[
  {
    "auditId": "E-ARC-03",
    "title": "Invariant and Equivariant Graph Networks",
    "year": 2018,
    "category": "Graphs & equivariance",
    "primaryUrl": "https://arxiv.org/abs/1812.09902",
    "citationCount": 583,
    "citationUrl": "https://www.semanticscholar.org/paper/6541afa4b4061a7d5c8387514bedea9dc249fd80",
    "citationDate": "2026-09-09",
    "issueTitle": "The equivariant basis count depends on graph size",
    "addressingSources": [
      {
        "label": "Pearce-Crump: Section 3.5 and Theorems 18–19",
        "url": "https://arxiv.org/html/2212.08648v3#S3.SS5"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-e-arc-03",
    "status": "addressed",
    "plainExplanation": "The Bell-number basis count requires enough nodes to realize every index-equality pattern; below that range, some proposed basis tensors vanish. Pearce-Crump’s later analysis explicitly derives the restricted count, states the condition n≥k+l, and explains how the earlier characterization must depend on n for smaller graphs.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "E-ARC-13",
    "title": "Theoretical Limitations of Self-Attention in Neural Sequence Models",
    "year": 2019,
    "category": "Transformers & sequence models",
    "primaryUrl": "https://arxiv.org/abs/1906.06755",
    "citationCount": 420,
    "citationUrl": "https://www.semanticscholar.org/paper/b3564be8b79f25585acb035f3deaf4ae93c26d8f",
    "citationDate": "2026-09-09",
    "issueTitle": "A missing attention-weight perturbation term",
    "addressingSources": [
      {
        "label": "Author supplement: Section S2, pages 8–9",
        "url": "https://mhahn.info/files/transformers-proof.pdf#page=8"
      }
    ],
    "resolutionType": "revision",
    "slug": "addressed-e-arc-13",
    "status": "addressed",
    "plainExplanation": "The soft-attention proof bounds changes in values but omits the contribution from changed attention weights. Hahn’s 2021 author supplement explicitly identifies that missing term and bounds it by O(1/n) under the paper’s boundedness and Lipschitz assumptions, completing that step.",
    "addressingRelationship": "later-version"
  },
  {
    "auditId": "IARC-021",
    "title": "On the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact Groups",
    "year": 2018,
    "category": "Graphs & equivariance",
    "primaryUrl": "https://arxiv.org/abs/1802.03690",
    "citationCount": 598,
    "citationUrl": "https://www.semanticscholar.org/paper/84032c19bad3493957d1319babd19bde2821fee3",
    "citationDate": "2026-09-09",
    "issueTitle": "Network equivariance does not force every weight matrix to commute",
    "addressingSources": [
      {
        "label": "Agrawal and Ostrowski: abstract and introduction, pages 1–2",
        "url": "https://jmlr.org/papers/volume24/23-0294/23-0294.pdf#page=1"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-iarc-021",
    "status": "addressed",
    "plainExplanation": "Equivariance on the hidden states reached by a network does not automatically make each given linear map equivariant on its entire ambient space. Agrawal and Ostrowski explicitly distinguish equivariant functions of the network input from equivariant individual weight matrices and construct networks using that distinction. Their later treatment addresses the conceptual issue, rather than issuing an erratum to the original convolution theorem.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IARC-027",
    "title": "GemNet: Universal Directional Graph Neural Networks for Molecules",
    "year": 2021,
    "category": "Molecular models & equivariance",
    "primaryUrl": "https://arxiv.org/abs/2106.08903",
    "citationCount": 651,
    "citationUrl": "https://www.semanticscholar.org/paper/5542d0ff99767f75f8c8a329fc3d88d73ff470c3",
    "citationDate": "2026-09-09",
    "issueTitle": "Spherical products omit reflection-odd interactions",
    "addressingSources": [
      {
        "label": "The Price of Freedom: Proposition 3.2, Section 6.2 and Appendix K",
        "url": "https://proceedings.mlr.press/v267/xie25c.html"
      },
      {
        "label": "Published paper: selection rule, page 5",
