainExplanation":"Let paired differences be \\(+20\\) on five of \\(100\\) prompts and zero on the other \\(95\\), identically across inference seeds. A bootstrap replicate has favorable count \\(B\\sim\\operatorname{Binomial}(100,0.05)\\) and mean difference \\(0.2B\\). Since \\(P(B=0)=0.95^{100}\\approx0.00592<0.025\\), its 95% percentile lower endpoint is positive, although all gain comes from five percent of prompts. Pairing, seed averaging and positive aggregate weights preserve this construction. The positive-mean inference can remain valid. Excluding zero does not establish how many prompts improve, and a positive mean alone does not always imply interval exclusion.","primaryUrl":"https://arxiv.org/html/2606.03971","noteUrl":"/review?paper=review-neurips-2026-156098","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156102","title":"Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models","authors":["Chen Xiong","Zhiyuan HE","Pin-Yu Chen","Ching-Yun Ko","Tsung-Yi Ho"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2602.04896","noteUrl":"/review?paper=review-neurips-2026-156102","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156103","title":"Offline-Online Reinforcement Learning for Linear Mixture MDPs","authors":["Zhongjun Zhang","Sean R. Sinclair"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Offline-Online Reinforcement Learning for Linear Mixture MDPs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2604.11994","noteUrl":"/review?paper=review-neurips-2026-156103","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156104","title":"FineRMoE: Dimension Expansion for Finer-Grained Expert with an Upcycling Approach","authors":["Ning Liao","Xiaoxing Wang","Xiaohan Qin","Junchi Yan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"FineRMoE: Dimension Expansion for Finer-Grained Expert with an Upcycling Approach","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.13364","noteUrl":"/review?paper=review-neurips-2026-156104","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156105","title":"Differentiable Systematic Resampling for Variational Sequential Monte Carlo","authors":["Fredrik Cumlin","Saikat Chatterjee"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"Differentiable Systematic Resampling for Variational Sequential Monte Carlo","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2610.12094","noteUrl":"/review?paper=review-neurips-2026-156105","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156106","title":"MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models","authors":["Zhichao Yang","Yuanze Hu","Haojie Hao","Longkun Hao","Dongshuo Huang","Hongyu Lin","Li","Lanqing HONG","Yihang Lou","Yan Bai"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Stationarity does not remove residual curvature from the loss expansion","plainExplanation":"The local world-model theory omits residual curvature, allows cancellation outside the observed subspace, and transfers auxiliary curvature to a joint objective without controlling its other term. A separate MSE expansion has inconsistent factors.","primaryUrl":"https://arxiv.org/html/2606.04627","noteUrl":"/review?paper=review-neurips-2026-156106","collection":"neurips-2026","severity":"major","issueTypes":["lemma_error","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156113","title":"How Much of a Model Do We Need? Redundancy and Slimmability in Remote Sensing Foundation Models","authors":["Leonard Hackel","Begum Demir","Tom Burgert"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"How Much of a Model Do We Need? Redundancy and Slimmability in Remote Sensing Foundation Models","plainExplanation":"","primaryUrl":"https://arxiv.org/pdf/2601.22841v2","noteUrl":"/review?paper=review-neurips-2026-156113","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156114","title":"Revisiting Embodied Chain-of-Thought for Generalizable Robot Manipulation","authors":["Nan Sun","Yuan Zhang","Yongkun Yang","Wentao Zhao","Peiyan Li","Jun Guo","Wenxuan Song","Pengxiang Ding","Runze Suo","Yifei Su","Xin Xiao","Xinghang Li","Huaping Liu"],"year":2026,"category":"Robotics & action models","status":"finding","issueTitle":"Revisiting Embodied Chain-of-Thought for Generalizable Robot Manipulation","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.03784","noteUrl":"/review?paper=review-neurips-2026-156114","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156115","title":"DyJR: Preserving Local Policy Plasticity in Reinforcement Learning with Verifiable Rewards via Dynamic Jensen-Shannon Replay","authors":["Long Li","Zhijian Zhou","Tianyi Wang","Weidi Xu","Zuming Huang","Wei Chu","Zhe Wang","Shirui Pan","Chao Qu","Yuan Qi"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"DyJR: Preserving Local Policy Plasticity in Reinforcement Learning with Verifiable Rewards via Dynamic Jensen-Shannon Replay","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.16157v2","noteUrl":"/review?paper=review-neurips-2026-156115","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156118","title":"Mechanistic Interpretability with Sparse Autoencoder Neural Operators","authors":["Bahareh Tolooshams","Ailsa Shen","Animashree Anandkumar"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An isometric lift does not remove out-of-range dictionary components","plainExplanation":"The lifted encoder equivalence omits the requirement that its dictionary lie in the lift range. Several appendix spectral-gradient formulas also omit the conjugation present in the correct main proposition.","primaryUrl":"https://arxiv.org/html/2509.03738","noteUrl":"/review?paper=review-neurips-2026-156118","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156119","title":"Few Channels Draw The Whole Picture: Revealing Massive Activations in Diffusion Transformers","authors":["Evelyn Turri","Davide Bucciarelli","Sara Sarto","Lorenzo Baraldi","Marcella Cornia"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"Few Channels Draw The Whole Picture: Revealing Massive Activations in Diffusion Transformers","plainExplanation":"With the inherited mask, Equation (3) retains the target on selected channels and injects the source on all other channels. This is the opposite of the replacement operation stated in the text. The same polarity enters the spatial mask in Equation (4).","primaryUrl":"https://arxiv.org/html/2605.13974","noteUrl":"/review?paper=review-neurips-2026-156119","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","formula_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156120","title":"UniVLR: Unifying Text and Vision in Visual Latent Reasoning for Multimodal LLMs","authors":["Houcheng Jiang","Jiajun Fu","Junfeng Fang","Chen Gao","Xiang Wang","Xiangnan He","Yong Li"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"UniVLR: Unifying Text and Vision in Visual Latent Reasoning for Multimodal LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.11856","noteUrl":"/review?paper=review-neurips-2026-156120","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156121","title":"Adaptive Coordinate Transforms for Neural Operators","authors":["Chaoyu Liu","Zhonghao Li","Gaohang Chen","Zakhar Shumaylov","Zhongying Deng","Zhang Qian","Zhonghua Qiao","Carola-Bibiane Schönlieb"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Adaptive Coordinate Transforms for Neural Operators","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.06203","noteUrl":"/review?paper=review-neurips-2026-156121","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156126","title":"Harnessing Agentic Evolution","authors":["Jiayi Zhang","Yongfeng Gu","Jianhao Ruan","Maojia Song","Yiran Peng","Zhiguang Han","Jinyu Xiang","Zhitao Wang","Caiyin Yang","Bang Liu","Chenglin Wu","Yuyu Luo"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Harnessing Agentic Evolution","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2605.13821","noteUrl":"/review?paper=review-neurips-2026-156126","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156131","title":"Models Designed to Forget: Machine Unlearning via Key Deletion","authors":["Sonia Laguna","Jorge da Silva Gonçalves","Moritz Vandenhirtz","Alain Ryser","Irene Cannistraci","Julia Vogt"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Models Designed to Forget: Machine Unlearning via Key Deletion","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.15033","noteUrl":"/review?paper=review-neurips-2026-156131","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156132","title":"GPU-Accelerated Synthesis of Mixed-Boolean Arithmetic: Beyond Caching","authors":["Gabriel Bathie","Nathanaël Fijalkow"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Finite sampled specifications do not guarantee universal equivalence","plainExplanation":"Sampled expression agreement does not guarantee universal modular equivalence. The decoder's claimed syntax domain also differs from its canonicalized enumeration, adjacent ranks can cross decoding branches, and parity alone does not prove the thread map is a permutation.","primaryUrl":"https://arxiv.org/html/2605.08243","noteUrl":"/review?paper=review-neurips-2026-156132","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","reporting_inconsistency"],"classifications":["major"]},{"slug":"review-neurips-2026-156134","title":"Bayesian Optimization with Fisher Information Geometry: Gradient Bounds and Trust-Region Methods","authors":["Saksham Kiroriwal","Julius Pfrommer","Jürgen Beyerer"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"Bayesian Optimization with Fisher Information Geometry: Gradient Bounds and Trust-Region Methods","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.31107","noteUrl":"/review?paper=review-neurips-2026-156134","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156136","title":"AsyncOPD: How Stale Can On-Policy Distillation Be?","authors":["Wonjun Kang","Kevin Galim","Seunghyuk Oh","Minjun Kang","Sanghyun Park","Donghoon Kim","Minjae Lee","Minseo Kim","Rishabh Tiwari","Yuchen Zeng","HYUNG IL KOO","Kangwook Lee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"AsyncOPD: How Stale Can On-Policy Distillation Be?","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.24143","noteUrl":"/review?paper=review-neurips-2026-156136","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156137","title":"Amortized Optimal Transport from Sliced Potentials","authors":["Minh-Phuc Truong","Khai Nguyen"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Sinkhorn updates swap source and target marginal weights","plainExplanation":"Equations (8)–(9) update target potential \\(g\\in\\mathbb R^m\\) using source weights \\(\\alpha\\in\\mathbb R^n\\), and source potential using target weights. The target-potential update uses source marginal weights and vice versa, giving the wrong constraints and incompatible vector dimensions when support sizes differ. This is a local label swap in the printed Sinkhorn formulas.","primaryUrl":"https://arxiv.org/html/2604.15114","noteUrl":"/review?paper=review-neurips-2026-156137","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156141","title":"Efficient Adaptive Data Acquisition via Pretrained Belief Representations","authors":["Daolang Huang","Zhuoyue Huang","Conor Hassan","Luigi Acerbi","Samuel Kaski","Thomas Rainforth"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"A target posterior need not suffice for the next acquisition","plainExplanation":"A target posterior alone need not determine the next acquisition: histories can retain nuisance information that changes which available sensor will reveal the target. Sequential sufficiency needs a predictive-state condition.","primaryUrl":"https://arxiv.org/html/2606.25197","noteUrl":"/review?paper=review-neurips-2026-156141","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-156145","title":"LogSTOP: Temporal Scores over Prediction Sequences for Matching and Retrieval","authors":["Avishree Khare","Hideki Okamoto","Bardh Hoxha","Georgios Fainekos","Rajeev Alur"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The closed smoothing interval contains one extra time point","plainExplanation":"Section 3 interprets the recursion probabilistically under independent local events and true local probabilities; Until is expanded into two logically disjoint branches. Until’s alternative branches are logically disjoint and share a local event with its complement, so independent-local-event assumptions do not justify multiplying their complements as if the branches were independent. The stated local-event assumptions do not justify the general probability interpretation for Until. The score can remain a useful retrieval heuristic without that exact interpretation.","primaryUrl":"https://arxiv.org/html/2510.06512","noteUrl":"/review?paper=review-neurips-2026-156145","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156146","title":"DaCe-DT: Data-Centric Offline Multi-Task Reinforcement Learning via Adaptive Prompts and Trajectory Correction for Heterogeneous Tasks","authors":["Shudong Wang","Xinfei Wang","Chenhao Zhang","Shanchen Pang","Wenhao Ji","Haiyuan Gui","Meng Han","Xiaojian Liao"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Joint expectile minimization favors the aggressive boundary","plainExplanation":"The prompt-length training rule specifies an STE for rounding but not the following hard mask. The joint expectile objective also favors aggressive expectiles under symmetric noise, contrary to its claimed conservative adaptation.","primaryUrl":"https://arxiv.org/html/2610.11085","noteUrl":"/review?paper=review-neurips-2026-156146","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156147","title":"Predictively-Oriented Kalman Filtering","authors":["Zheyang Shen","Gerardo Duran-Martin","Chris Oates"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The covariance-order chain does not establish mean-update conservativity","plainExplanation":"Appendix A.4 infers Euclidean mean-update conservativity by replacing covariance factors on both sides of a matrix sandwich and identifying its final expression with the Kalman gain square. Covariance order does not justify replacing both sides of an arbitrary matrix sandwich, and the chain’s final expression is not the Kalman gain square, leaving conservativity unproved.","primaryUrl":"https://arxiv.org/html/2606.03230","noteUrl":"/review?paper=review-neurips-2026-156147","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156148","title":"Knowledge-Graph Paths as Intermediate Supervision for Self-Evolving Search Agents","authors":["Huyu Wu","Jun Liu","Xiaochi Wei","Yan Gao","YIWU","Yao Hu"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Knowledge-Graph Paths as Intermediate Supervision for Self-Evolving Search Agents","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2605.05702","noteUrl":"/review?paper=review-neurips-2026-156148","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156150","title":"Strategic Decision Focused Learning","authors":["Tinashe Handina","Yuehan Diao","Adam Wierman","Eric Mazumdar"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Strategic Decision Focused Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2609.14907","noteUrl":"/review?paper=review-neurips-2026-156150","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156633","title":"DiPMInd: Distance profile based mutual independence testing for random objects","authors":["Yaqing Chen","Paromita Dubey"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"DiPMInd: Distance profile based mutual independence testing for random objects","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2412.06766","noteUrl":"/review?paper=review-neurips-2026-156633","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156634","title":"Generalized Robust Adaptive-Bandwidth Multi-View Manifold Learning in High Dimensions with Noise","authors":["Xiucai Ding","Chao Shen","Hau-Tieng Wu"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"Squared-distance bandwidths scale with half the manifold dimension","plainExplanation":"The multiview expansion uses the wrong bandwidth-volume power, and the signal-energy proposition identifies distance quantiles with covariance trace without justification. The stated common bandwidth range also does not force its variance bound to vanish.","primaryUrl":"https://arxiv.org/html/2602.10530","noteUrl":"/review?paper=review-neurips-2026-156634","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156635","title":"Gradient descent inference in empirical risk minimization","authors":["Qiyang Han","Xiaocong Xu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The information identity differentiates the wrong link placement","plainExplanation":"The general-link information formula differentiates a different loss composition than the model defines. An auxiliary triangular-inverse bound also lacks the inverse scale required even for a scalar matrix.","primaryUrl":"https://arxiv.org/html/2412.09498","noteUrl":"/review?paper=review-neurips-2026-156635","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-156636","title":"Smooth Flow Matching for Synthesizing Functional Data","authors":["Jianbin Tan","Anru Zhang"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The composition and covariance arguments still need stronger hypotheses","plainExplanation":"Fixed-input derivative bounds do not ensure Sobolev regularity along a moving path, and operator-norm covariance convergence does not ensure Gaussian Wasserstein convergence in infinite dimension. The stronger functional-space assumptions and possible estimator-specific trace control are separate conditions, not consequences of those implications.","primaryUrl":"https://arxiv.org/html/2508.13831","noteUrl":"/review?paper=review-neurips-2026-156636","collection":"neurips-2026","severity":"major","issueTypes":["missing_assumption","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156637","title":"Fast