        "url": "https://raw.githubusercontent.com/mlresearch/v267/main/assets/xie25c/xie25c.pdf#page=5"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-iarc-027",
    "status": "addressed",
    "plainExplanation": "The spherical-product universality argument loses odd-parity Clebsch–Gordan couplings, including the cross-product channel, so it does not cover all rotation-invariant scalar targets that change sign under reflection. The Price of Freedom explicitly derives this selection rule and explains the missing cross-product and pseudoscalar interactions for Gaunt products. It addresses the same mathematical obstruction through later analysis, not an explicit GemNet erratum.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IARC-038",
    "title": "FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling",
    "year": 2018,
    "category": "Graphs & sampling",
    "primaryUrl": "https://arxiv.org/abs/1801.10247",
    "citationCount": 1768,
    "citationUrl": "https://www.semanticscholar.org/paper/2503dff90685857ce7295e37d0045e2eef41c8b8",
    "citationDate": "2026-09-09",
    "issueTitle": "Shared sampled nodes create covariance between outputs",
    "addressingSources": [
      {
        "label": "Neighborhood Sampling Does Not Learn the Same Graph Neural Network: Theorem 4, equations 16–18",
        "url": "https://arxiv.org/html/2509.22868#S4.SS1"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-iarc-038",
    "status": "addressed",
    "plainExplanation": "FastGCN’s variance derivation omits covariance between outputs that reuse the same inner node sample. Niu, Anitescu and Chen’s later analysis retains these cross-node second moments for layer-wise sampling in its covariance and neural-tangent-kernel formulas. This addresses the missing mechanism in a Gaussian-process analysis; it is not an explicit correction of FastGCN’s variance proposition or its sampling optimum.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IARC-060",
    "title": "Axiomatic Attribution for Deep Networks",
    "year": 2017,
    "category": "Explainability",
    "primaryUrl": "https://arxiv.org/abs/1703.01365",
    "citationCount": 8651,
    "citationUrl": "https://www.semanticscholar.org/paper/f302e136c41db5de1d624412f68c9174cf7ae8be",
    "citationDate": "2026-09-09",
    "issueTitle": "Integrated-gradients axioms require additional mathematical conditions",
    "addressingSources": [
      {
        "label": "Symmetry-Preserving Paths in Integrated Gradients: Sections 1.2 and 2",
        "url": "https://arxiv.org/html/2103.13533"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-iarc-060",
    "status": "addressed",
    "plainExplanation": "The original completeness argument needs stronger regularity along the integration path, and the symmetry-based path-uniqueness statement needs qualification. Lerma and Lucas explicitly discuss these proof issues, give a gradient theorem with Lipschitz and path-differentiability assumptions, and formulate the symmetry result precisely. This treatment clarifies the guarantees without invalidating ordinary straight-line integrated gradients on admissible paths.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "MBIO-ATOM-002",
    "title": "TorchMD-NET: Equivariant Transformers for Neural Network based Molecular Potentials",
    "year": 2022,
    "category": "Molecular models & equivariance",
    "primaryUrl": "https://arxiv.org/abs/2202.02541",
    "citationCount": 303,
    "citationUrl": "https://www.semanticscholar.org/paper/68ab47dbf07dcff861548ee37c906b6793dfd75c",
    "citationDate": "2026-09-09",
    "issueTitle": "A vector message can remain nonzero at the distance cutoff",
    "addressingSources": [
      {
        "label": "TorchMD-Net 0.14.1: vector_cutoff documentation",
        "url": "https://torchmd-net.readthedocs.io/en/0.14.1/models.html#equivariant-transformer"
      },
      {
        "label": "Author source: cutoff diagnosis and optional vector branch",
        "url": "https://github.com/torchmd/torchmd-net/blob/main/torchmdnet/models/torchmd_et.py"
      }
    ],
    "resolutionType": "implementation",
    "slug": "addressed-mbio-atom-002",