Wasserstein rates for estimating probability distributions of probabilistic graphical models","authors":["Daniel Bartl","Stephan Eckstein"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Fast Wasserstein rates for estimating probability distributions of probabilistic graphical models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.09270","noteUrl":"/review?paper=review-neurips-2026-156637","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156638","title":"Privacy Guarantees in Posterior Sampling under Contamination","authors":["Shenggang Hu","Hongsheng Dai","Louis Aslett","Murray Pollock","Gareth Roberts"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Privacy Guarantees in Posterior Sampling under Contamination","plainExplanation":"The displayed expression integrates to the scale parameter rather than one.","primaryUrl":"https://arxiv.org/html/2403.07772","noteUrl":"/review?paper=review-neurips-2026-156638","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error"],"classifications":["major"]},{"slug":"review-neurips-2026-156639","title":"Consistent Bayesian Spatial Domain Partitioning Using Predictive Spanning Tree Methods","authors":["Kun Huang","Huiyan Sang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The square-grid edge count is too small for the counting bound","plainExplanation":"Proposition 1 asserts existence of a modeled partition with \\(k_0\\) nonempty clusters under geometric Assumption 2 alone. Geometric regularity alone does not ensure the grid contains enough indivisible blocks to form the claimed number of nonempty clusters. The standalone proposition omits a grid-resolution premise. The stronger main asymptotic assumptions exclude this coarse-grid example and are not refuted.","primaryUrl":"https://arxiv.org/html/2508.08324","noteUrl":"/review?paper=review-neurips-2026-156639","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156657","title":"Almost Sure Convergence of Linear Temporal Difference Learning with Arbitrary Features","authors":["Jiuqi Wang","Shangtong Zhang"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A local algebraic error in the linear-growth proof is repairable","plainExplanation":"The displayed constant need not bound the update. This is a local algebraic error: the existence of a linear-growth constant, and the convergence argument that needs its existence, can be recovered directly.","primaryUrl":"https://arxiv.org/html/2409.12135","noteUrl":"/review?paper=review-neurips-2026-156657","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-156658","title":"The Z-Gromov-Wasserstein Distance","authors":["Martin Bauer","Facundo Mémoli","Tom Needham","Mao Nishino"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The infinity-distance network space is not separable","plainExplanation":"The separability claim fails at the infinity endpoint, which contains an uncountable separated binary-kernel family. A separate lifting proof asks an augmented graph map to realize a product measure outside its image.","primaryUrl":"https://arxiv.org/html/2408.08233","noteUrl":"/review?paper=review-neurips-2026-156658","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156659","title":"Corruptions of Supervised Learning Problems: Typology and Mitigations","authors":["Laura Iacovissi","Nan Lu","Robert Williamson"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A marginal label inverse does not correct the joint supervised law","plainExplanation":"The marginal label Bayesian inverse does not generally correct supervised risk when inputs retain label information. The correction theorem also mixes forward and reverse factors, and its example matrices compose to a label swap rather than identity.","primaryUrl":"https://arxiv.org/pdf/2307.08643","noteUrl":"/review?paper=review-neurips-2026-156659","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156660","title":"Demographic parity in regression and classification within the unawareness framework","authors":["Vincent Divol","Solenne Gaucher"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Demographic parity in regression and classification within the unawareness framework","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2409.02471","noteUrl":"/review?paper=review-neurips-2026-156660","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156661","title":"DINOv3","authors":["Oriane Siméoni","Huy Vo","Maximilian Seitzer","Federico Baldassarre","Maxime Oquab","Cijo Jose","Vasil Khalidov","Marc Szafraniec","Seung Eun Yi","Michaël Ramamonjisoa","Francisco Massa","Daniel Haziza","Luca Wehrstedt","Jianyuan Wang","Timothée Darcet","Théo Moutakanni","Leonel Sentana","Claire Roberts","Andrea Vedaldi","Jamie Tolan","John Brandt","Camille Couprie","Julien Mairal","Herve Jegou","Patrick Labatut","Piotr Bojanowski"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"DINOv3","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2508.10104","noteUrl":"/review?paper=review-neurips-2026-156661","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156662","title":"Gradient Span Algorithms Make Predictable Progress in High Dimension","authors":["Felix Benning","Leif Döring"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Gradient Span Algorithms Make Predictable Progress in High Dimension","plainExplanation":"Dimension-valid covariance requires a positive mixture, without the moment conditions needed for derivatives.","primaryUrl":"https://arxiv.org/html/2410.09973","noteUrl":"/review?paper=review-neurips-2026-156662","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156663","title":"An Anytime Algorithm for Good Arm Identification","authors":["Marc Jourdan","Andrée Delahaye-Duriez","Clémence Réda"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The anytime error bound and supporting thresholds need correction","plainExplanation":"The right-hand side tends to zero when a positive gap tends to zero at fixed admissible time. The error of a nearly threshold Gaussian arm instead tends to one half. The large-branch inversion used in the proof cannot justify the stated all-time domain.","primaryUrl":"https://arxiv.org/html/2310.10359","noteUrl":"/review?paper=review-neurips-2026-156663","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","incorrect_bound"],"classifications":["major"]},{"slug":"review-neurips-2026-156664","title":"Online Bernstein-von Mises theorem","authors":["Jeyong Lee","Junhyeok Choi","Minwoo Chae"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The Gaussian covariance bound fails at the allowed square aspect ratio","plainExplanation":"The auxiliary theory applies an unselected Gaussian covariance bound after slab selection, allows an invalid square aspect ratio, and treats a nonsymmetric covariance product as a contraction. Separate tensor scaling, unlocalized exponential remainder and precision-concavity claims also fail. Two local displays omit a square and confuse Gaussian fourth and sixth moments; these findings do not constitute an end-to-end counterexample to the main theorem.","primaryUrl":"https://arxiv.org/html/2504.05661","noteUrl":"/review?paper=review-neurips-2026-156664","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","lemma_error","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156665","title":"Sublinear Variational Optimization of Gaussian Mixture Models with Millions to Billions of Parameters","authors":["Sebastian Salwig","Till Kahlke","Florian Hirschberger","Dennis Forster","Jörg Lücke"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A negative ELBO makes the relative-improvement stopping test stop too early","plainExplanation":"Equation (14) divides monotone ELBO improvement by the previous ELBO itself and stops when the ratio is below a positive tolerance. A negative previous ELBO makes every positive improvement produce a nonpositive relative ratio, automatically triggering the printed stopping criterion before convergence.","primaryUrl":"https://arxiv.org/html/2501.12299","noteUrl":"/review?paper=review-neurips-2026-156665","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156666","title":"\"What is Different Between These Datasets?\" A Framework for Explaining Data Distribution Shifts","authors":["Varun Babbar","Zhicheng Guo","Cynthia Rudin"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Evaluating aggregate influence on the fitted training loss gives zero scores","plainExplanation":"The influence definition has deletion sign and sample-normalization inconsistencies. More fundamentally, using the same optimized discriminator loss for aggregate evaluation makes every first-order point score zero at the exact empirical optimum.","primaryUrl":"https://arxiv.org/html/2403.05652","noteUrl":"/review?paper=review-neurips-2026-156666","collection":"neurips-2026","severity":"major","issueTypes":["algorithm_specification","mathematical_error"],"classifications":["major"]},{"slug":"review-neurips-2026-156667","title":"Global convergence of adjoint-optimized neural PDEs","authors":["Konstantin Riedl","Justin Sirignano","Konstantinos Spiliopoulos"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"An energy estimate controls adjoint norms by the square root of loss","plainExplanation":"Lemma 6.2 turns a bound on squared adjoint norms by residual energy into a bound on unsquared norms proportional to the loss itself. Taking an unsquared norm from a squared energy estimate requires a square root; the printed loss-linear bound fails under arbitrarily small residual scaling.","primaryUrl":"https://arxiv.org/html/2506.13633","noteUrl":"/review?paper=review-neurips-2026-156667","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156668","title":"Guaranteed Nonconvex Low-Rank Tensor Estimation via Scaled Gradient Descent","authors":["Tong Wu"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The weighted factor distance ignores rank-deficient slice errors","plainExplanation":"The target-weighted factor distance can miss nonzero product errors in rank-deficient transformed slices, and the inverse iteration can then be undefined. An auxiliary singular-value lower bound also ignores nullspaces, while an alignment lemma inverts a difference that vanishes at the optimum.","primaryUrl":"https://arxiv.org/html/2501.01696","noteUrl":"/review?paper=review-neurips-2026-156668","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156669","title":"Minimax Optimal Two-Sample Testing under Local Differential Privacy","authors":["Jongmin Mun","Seungwoo Kwak","Ilmun Kim"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"The scaled discrete-noise mechanism has disjoint supports","plainExplanation":"The adaptive resolution count can be nonpositive in an allowed low-information regime. The printed scaled discrete-noise mechanism also has category-dependent lattice supports and fails local differential privacy. Appendix formulas additionally misdefine chi-square divergence and omit the normalization in a projection used for a baseline null law.","primaryUrl":"https://arxiv.org/html/2411.09064","noteUrl":"/review?paper=review-neurips-2026-156669","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156670","title":"Multiple Instance Verification","authors":["Xin Xu","Eibe Frank","Geoff Holmes"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Scalar conditional entropy cannot support the proof’s pointwise comparison","plainExplanation":"Proposition 4 uses positive-probability crossings between conditional entropies to claim strict benefit from combining query and target information. The proof treats scalar averaged entropies as pointwise posterior entropies; the literal assumption is impossible, and the natural posterior interpretation still does not guarantee strict information gain.","primaryUrl":"https://arxiv.org/html/2407.06544","noteUrl":"/review?paper=review-neurips-2026-156670","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156671","title":"Nonparametric Estimation of a Factorizable Density using Diffusion Models","authors":["Hyeok Kyu Kwon","Dongha Kim","Ilsang Ohn","Minwoo Chae"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"The Langevin exposition reverses the score drift","plainExplanation":"The Langevin background uses the opposite score sign from its stated stationary law. The OU proof also treats a variable rate as time homogeneous and integrates over overlapping cells as though they formed a partition.","primaryUrl":"https://arxiv.org/html/2501.01783","noteUrl":"/review?paper=review-neurips-2026-156671","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156737","title":"A Multi-Fidelity Control Variate Approach for Policy Gradient Estimation","authors":["Xinjie Liu","Cyrus Neary","Kushagra Gupta","Wesley Suttle","Christian Ellis","Ufuk Topcu","David Fridovich-Keil"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Summability does not ensure positive finite-sample descent weights","plainExplanation":"The proof transfers scalar-loss correlation to vector-gradient variance without justification. Its finite-sample theorem also lacks a per-step positivity condition, and an auxiliary gradient bound treats a return-to-go as a single reward.","primaryUrl":"https://arxiv.org/html/2503.05696","noteUrl":"/review?paper=review-neurips-2026-156737","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156738","title":"Accelerating Inference of Discrete Autoregressive Normalizing Flows by Selective Jacobi Decoding","authors":["Jiaru Zhang","Juanwu Lu","Xiaoyu Wu","Ziran Wang","Ruqi Zhang"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"Finite termination does not guarantee shrinking transient errors","plainExplanation":"Appendix B.2 correctly proves that the strictly autoregressive map reaches its fixed point in at most one sweep per coordinate. The auxiliary zero-spectral-radius criterion is too broad, and errors can grow before finite termination. Positive-tolerance stopping can also produce a noninjective approximate decoder despite an exact bijective underlying flow. A separate notation correction aligns the density formula with the direction of the encoder maps.","primaryUrl":"https://arxiv.org/html/2505.24791","noteUrl":"/review?paper=review-neurips-2026-156738","collection":"neurips-2026","severity":"repairable","issueTypes":["algorithm_specification","false_statement","formula_error","missing_assumption","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156739","title":"Achieving adaptivity and optimality for multi-armed bandits using Exponential-Kullback Leibler Maillard Sampling","authors":["Hao Qin","Kwang-Sung Jun","Chicheng Zhang"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"A misplaced exponential factor interrupts one regret-proof calculation","plainExplanation":"Integration does not turn a decaying exponential into a growing one. The printed upper bound with the larger factor can still be true; the defect is the equality and the subsequent use of a decaying series to obtain Proposition 22.","primaryUrl":"https://arxiv.org/html/2502.14379","noteUrl":"/review?paper=review-neurips-2026-156739","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["formula_error"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-156740","title":"Align and Distill: Unifying and Improving Domain Adaptive Object Detection","authors":["Justin Kay","Timm Haucke","Suzanne Stathatos","Siqi Deng","Erik Young","Pietro Perona","Sara Beery","Grant Van Horn"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Align and Distill: Unifying and Improving Domain Adaptive Object Detection","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2403.12029","noteUrl":"/review?paper=review-neurips-2026-156740","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156742","title":"Discovering Symbolic Differential Equations with Symmetry Invariants","authors":["Jianke Yang","Manu Bhat","Bryan Hu","Yadi Cao","Nima Dehmamy","Robin Walters","Rose Yu"],"year":2026,"category":"Graphs & equivariance","status":"no-error","issueTitle":"Discovering Symbolic Differential Equations with Symmetry Invariants","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.12083","noteUrl":"/review?paper=review-neurips-2026-156742","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156744","title":"Generalized Smooth Stochastic Variational Inequalities: Almost Sure Convergence and Convergence Rates","authors":["Daniil Vankov","Angelia Nedich","Lalitha Sankar"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"Conditioning on the clipping sample invalidates the deterministic variance bound","plainExplanation":"Equations (101) and (176) turn a global Markov event of probability at least one half into a lower bound proportional to the unconditional distance moment. The high-probability clipping event is correlated with iterate distance, so its probability bound cannot replace a conditional distance moment by the unconditional one. The rate proof’s progress lower bound is unsupported. This does not independently establish that the final convergence-rate conclusions are false.","primaryUrl":"https://arxiv.org/html/2410.12334","noteUrl":"/review?paper=review-neurips-2026-156744","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156746","title":"On the Convergence Analysis of Muon","authors":["Wei Shen","Ruichuan Huang","Minhui Huang","Cong Shen","Jiawei zhang"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"On the Convergence Analysis of Muon","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.23737","noteUrl":"/review?paper=review-neurips-2026-156746","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156747","title":"OpInf-LLM: Parametric PDE Solving with LLMs via Operator Inference","authors":["Zhuoyuan &amp;quot;Jacob&amp;quot; Wang","Hanjiang Hu","Xiyu Deng","Saviz Mowlavi","Yorie