    "status": "addressed",
    "plainExplanation": "Value-projection biases can leave nonzero vector messages at the distance cutoff, allowing an energy jump when an edge disappears. The official implementation explicitly diagnoses this mechanism and provides vector_cutoff to apply the cutoff to the vector pathway. The fix is optional and defaults to false in the linked version; rotational equivariance itself is not the issue.",
    "addressingRelationship": "separate-code-source"
  },
  {
    "auditId": "E-GEN-20",
    "title": "Representation Learning with Contrastive Predictive Coding",
    "year": 2018,
    "primaryUrl": "https://arxiv.org/abs/1807.03748",
    "citationCount": 14412,
    "citationUrl": "https://www.semanticscholar.org/paper/b227f3e4c0dc96e5ac5426b85485a70f2175a205",
    "citationDate": "2026-09-09",
    "category": "Representation learning",
    "issueTitle": "Finite-minibatch InfoNCE lower-bound derivation",
    "resolutionType": "later-work",
    "addressingSources": [
      {
        "label": "On Variational Bounds of Mutual Information — Section 2.3 and appendix derivation",
        "url": "https://arxiv.org/abs/1905.06922"
      }
    ],
    "slug": "addressed-e-gen-20",
    "status": "addressed",
    "plainExplanation": "CPC's appendix replaces a random negative-sample sum inside a logarithm by its expectation, which does not establish the stated finite-minibatch mutual-information bound. Poole et al. give a valid variational derivation of the InfoNCE lower bound, addressing the proof mechanism while preserving the bound itself.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "O-2303.01469",
    "title": "Consistency Models",
    "year": 2023,
    "primaryUrl": "https://proceedings.mlr.press/v202/song23a/song23a.pdf",
    "citationCount": 2155,
    "citationUrl": "https://www.semanticscholar.org/paper/ac974291d7e3a152067382675524f3e3c2ded11b",
    "citationDate": "2026-09-09",
    "category": "Generative models",
    "issueTitle": "EMA teacher mismatch in the consistency-training limit",
    "resolutionType": "later-work",
    "addressingSources": [
      {
        "label": "Improved Techniques for Training Consistency Models — Proposition 1 and Section 3.2",
        "url": "https://proceedings.iclr.cc/paper_files/paper/2024/file/41bd71e7bf7f9fe68f1c936940fd06bd-Paper-Conference.pdf"
      }
    ],
    "slug": "addressed-o-2303-01469",
    "status": "addressed",
    "plainExplanation": "The consistency-training limit can lose its dependence on the data when the moving-average teacher differs from the student; this is separate from the consistency-distillation error proof. The authors' follow-up analyzes an explicit example in Proposition 1 and removes the teacher's exponential moving average to recover the intended learning signal.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IG002",
    "title": "Importance Weighted Autoencoders",
    "year": 2015,
    "primaryUrl": "https://arxiv.org/pdf/1509.00519",
    "citationCount": 1369,
    "citationUrl": "https://www.semanticscholar.org/paper/3e47c4c2dd98c49b7771c7228812d5fd9eee56a3",
    "citationDate": "2026-09-09",
    "category": "Generative models",
    "issueTitle": "Expected-log convergence requires tail control",
    "resolutionType": "later-work",
    "addressingSources": [
      {
        "label": "Filtering Variational Objectives — Proposition 1(b) and proof",
        "url": "https://arxiv.org/abs/1705.09279"
      },
      {
        "label": "Hierarchical Importance Weighted Autoencoders — supplement, Section 4 and Appendix C",
        "url": "https://proceedings.mlr.press/v97/huang19d/huang19d-supp.pdf"
      }
    ],
    "slug": "addressed-ig002",
    "status": "addressed",
    "plainExplanation": "Bounded positive importance weights alone do not justify convergence of the expected logarithm: sufficiently heavy lower tails can leave every finite-sample objective at minus infinity. Filtering Variational Objectives supplies uniform integrability of the logarithms as a sufficient condition, and Hierarchical IWAE explicitly discusses this missing ingredient; the lower-bound and monotonicity clauses are unaffected.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IG013",