Nakahira"],"year":2026,"category":"Statistical learning","status":"finding","issueTitle":"The projection and lift use incompatible POD normalizations","plainExplanation":"Algorithm 1 identifies the POD basis directly with ordinary unweighted left singular vectors, while the code-generation prompt projects with spatial quadrature weight and lifts with the same basis. Unweighted SVD vectors are Euclidean-orthonormal rather than quadrature-orthonormal, so the stated projection and lift scale even in-subspace data by the spatial weight.","primaryUrl":"https://arxiv.org/html/2602.01493","noteUrl":"/review?paper=review-neurips-2026-156747","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156748","title":"Paradoxical noise preference in RNNs","authors":["Noah Eckstein","Manoj Srinivasan"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Noise rescaling is not without loss of generality for fixed tanh","plainExplanation":"The noise-scale reduction is valid for positively homogeneous ReLU but not fixed-amplitude tanh, whose saturation scale leaves an additional absolute parameter.","primaryUrl":"https://arxiv.org/html/2601.04539","noteUrl":"/review?paper=review-neurips-2026-156748","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-156749","title":"Rational Tuning of LLM Cascades via Probabilistic Modeling","authors":["Michael Zellinger","Matt Thomson"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A minimum threshold does not accept the minimum-score atom","plainExplanation":"Cascade risk requires calibration conditional on routing history, beyond each model's own confidence calibration. The minimum-threshold boundary also fails to reproduce accept-all when the fitted score law has an endpoint atom.","primaryUrl":"https://arxiv.org/html/2501.09345","noteUrl":"/review?paper=review-neurips-2026-156749","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156750","title":"RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance","authors":["Miao Lin","Feng Yu","Rui Ning","Lusi Li","Jiawei Chen","Qian Lou","Mengxin Zheng","Chunsheng Xin","Hongyi Wu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance","plainExplanation":"The Gaussian transfer bound remains valid, but it does not imply the detector’s decision. Correct rank formulas, explicit confidence over calibration, and a strict or randomized tie rule provide different repairs to the separate false-positive claims.","primaryUrl":"https://arxiv.org/html/2602.00183","noteUrl":"/review?paper=review-neurips-2026-156750","collection":"neurips-2026","severity":"critical","issueTypes":["algorithm_specification","false_statement","formula_error","incorrect_bound","proof_gap"],"classifications":["critical"]},{"slug":"review-neurips-2026-156751","title":"TextRegion: Text-Aligned Region Tokens from Frozen Image-Text Models","authors":["Yao Xiao","Qiqian Fu","Heyi Tao","Yuqun Wu","Zhen Zhu","Derek Hoiem"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"TextRegion: Text-Aligned Region Tokens from Frozen Image-Text Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.23769","noteUrl":"/review?paper=review-neurips-2026-156751","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156753","title":"What Should Embeddings Embed? Autoregressive Models Represent Latent Generating Distributions","authors":["Liyi Zhang","Michael Li","R. Thomas McCoy","Ted Sumers","Jian-Qiao Zhu","Tom Griffiths"],"year":2026,"category":"Representation learning","status":"no-error","issueTitle":"What Should Embeddings Embed? Autoregressive Models Represent Latent Generating Distributions","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2406.03707","noteUrl":"/review?paper=review-neurips-2026-156753","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156754","title":"AdaCubic: An Adaptive Cubic Regularization Optimizer for Deep Learning","authors":["Ioannis Tsingalis","Constantine Kotropoulos","Corentin Briat"],"year":2026,"category":"Optimization & generalization","status":"finding","issueTitle":"A larger mini-batch does not remove shared-probe Hessian noise","plainExplanation":"The shared-probe Hessian bound cannot gain accuracy from mini-batch size alone, and the diagonal Taylor model omits quadratic cross-curvature terms needed by the convergence proof. A fixed-iterate concentration bound separates probe and data errors; explicit Taylor corrections retain the omitted cross-curvature. Separate local errors concern Bernstein constants and a printed root-finder guard; the author code already uses the correct stopping test.","primaryUrl":"https://arxiv.org/html/2604.09437","noteUrl":"/review?paper=review-neurips-2026-156754","collection":"neurips-2026","severity":"repairable","issueTypes":["formula_error","incorrect_bound","invalid_inference","notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156755","title":"ARC-Encoder: learning compressed text representations for large language models","authors":["Hippolyte Pilchen","Edouard Grave","Patrick Perez"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"ARC-Encoder: learning compressed text representations for large language models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.20535","noteUrl":"/review?paper=review-neurips-2026-156755","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156756","title":"Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning","authors":["Charlotte Beylier","Hannah Selder","Arthur Fleig","Simon M. Hofmann","Nico Scherf"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2511.20591","noteUrl":"/review?paper=review-neurips-2026-156756","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156757","title":"Autofocus Retrieval: An Effective Pipeline for Multi-Hop Question Answering With Semi-Structured Knowledge","authors":["Derian Boer","Stephen Roth","Stephen Roth"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Outgoing-neighbor constraints restrict tails rather than heads","plainExplanation":"Algorithm 2 enforces a relational triplet by intersecting head candidates with outgoing neighbors of the heads, then tail candidates with incoming neighbors of the tails. The grounding update applies outgoing-neighbor constraints to heads instead of tails, and incoming constraints to tails instead of heads, deleting valid triplet candidates.","primaryUrl":"https://arxiv.org/html/2505.09246","noteUrl":"/review?paper=review-neurips-2026-156757","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156758","title":"Beyond the Linear Separability Ceiling: Aligning Representations in VLMs","authors":["Enrico Vompa","Tanel Tammet","Mohit Vaishnav"],"year":2026,"category":"Representation learning","status":"finding","issueTitle":"Beating one centroid probe does not establish failure of all linear decoders","plainExplanation":"Section 5.1 interprets success beyond its normalized nearest-centroid probe as a solution not linearly readable from the visual embeddings. The centroid probe tests one linear weight vector rather than all linear decoders, so outperforming it does not establish that the solution requires nonlinear readout. The nonlinear-only mechanistic inference is underdetermined. The measured improvement over the particular centroid probe remains valid.","primaryUrl":"https://arxiv.org/html/2507.07574","noteUrl":"/review?paper=review-neurips-2026-156758","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156760","title":"Exploring the Limitations of Layer Synchronization in Spiking Neural Networks","authors":["Roel Koopman","Amirreza Yousefzadeh","Mahyar Shahsavari","Guangzhi Tang","Manolis Sifalakis"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The displayed cross-entropy loss lacks its leading minus sign","plainExplanation":"Appendix A.4 writes \\(\\sum_i y_i\\log p_i\\) as the cross-entropy loss minimized in training. The printed loss is positive log correct-class probability, whose minimization reverses the intended cross-entropy update because the leading minus sign is missing. This is a local printed loss-sign defect.","primaryUrl":"https://arxiv.org/html/2408.05098","noteUrl":"/review?paper=review-neurips-2026-156760","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156762","title":"Many Circuits, One Mechanism: Input Variation and Evaluation Granularity in Circuit Discovery","authors":["Alireza Bayat Makou","Jingcheng Niu","Subhabrata Dutta","Iryna Gurevych"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Sufficient answer-transfer dimension does not identify all band information","plainExplanation":"Appendix F.9 interprets a minimal subspace trained to transfer the source target token as the number of dimensions encoding band identity. The intervention is optimized for source-answer transfer, so a small sufficient transfer subspace does not identify all remaining band information or its total encoding dimension.","primaryUrl":"https://arxiv.org/html/2606.06267","noteUrl":"/review?paper=review-neurips-2026-156762","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156765","title":"On