    "title": "GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium",
    "year": 2017,
    "primaryUrl": "https://arxiv.org/pdf/1706.08500",
    "citationCount": 514,
    "citationUrl": "https://www.semanticscholar.org/paper/56f5005c4be6f816f6f43795cc4825d798cd53ef",
    "citationDate": "2026-09-09",
    "category": "Generative models",
    "issueTitle": "Two-timescale stability does not imply a Nash equilibrium",
    "resolutionType": "later-work",
    "addressingSources": [
      {
        "label": "Implicit Learning Dynamics in Stackelberg Games — stable non-Nash equilibria and timescale separation",
        "url": "https://proceedings.mlr.press/v119/fiez20a/fiez20a.pdf"
      },
      {
        "label": "Adversarial Monte Carlo Meta-Learning of Optimal Prediction Procedures — discussion, p. 54",
        "url": "https://jmlr.org/papers/volume22/20-1065/20-1065.pdf"
      }
    ],
    "slug": "addressed-ig013",
    "status": "addressed",
    "plainExplanation": "A stable fast-player equilibrium and a stable reduced slow-player dynamic can converge to a point that is not a local Nash equilibrium. Later Stackelberg analyses distinguish these equilibrium notions and explicitly discuss two-timescale learning, addressing the inference from convergence to Nash optimality without invalidating convergence to the reduced dynamic's attractor.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IG055",
    "title": "Domain-Adversarial Training of Neural Networks",
    "year": 2015,
    "primaryUrl": "https://arxiv.org/pdf/1505.07818",
    "citationCount": 12117,
    "citationUrl": "https://www.semanticscholar.org/paper/1d5972b32a9b5a455a6eef389de5b7fca25771ad",
    "citationDate": "2026-09-09",
    "category": "Domain adaptation",
    "issueTitle": "H-divergence does not automatically control disagreement",
    "resolutionType": "later-work",
    "addressingSources": [
      {
        "label": "Bridging Theory and Algorithm for Domain Adaptation — supplement, Proposition A.3",
        "url": "https://proceedings.mlr.press/v97/zhang19i/zhang19i-supp.pdf"
      }
    ],
    "slug": "addressed-ig055",
    "status": "addressed",
    "plainExplanation": "The target-risk argument requires disagreement functions from HΔH, whereas the printed guarantee uses H-divergence without the needed class relationship. The MDD supplement supplies HΔH=H as a sufficient closure condition for equality of the divergences and notes that neural classes generally lack it; this is a later qualification of the mechanism, not an explicit DANN erratum.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IG059",
    "title": "Maximum Classifier Discrepancy for Unsupervised Domain Adaptation",
    "year": 2017,
    "primaryUrl": "https://arxiv.org/pdf/1712.02560",
    "citationCount": 2030,
    "citationUrl": "https://www.semanticscholar.org/paper/0d725e4fea8bbaf332d6a8d424ebecbd547a3851",
    "citationDate": "2026-09-09",
    "category": "Domain adaptation",
    "issueTitle": "Unqualified replacement of HΔH by H-divergence",
    "resolutionType": "later-work",
    "addressingSources": [
      {
        "label": "Bridging Theory and Algorithm for Domain Adaptation — supplement, Proposition A.3",
        "url": "https://proceedings.mlr.press/v97/zhang19i/zhang19i-supp.pdf"
      }
    ],
    "slug": "addressed-ig059",
    "status": "addressed",
    "plainExplanation": "The first disagreement-based target-risk bound is valid, but the subsequent comparison with H-divergence does not hold for arbitrary hypothesis classes. The MDD supplement states a sufficient class-closure condition and explains why neural classes need not satisfy it; this addresses the missing qualification without claiming that MCD's practical discrepancy objective is invalid.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "BIO-GEN-04",
    "title": "GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation",
    "year": 2020,
    "primaryUrl": "https://arxiv.org/pdf/2001.09382",
    "citationCount": 568,
    "citationUrl": "https://www.semanticscholar.org/paper/036d743c7ca1e513adf0a91594fc8111e03dc30c",
    "citationDate": "2026-09-09",
    "category": "Molecular generation",
    "issueTitle": "BFS level width does not bound every edge-index gap",
    "resolutionType": "implementation",
    "addressingSources": [
      {
        "label": "GraphAF README — link to the authors' TorchDrug implementation",
        "url": "https://github.com/DeepGraphLearning/GraphAF/blob/master/README.md"
      },
      {
        "label": "TorchDrug AutoregressiveGeneration — max_edge_unroll and preprocessing",
        "url": "https://github.com/DeepGraphLearning/torchdrug/blob/master/torchdrug/tasks/generation.py"
      }
    ],
    "slug": "addressed-bio-gen-04",
    "status": "addressed",
    "plainExplanation": "GraphAF's level-width argument does not justify an equally sized autoregressive lookback window: an edge can span more positions than the largest BFS level. The authors' linked TorchDrug implementation measures actual endpoint-index gaps after BFS instead, addressing this mechanism for the measured orderings. That measurement does not certify every possible future ordering, and no dataset failure is asserted.",
    "addressingRelationship": "separate-code-source"
  },
  {
    "auditId": "BIO-GEN-07",
    "title": "Junction Tree Variational Autoencoder for Molecular Graph Generation",
    "year": 2018,
    "primaryUrl": "https://arxiv.org/pdf/1802.04364",
    "citationCount": 1716,
    "citationUrl": "https://www.semanticscholar.org/paper/fd17bd9a5dc24a081ad9743570f50dd6750f54b2",
    "citationDate": "2026-09-09",
    "category": "Molecular generation",
    "issueTitle": "Blanket singleton insertion can break running intersection",
    "resolutionType": "implementation",
    "addressingSources": [
      {
        "label": "Authors' released tree_decomp implementation — bond/ring branching and overlap weights",
        "url": "https://github.com/wengong-jin/icml18-jtnn/blob/master/fast_jtnn/chemutils.py"
      }
    ],
    "slug": "addressed-bio-gen-07",
    "status": "addressed",
    "plainExplanation": "The printed junction-tree procedure can separate two fused rings through a singleton that omits another shared atom, violating running intersection when a pendant bond is also present. The released tree_decomp code distinguishes bond and ring cases and uses overlap-cardinality edge weights, avoiding that exact counterexample. This is implementation-level addressing, not a proof that every molecular graph is decomposed correctly.",
    "addressingRelationship": "separate-code-source"
  },
  {
    "auditId": "BIO-ML-01",
    "title": "Pre-training Molecular Graph Representation with 3D Geometry",
    "year": 2022,
    "primaryUrl": "https://arxiv.org/abs/2110.07728v2",
    "citationCount": 489,
    "citationUrl": "https://www.semanticscholar.org/paper/b8f816e23ff40d6afabccca2ee4770087ef0ef57",
    "citationDate": "2026-09-09",
    "category": "Molecular representation learning",
    "issueTitle": "Differential-entropy offsets and β-weighted ELBO claims",
    "resolutionType": "later-work",
    "addressingSources": [
      {
        "label": "A Group Symmetric Stochastic Differential Equation Model for Molecule Multi-modal Pretraining — Appendix F, Eqs. (65), (68)–(69)",
        "url": "https://proceedings.mlr.press/v202/liu23h/liu23h.pdf"
      }
    ],
    "slug": "addressed-bio-ml-01",
    "status": "addressed",
    "plainExplanation": "GraphMVP's mutual-information derivation needs entropy offsets for continuous variables, and reducing the KL weight below one need not preserve an evidence lower bound. The authors' later MoleculeSDE appendix includes possibly negative entropy lower constants, keeps unit KL weight in its ELBO, and labels the β-weighted reconstruction loss a proxy objective. These distinctions address the selected formulas without disputing pretraining utility.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "BIO-ML-03",
    "title": "Graph Contrastive Learning with Augmentations",
    "year": 2020,
    "primaryUrl": "https://arxiv.org/abs/2010.13902v3",
    "citationCount": 2978,
    "citationUrl": "https://www.semanticscholar.org/paper/76c124786ccf4263e6403a15a8e350ac28be4e65",