Rate-Optimal Partitioning Classification from Observable and from Privatised Data","authors":["Balázs Cs. Csáji","László Györfi","Ambrus Tamás","Harro Walk"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Top-class ties leave positive third-class regret outside the margin bound","plainExplanation":"The multiclass rate permits top-class ties that its positive-gap margin conditions do not control. A third class can win strictly with enough probability to produce slower regret than the claimed rate.","primaryUrl":"https://arxiv.org/html/2312.14889","noteUrl":"/review?paper=review-neurips-2026-156765","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-156766","title":"On the Expressive Power and Limitations of Multi-Layer SSMs","authors":["Nikola Zubic","Yuyi Wang","Yuyi Wang","Davide Scaramuzza"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The proofs need explicit arithmetic closure and consistent time dependence","plainExplanation":"Closure under affine composition does not imply closure under a fixed finite encoding.","primaryUrl":"https://arxiv.org/html/2604.14501","noteUrl":"/review?paper=review-neurips-2026-156766","collection":"neurips-2026","severity":"repairable","issueTypes":["invalid_inference","missing_assumption"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156767","title":"Regret minimization in Linear Bandits with offline data via extended D-optimal exploration.","authors":["Sushant Vijayan","Arun Suggala","Karthikeyan Shanmugam","Soumyabrata Pal"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Unbounded reward gaps defeat the parameter-free regret bound","plainExplanation":"The parameter-free regret bound omits arbitrarily large initial exploration gaps. The exact mixing prescription can also have no integer full-phase schedule. Auxiliary effective-dimension and ellipsoid arguments respectively use the wrong extreme eigenvalue and ignore offline support outside the live set.","primaryUrl":"https://arxiv.org/html/2508.08420","noteUrl":"/review?paper=review-neurips-2026-156767","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156769","title":"TropNNC: Structured Neural Network Compression Using Tropical Geometry","authors":["Konstantinos Fotopoulos","Petros Maragos","Panagiotis Misiakos"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The propagation proof needs its input radius; two examples need local corrections","plainExplanation":"The final substitution replaces an input-radius factor by an output-radius expression. A contractive layer can make its output arbitrarily smaller than its input, so the needed inequality does not follow.","primaryUrl":"https://arxiv.org/html/2409.03945","noteUrl":"/review?paper=review-neurips-2026-156769","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156770","title":"Active Learning via Classifier Impact and Greedy Selection for Interactive Image Retrieval","authors":["Leah Bar","Boaz Lerner","Nir Darshan","Rami Ben-Ari"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The greedy guarantee fails for the signed variance objective","plainExplanation":"The signed posterior-variance objective cannot satisfy its claimed nonnegative multiplicative greedy guarantee, and normalization does not restore submodularity in general. A separate motivational probability formula averages prefixes rather than the all-correct event.","primaryUrl":"https://arxiv.org/html/2412.02310","noteUrl":"/review?paper=review-neurips-2026-156770","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","notation"],"classifications":["major"]},{"slug":"review-neurips-2026-156771","title":"ARROW: Augmented Replay for RObust World models","authors":["Abdulaziz Alyahya","Abdallah Al Siyabi","Markus R. Ernst","Luke Yang","Levin Kuhlmann","Gideon Kowadlo"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"ARROW: Augmented Replay for RObust World models","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2603.11395","noteUrl":"/review?paper=review-neurips-2026-156771","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156772","title":"ASAT: Adaptive Scoring and Thresholding with Human Feedback for Robust Out-of-Distribution Detection","authors":["Daisuke Yamada","Harit Vishwakarma","Ramya Korlakai Vinayak"],"year":2026,"category":"Statistical learning","status":"no-error","issueTitle":"ASAT: Adaptive Scoring and Thresholding with Human Feedback for Robust Out-of-Distribution Detection","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2505.02299","noteUrl":"/review?paper=review-neurips-2026-156772","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156774","title":"Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens","authors":["Karthik Valmeekam","Vardhan Palod","Kaya Stechly","Atharva Gundawar","Subbarao Kambhampati"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Correlation does not require closeness to the identity line","plainExplanation":"Section 7 says that correlation between generated and ground-truth trace lengths would require generated lengths to cluster near the ground-truth values. Distance from the identity line measures length calibration rather than correlation, so dispersion alone does not establish absence of association.","primaryUrl":"https://arxiv.org/html/2505.13775","noteUrl":"/review?paper=review-neurips-2026-156774","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156775","title":"Bigger Isn't Always Memorizing: Early Stopping Overparameterized Diffusion Models","authors":["Alessandro Favero","Antonio Sclocchi","Matthieu Wyart"],"year":2026,"category":"Generative models","status":"finding","issueTitle":"A Rayleigh quotient does not establish a learnable score eigenmode","plainExplanation":"Appendix E treats local cutoff functions as eigenfunctions and extends their inverse-Rayleigh-quotient fitting time to any isotropic kernel. A local Rayleigh quotient does not prove eigenfunction status or target spectral support, so it cannot imply a universal finite fitting time for isotropic kernels. The universal spectral inference is unsupported. The local power counting and observed NTK/ReLU scaling are not refuted.","primaryUrl":"https://arxiv.org/html/2505.16959","noteUrl":"/review?paper=review-neurips-2026-156775","collection":"neurips-2026","severity":"repairable","issueTypes":["notation","proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156776","title":"Classification of high-dimensional data with spiked covariance matrix structure","authors":["Yin-Jen Chen","Minh Tang"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Classwise sample sizes need not converge to the test mixture proportions","plainExplanation":"Theorem 3 uses \\(\\log(n_1/n_2)\\to\\log(\\pi_1/(1-\\pi_1))\\) by the strong law, while its stated sampling assumptions allow classwise fixed sample sizes with only a bounded ratio. Classwise iid sampling does not force class counts to match test priors, so the strong-law threshold step and Bayes-consistency conclusion need an additional sampling/prior hypothesis. The Bayes-consistency assertion needs a sampling/prior hypothesis. The separate support-recovery statement is unaffected.","primaryUrl":"https://arxiv.org/html/2110.01950","noteUrl":"/review?paper=review-neurips-2026-156776","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156777","title":"DecompDreamer: A Composition-Aware Curriculum for Structured 3D Asset Generation","authors":["Utkarsh Nath","Rajeev Goel","Rahul Khurana","Kyle Min","Mark Ollila","Pavan Turaga","Varun Jampani","Tejaswi Gowda"],"year":2026,"category":"Generative models","status":"no-error","issueTitle":"DecompDreamer: A Composition-Aware Curriculum for Structured 3D Asset Generation","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2503.11981","noteUrl":"/review?paper=review-neurips-2026-156777","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156778","title":"Fed-SB: A Silver Bullet for Extreme Communication Efficiency and Performance in (Private) Federated LoRA Fine-Tuning","authors":["Raghav Singhal","Kaustubh Ponkshe","Rohit Vartak","Lav Varshney","Praneeth Vepakomma"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Privacy accounting and trajectory claims need precise boundaries","plainExplanation":"A deterministic subspace selected from protected records is itself an output of the mechanism. Later noise confined to that subspace does not automatically protect its selection.","primaryUrl":"https://arxiv.org/html/2502.15436","noteUrl":"/review?paper=review-neurips-2026-156778","collection":"neurips-2026","severity":"major","issueTypes":["definition_ambiguity","invalid_inference","missing_assumption","notation","proof_gap"],"classifications":["major"]},{"slug":"review-neurips-2026-156780","title":"In-context Learning in Presence of Spurious Correlations","authors":["Hrayr Harutyunyan","Rafayel Darbinyan","Samvel Karapetyan","Hrant Khachatrian"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"In-context