    "citationDate": "2026-09-09",
    "category": "Molecular representation learning",
    "issueTitle": "Negative-only denominator is not the standard InfoNCE bound",
    "resolutionType": "discussion",
    "addressingSources": [
      {
        "label": "GraphCL issue 27 — discrepancy with the InfoNCE denominator",
        "url": "https://github.com/Shen-Lab/GraphCL/issues/27"
      },
      {
        "label": "GraphCL issue 2 — earlier question about positive-pair subtraction",
        "url": "https://github.com/Shen-Lab/GraphCL/issues/2"
      }
    ],
    "slug": "addressed-bio-ml-03",
    "status": "addressed",
    "plainExplanation": "GraphCL's finite-batch negative-only denominator does not justify the supplement's conversion to a population mutual-information bound. Public GitHub issues identify the positive-pair subtraction and its inconsistency with the standard InfoNCE denominator. This discussion addresses the mechanism; it is not a verified theorem repair or a claim that the empirical contrastive objective is unusable.",
    "addressingRelationship": "separate-discussion"
  },
  {
    "auditId": "E-OPT-13",
    "title": "Deep Learning without Poor Local Minima",
    "year": 2016,
    "category": "Optimization & generalization",
    "primaryUrl": "https://arxiv.org/pdf/1605.07110v3",
    "citationCount": 1015,
    "citationUrl": "https://www.semanticscholar.org/paper/07925910d45761d96269fc3bdfdc21b1d20d84ad",
    "citationDate": "2026-09-09",
    "issueTitle": "Overbroad zero-Hessian saddle family",
    "addressingSources": [
      {
        "label": "Later saddle classification: Theorem 7",
        "url": "https://www.jmlr.org/papers/volume25/23-0493/23-0493.pdf"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-e-opt-13",
    "status": "addressed",
    "plainExplanation": "The proof of Corollary 2.4 includes an all-but-one-zero weight family whose Hessian need not vanish; the three-factor scalar example (0,0,1) is a strict saddle. Achour, Malgouyres and Gerchinovitz's later Theorem 7 correctly classifies these points. This is a later mathematical treatment, not an explicit erratum, and the original existence of non-strict saddles survives.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "E-RL-02",
    "title": "Trajectory balance: Improved credit assignment in GFlowNets",
    "year": 2022,
    "category": "Reinforcement learning",
    "primaryUrl": "https://arxiv.org/pdf/2201.13259",
    "citationCount": 302,
    "citationUrl": "https://www.semanticscholar.org/paper/0116eb4208e9abff3064934d178e655e716d92ad",
    "citationDate": "2026-09-09",
    "issueTitle": "Expected-gradient notation and variance-baseline sign",
    "addressingSources": [
      {
        "label": "Revised Appendix A.3: gradient inside expectation",
        "url": "https://arxiv.org/pdf/2201.13259v4"
      },
      {
        "label": "Expected-gradient identities: Proposition 1",
        "url": "https://arxiv.org/pdf/2210.00580v4"
      },
      {
        "label": "Corrected variance comparison: Appendix B.2",
        "url": "https://arxiv.org/pdf/2605.15417"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-e-rl-02",
    "status": "addressed",
    "plainExplanation": "Older Appendix A.3 notation conflated the expected trajectory gradient with the gradient of an expected loss; v4 makes the former explicit, as does Proposition 1 of GFlowNets and variational inference. The appendix also reverses the log-ratio sign in a variance comparison; f-Trajectory Balance, Appendix B.2, supplies the corrected expression and a local variance guarantee. These address two auxiliary derivations, not the core trajectory-balance sampling identity.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "E-RL-08",
    "title": "ORPO: Monolithic Preference Optimization without Reference Model",
    "year": 2024,
    "category": "Reinforcement learning",
    "primaryUrl": "https://arxiv.org/pdf/2403.07691",
    "citationCount": 689,
    "citationUrl": "https://www.semanticscholar.org/paper/973814cd535facbf4f27c3de477b05bf19366030",
    "citationDate": "2026-09-09",