Learning in Presence of Spurious Correlations","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2410.03140","noteUrl":"/review?paper=review-neurips-2026-156780","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156781","title":"Memento No More: Coaching AI Agents to Master Multiple Tasks via Hints Internalization","authors":["Minttu Alakuijala","Ya Gao","Georgy Ananov","Samuel Kaski","Pekka Marttinen","Alexander Ilin","Harri Valpola"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The return index omits the final constructed iterate","plainExplanation":"The algorithm returns the iterate preceding its final update. Its inclusive loop constructs \\(\\theta_{I+1}\\), but the return statement selects \\(\\theta_I\\). This is a pseudocode off-by-one. It does not show that the executed experiments discarded an update round.","primaryUrl":"https://arxiv.org/html/2502.01562","noteUrl":"/review?paper=review-neurips-2026-156781","collection":"neurips-2026","severity":"minor_cosmetic","issueTypes":["notation"],"classifications":["minor_cosmetic"]},{"slug":"review-neurips-2026-156782","title":"Mollifier Layers: Enabling Efficient High-Order Derivatives in Inverse PDE Learning","authors":["Vinayak Vinayak","Ananyae bhartari","Vivek Shenoy"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Lipschitz continuity of a field does not give linear derivative bias","plainExplanation":"The derivative theorem needs the derivative’s modulus of continuity and an inverse-bandwidth noise term. Interior counterexamples identify both defects, and a normalized-grid bound recovers convergence under explicit regularity and noise conditions. Additional local corrections concern the kernel family, the A.7 normalization and the printed cosine schedule; Appendix A.8 already states a valid smooth unit-mass kernel.","primaryUrl":"https://arxiv.org/html/2505.11682","noteUrl":"/review?paper=review-neurips-2026-156782","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","formula_error","incorrect_bound","missing_assumption","notation"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156783","title":"No Detail Left Behind: Revisiting Self-Retrieval for Fine-Grained Image Captioning","authors":["Manu Gaur","Darshan Singh","Makarand Tapaswi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"No Detail Left Behind: Revisiting Self-Retrieval for Fine-Grained Image Captioning","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2409.03025","noteUrl":"/review?paper=review-neurips-2026-156783","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156784","title":"On Fitting Flow Models with Large Sinkhorn Couplings","authors":["Stephen Zhang","Alireza Mousavi-Hosseini","Michal Klein","Marco Cuturi"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Uniform conditional variance needs stronger small-ball control","plainExplanation":"The pair-averaged condition allows the target distribution to concentrate near some points more strongly than a uniform small-ball condition permits. Conditional matching near such points can have smaller variance.","primaryUrl":"https://arxiv.org/html/2506.05526","noteUrl":"/review?paper=review-neurips-2026-156784","collection":"neurips-2026","severity":"major","issueTypes":["false_statement","mathematical_error"],"classifications":["major"]},{"slug":"review-neurips-2026-156785","title":"OTIS: Learning High-Quality Time Series Features With Tiny Encoders","authors":["Özgün Turgut","Philip Müller","Martin Menten","Daniel Rueckert"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"OTIS: Learning High-Quality Time Series Features With Tiny Encoders","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2410.07299","noteUrl":"/review?paper=review-neurips-2026-156785","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156786","title":"Random features for Grassmannian kernel approximation with bounded rank-one projections","authors":["Remi Delogne","Laurent Jacques"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A rank-based pseudodeterminant drops zero principal cosines","plainExplanation":"The rank-based pseudodeterminant is not the basis cross-Gram determinant when some principal cosines vanish. Dropping zero eigenvalues changes the claimed Grassmannian kernel and distance identity.","primaryUrl":"https://arxiv.org/html/2608.04227","noteUrl":"/review?paper=review-neurips-2026-156786","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-156787","title":"Reasoning Under 1 Billion: Memory-Augmented Reinforcement Learning for Large Language Models","authors":["Hung Le","Van Dai Do","Dung Nguyen","Svetha Venkatesh"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Reasoning Under 1 Billion: Memory-Augmented Reinforcement Learning for Large Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2504.02273","noteUrl":"/review?paper=review-neurips-2026-156787","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156788","title":"Replicability is Asymptotically Free in Multi-armed Bandits","authors":["Junpei Komiyama","Shinji Ito","Yuichi Yoshida","Souta Koshino"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"Replicable allocation does not standardize the two-sample statistic","plainExplanation":"Lemma 7 claims simultaneous coverage of all active linear-bandit arms with radius lacking an arm-count adjustment. The radius lacks the adjustment needed for simultaneous coverage across many independent arm errors, so a valid design can exceed the stated joint failure probability. The lemma’s uniform coverage is false at the printed radius. A fixed-arm interpretation can hold, and later main-proof union bounds may already supply the needed adjustment.","primaryUrl":"https://arxiv.org/html/2402.07391","noteUrl":"/review?paper=review-neurips-2026-156788","collection":"neurips-2026","severity":"repairable","issueTypes":["false_statement","lemma_error"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156789","title":"RESIST: Resilient Decentralized Learning Using Consensus Gradient Descent","authors":["Cheng Fang","Rishabh Dixit","Waheed Bajwa","Mert Gurbuzbalaban"],"year":2026,"category":"Optimization & generalization","status":"no-error","issueTitle":"RESIST: Resilient Decentralized Learning Using Consensus Gradient Descent","plainExplanation":"No issue was found in the checked claims. Let u=1/M times the all-ones vector, Q_k^pi=1 c_k^T, P=1 u^T. Row stochasticity gives c_k>=0 and sum c_k=1, so ||Q_k^pi-P||_2=sqrt(M)||c_k-u||_2<=sqrt(M-1), for every M>=1. Equation (20) identifies T=gradient F(W). Let What=1 what^T, Wstar=1 wstar^T, and Wtilde have rows w_jstar. Since each local minimizer is unconstrained, gradient F(Wtilde)=0. L-smoothness and the Frobenius triangle inequality give ||T||_F<=L(||W-What||_F+||What-Wstar||_F+||Wstar-Wtilde||_F)<=L[sqrt(sum_l (xi_l^1)^2)+sqrt(M)xi^6+Delta]<=L[sum_l xi_l^1+sqrt(M)xi^6+Delta]. Hence xi_k^4<=sqrt(M-1)L[sum_l xi_l^1+sqrt(M)xi^6+Delta]. These three coefficients are respectively no larger than (sqrt(M)+2)sqrt(2), (sqrt(M)+2)M, and sqrt(M)+2 in (34). This proves (34) without changing its coefficients, for arbitrary M,d and coordinate-dependent stochastic weights.","primaryUrl":"https://arxiv.org/html/2502.07977","noteUrl":"/review?paper=review-neurips-2026-156789","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156790","title":"Rethinking Reward Models for Multi-Domain Test-Time Scaling","authors":["Dong Bok Lee","Seanie Lee","Sangwoo Park","Minki Kang","Jinheon Baek","Dongki Kim","Dominik Wagner","Jiongdao Jin","Heejun Lee","Tobias Bocklet","Jinyu Wang","Jingjing Fu","Sung Ju Hwang","Jiang Bian","Lei Song"],"year":2026,"category":"Reinforcement learning","status":"no-error","issueTitle":"Rethinking Reward Models for Multi-Domain Test-Time Scaling","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2510.00492","noteUrl":"/review?paper=review-neurips-2026-156790","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156791","title":"Rethinking the Mixture of Vision Encoders Paradigm for Enhanced Visual Understanding in Multimodal LLMs","authors":["Mozhgan Azadani","James Riddell","Sean Sedwards","Krzysztof Czarnecki"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Rethinking the Mixture of Vision Encoders Paradigm for Enhanced Visual Understanding in Multimodal LLMs","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2501.06986","noteUrl":"/review?paper=review-neurips-2026-156791","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-156792","title":"RSQ: Learning from Important Tokens Leads to Better Quantized LLMs","authors":["Yi-Lin Sung","Prateek Yadav","Jialu Li","Jaehong Yoon","Mohit Bansal"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"RSQ: Learning from Important Tokens Leads to Better Quantized LLMs","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2503.01820","noteUrl":"/review?paper=review-neurips-2026-156792","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156795","title":"The