    "issueTitle": "Sign error in the odds-ratio loss gradient",
    "addressingSources": [
      {
        "label": "Public discussion of the missing minus sign; not an erratum",
        "url": "https://www.alphaxiv.org/abs/2403.07691v1"
      },
      {
        "label": "Later ORPO gradient derivation: Appendix E",
        "url": "https://arxiv.org/pdf/2502.01237"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-e-rl-08",
    "status": "addressed",
    "plainExplanation": "The manuscript's gradient derivation has a missing leading minus sign and inconsistent intermediate derivatives of log(1−p). Public discussion identifies the appendix sign issue, and The Differences Between Direct Alignment Algorithms are a Blur, Appendix E, derives the correctly signed ORPO gradient. This does not establish that the training implementation differentiated the loss incorrectly.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "O-1710.10345",
    "title": "The Implicit Bias of Gradient Descent on Separable Data",
    "year": 2017,
    "category": "Optimization & generalization",
    "primaryUrl": "https://arxiv.org/pdf/1710.10345v7",
    "citationCount": 1160,
    "citationUrl": "https://www.semanticscholar.org/paper/11adc8bd35bd897502f9b5452ab7ac668ec9b0fb",
    "citationDate": "2026-09-09",
    "issueTitle": "Missing compatibility assumption in the multiclass limit",
    "addressingSources": [
      {
        "label": "Authors' revised multiclass Theorem 7",
        "url": "https://arxiv.org/pdf/1710.10345v7"
      }
    ],
    "resolutionType": "revision",
    "slug": "addressed-o-1710-10345",
    "status": "addressed",
    "plainExplanation": "The multiclass residual-limit argument needs a compatible solution to its support-vector equations; separability alone does not supply that compatibility. The authors' v7 explicitly adds the assumption to Theorem 7. The issue concerns the multiclass extension, not a refutation of the main binary implicit-bias result.",
    "addressingRelationship": "later-version"
  },
  {
    "auditId": "IR005",
    "title": "A Distributional Perspective on Reinforcement Learning",
    "year": 2017,
    "category": "Reinforcement learning",
    "primaryUrl": "https://arxiv.org/abs/1707.06887",
    "citationCount": 1952,
    "citationUrl": "https://www.semanticscholar.org/paper/c1f4ef741242d629d1f56e442a09a7ba29595a0e",
    "citationDate": "2026-09-09",
    "issueTitle": "Omitted variance of next-state conditional means",
    "addressingSources": [
      {
        "label": "Authors' total-variance derivation: Section 5.4, Eq. (5.19)",
        "url": "https://www.distributional-rl.org/contents/chapter5"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-ir005",
    "status": "addressed",
    "plainExplanation": "The appendix's variance-contraction argument drops the variance of the return's conditional mean across successor states. The authors' Distributional Reinforcement Learning book restores that term in Section 5.4, equation (5.19), through the law of total variance. This is a later correct derivation rather than a labelled erratum; the paper's Wasserstein contraction result is unaffected.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IR022",
    "title": "Learning Robust Rewards with Adversarial Inverse Reinforcement Learning",
    "year": 2017,
    "category": "Reinforcement learning",
    "primaryUrl": "https://arxiv.org/abs/1710.11248",
    "citationCount": 882,
    "citationUrl": "https://www.semanticscholar.org/paper/5e2c4e7b3302549b3718601c44d9af6c7554efef",
    "citationDate": "2026-09-09",
    "issueTitle": "Stochastic reward recovery needs kernel completeness",
    "addressingSources": [
      {
        "label": "Later stochastic identification analysis: Section 3.3",
        "url": "https://arxiv.org/pdf/2605.30843"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-ir022",
    "status": "addressed",
    "plainExplanation": "The deterministic recovery proof equates successor potential values, but equality of their expectations under stochastic transitions does not force those values to be constant. The later analysis in Section 3.3, Proposition 3.8 and equations (43)–(46), explicitly supplies a transition-kernel completeness requirement. This addresses the stochastic extension; the original deterministic argument survives.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IO-1607.00133",