Confusion is Real: GRAPHIC – A Network Science Approach to Confusion Matrices in Deep Learning","authors":["Johanna S. Fröhlich","Bastian Heinlein","Jan Claar","Hans Rosenberger","Vasileios Belagiannis","Ralf R Müller"],"year":2026,"category":"Graphs & equivariance","status":"finding","issueTitle":"A pseudo-label probe can exceed its teacher's true-label accuracy","plainExplanation":"A linear probe trained only on pseudo-labels can denoise nonlinear teacher errors in frozen features and exceed the teacher's true-label accuracy. The claimed universal ceiling therefore does not follow from label provenance.","primaryUrl":"https://arxiv.org/pdf/2602.19770v2","noteUrl":"/review?paper=review-neurips-2026-156795","collection":"neurips-2026","severity":"major","issueTypes":["false_statement"],"classifications":["major"]},{"slug":"review-neurips-2026-156796","title":"The Foundation Model Transparency Index","authors":["Rishi Bommasani","Kevin Klyman","Shayne Longpre","Sayash Kapoor","Nestor Maslej","Betty Xiong","Daniel Zhang","Percy Liang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"The Foundation Model Transparency Index","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2310.12941","noteUrl":"/review?paper=review-neurips-2026-156796","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-156798","title":"Towards Principled Task Grouping for Multi-Task Learning","authors":["Chenguang Wang","Xuanhao Pan","Tianshu Yu"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"The transfer-gain orientation is reversed for raw accuracy","plainExplanation":"Definition 3 permits loss or validation accuracy in \\(1-L_{\\rm paired}/L_{\\rm solo}\\), while maximizing the resulting transfer gain. Maximizing the displayed ratio-based gain prefers a smaller paired metric, so permitting raw accuracy reverses the intended preference for improvement. The accuracy-permitted definition has the wrong orientation. The loss/error-based experiments are not invalidated.","primaryUrl":"https://arxiv.org/html/2402.15328","noteUrl":"/review?paper=review-neurips-2026-156798","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156799","title":"Transformers in the Dark: Navigating Unknown Search Spaces via Bandit Feedback","authors":["Jungtaek Kim","Thomas Zeng","Ziqian Lin","Minjae Lee","Chungpa Lee","Jy-yong Sohn","Hyung Koo","Kangwook Lee"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"The explicit transformer constructions need corrected reward and membership registers","plainExplanation":"This does not store a frontier state’s reward in its own coordinate. Subsequent normalized attention therefore mixes token occurrence counts with a common offset rather than comparing identity-specific inherited rewards.","primaryUrl":"https://arxiv.org/html/2603.24780","noteUrl":"/review?paper=review-neurips-2026-156799","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-156800","title":"VLMGuard: Bootstrapping Malicious Prompt Detectors from Unlabeled Vision-Language Prompts in the Wild","authors":["Junlin Fang","Wenyu Chen","Reshmi Ghosh","Robert Sim","Ahmed Salem","Vitor Carvalho","Emily Lawton","Sharon Li","Jack Stokes","Sean Du"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"An unweighted projection theorem does not justify the singular-value-weighted score","plainExplanation":"Appendix K.4 identifies its valid unweighted top-subspace score-gap theorem with the actual singular-value-weighted score in Equation (6). Singular-value weights can reverse the class score gap, so a positive gap theorem for unweighted projection energy does not justify the method’s weighted score.","primaryUrl":"https://arxiv.org/html/2410.00296","noteUrl":"/review?paper=review-neurips-2026-156800","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-157013","title":"XFeat Revisited: Reproducibility and Evaluation of a Lightweight Image Matcher","authors":["Lazar Đoković","Aimee Lin"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"XFeat Revisited: Reproducibility and Evaluation of a Lightweight Image Matcher","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2608.09519","noteUrl":"/review?paper=review-neurips-2026-157013","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-157017","title":"[Re] FairDICE: A Fair Tradeoff in Multi-objective Offline RL","authors":["Peter Adema","Karim Galliamov","Aleksey Evstratovskiy","Ross Geurts"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"[Re] FairDICE: A Fair Tradeoff in Multi-objective Offline RL","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2603.03454","noteUrl":"/review?paper=review-neurips-2026-157017","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-157020","title":"WAREX: Web Agent Reliability Evaluation on Existing Benchmarks","authors":["Su Kara","Fazle Faisal","Suman Nath"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"WAREX: Web Agent Reliability Evaluation on Existing Benchmarks","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2510.03285","noteUrl":"/review?paper=review-neurips-2026-157020","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-157021","title":"[Re] Boosting the Visual Interpretability of CLIP via Adversarial Fine-Tuning","authors":["Anton Nuzhdin","Andrei Gruzitski","Balázs Egyed","Aleksandr Raudvee"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"A parameter-dependent saliency norm changes the optimization objective","plainExplanation":"Proposition 1 converts background worst-case score change into normalized EPG and claims equivalent minimization over encoder parameters. The normalization factor depends on the optimized encoder, so the valid first-order identity does not imply equal minimizers for normalized saliency and unnormalized background damage.","primaryUrl":"https://arxiv.org/html/2610.14832","noteUrl":"/review?paper=review-neurips-2026-157021","collection":"neurips-2026","severity":"repairable","issueTypes":["proof_gap"],"classifications":["repairable"]},{"slug":"review-neurips-2026-157633","title":"Reproducing FACTER: Fairness via Conformal Thresholding and Prompt Repair","authors":["Oscar Miró López-Feliu","Daimy van Loo","Xanthos Kekkos","Mikel Blom","Clara Rus"],"year":2026,"category":"Machine learning","status":"finding","issueTitle":"Reproducing FACTER: Fairness via Conformal Thresholding and Prompt Repair","plainExplanation":"","primaryUrl":"https://arxiv.org/html/2606.28620","noteUrl":"/review?paper=review-neurips-2026-157633","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-166291","title":"Min-p, Max Exaggeration: A Critical Analysis of Min-p Sampling in Language Models","authors":["Rylan Schaeffer","Joshua Kazdan","Yegor Denisov-Blanch"],"year":2026,"category":"Transformers & sequence models","status":"no-error","issueTitle":"Min-p, Max Exaggeration: A Critical Analysis of Min-p Sampling in Language Models","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2506.13681","noteUrl":"/review?paper=review-neurips-2026-166291","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-178475","title":"Reviewer Scores Are Not Comparable Across Research Areas in ML Peer Review","authors":["Binyan Xu","Xilin Dai","Fan Yang","Kehuan Zhang"],"year":2026,"category":"Machine learning","status":"no-error","issueTitle":"Reviewer Scores Are Not Comparable Across Research Areas in ML Peer Review","plainExplanation":"No mathematical error was identified in the statements screened.","primaryUrl":"https://arxiv.org/html/2607.27209","noteUrl":"/review?paper=review-neurips-2026-178475","collection":"neurips-2026","severity":"no_issues","issueTypes":[],"classifications":["no_issues"]},{"slug":"review-neurips-2026-179728","title":"Don't Claim Benchmark-Oriented Optimization Improves General Coding Capability — Diverse Evaluation Is Required","authors":["Egor Shibaev","Vera Kudrevskaia","Timur Galimzyanov","Mikhail Evtikhiev","Ana Terna","Rastislav Rabatin","Timur Kudashev","Timofey Bryksin","Arina Puchkova","Patrik Bartak","Egor Bogomolov","Sergey Titov"],"year":2026,"category":"Transformers & sequence models","status":"finding","issueTitle":"Don't Claim Benchmark-Oriented Optimization Improves General Coding Capability — Diverse Evaluation Is Required","plainExplanation":"","primaryUrl":"https://arxiv.org/abs/2608.13566v1","noteUrl":"/review?paper=review-neurips-2026-179728","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]},{"slug":"review-neurips-2026-180254","title":"RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments","authors":["Babak Rahmani","Sebastian Dziadzio","Joschka Strüber","Sergio Hernández-Gutiérrez","Matthias Bethge"],"year":2026,"category":"Reinforcement learning","status":"finding","issueTitle":"RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments","plainExplanation":"","primaryUrl":"https://arxiv.org/abs/2606.26094v1","noteUrl":"/review?paper=review-neurips-2026-180254","collection":"neurips-2026","severity":null,"issueTypes":[],"classifications":[]}]}