    "title": "Deep Learning with Differential Privacy",
    "year": 2016,
    "category": "Optimization & generalization",
    "primaryUrl": "https://arxiv.org/pdf/1607.00133v2",
    "citationCount": 8326,
    "citationUrl": "https://www.semanticscholar.org/paper/e9a986c8ff6c2f381d026fe014f6aaa865f34da7",
    "citationDate": "2026-09-09",
    "issueTitle": "Independence shortcut in adaptive privacy composition",
    "addressingSources": [
      {
        "label": "Conditional composition proof: Proposition 1, pp4–5",
        "url": "https://arxiv.org/pdf/1702.07476v3"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-io-1607-00133",
    "status": "addressed",
    "plainExplanation": "The adaptive-composition proof factors an expectation of a product even though later privacy losses can depend on earlier released outputs; independent noise does not make those losses independent. Mironov's Rényi Differential Privacy, Proposition 1, supplies the correct conditional-density argument. The composition theorem can be recovered without that independence shortcut.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IO-1802.08908",
    "title": "Scalable Private Learning with PATE",
    "year": 2018,
    "category": "Optimization & generalization",
    "primaryUrl": "https://arxiv.org/pdf/1802.08908v1",
    "citationCount": 734,
    "citationUrl": "https://www.semanticscholar.org/paper/44058a625cb64c311043145655645d8206e272c2",
    "citationDate": "2026-09-09",
    "issueTitle": "Missing positivity guard in the RDP bound",
    "addressingSources": [
      {
        "label": "Author code: positivity check, lines 176–179",
        "url": "https://github.com/tensorflow/privacy/blob/3d038a490a0cdd809c030435281c2e94e44c2f83/research/pate_2018/core.py#L176-L179"
      }
    ],
    "resolutionType": "implementation",
    "slug": "addressed-io-1802-08908",
    "status": "addressed",
    "plainExplanation": "Theorem 6 omits a positivity condition needed for its likely-outcome denominator, allowing a nonsensical negative privacy upper bound under the printed assumptions. The author implementation explicitly checks that condition before using the refined bound. This is an implementation safeguard, not evidence that the released PATE accountant underestimates privacy loss.",
    "addressingRelationship": "separate-code-source"
  },
  {
    "auditId": "IO-1902.09843",
    "title": "Adaptive Gradient Methods with Dynamic Bound of Learning Rate",
    "year": 2019,
    "category": "Optimization & generalization",
    "primaryUrl": "https://arxiv.org/pdf/1902.09843v1",
    "citationCount": 696,
    "citationUrl": "https://www.semanticscholar.org/paper/03af562fb8e69677865dbe94910e464443dd4623",
    "citationDate": "2026-09-09",
    "issueTitle": "Unjustified inverse-step-size monotonicity",
    "addressingSources": [
      {
        "label": "Savarese's correction and repaired conditions",
        "url": "https://arxiv.org/pdf/1908.04457"
      }
    ],
    "resolutionType": "later-work",
    "slug": "addressed-io-1902-09843",
    "status": "addressed",
    "plainExplanation": "The regret proof treats inverse adaptive step sizes as nondecreasing, although the stated clipping assumptions do not ensure this; the resulting signed telescoping terms cannot be bounded as written. Savarese's On the Convergence of AdaBound and its Connection to SGD analyzes the defect, gives a slow-convergence counterexample and supplies stronger sufficient conditions. This is an explicit later mathematical correction.",
    "addressingRelationship": "different-work"
  },
  {
    "auditId": "IO-1801.09344",
    "title": "Certified Defenses against Adversarial Examples",
    "year": 2018,
    "category": "Optimization & generalization",
    "primaryUrl": "https://arxiv.org/pdf/1801.09344v2",
    "citationCount": 1021,
    "citationUrl": "https://www.semanticscholar.org/paper/966e3c7a65ec75a6359b55c0cecaf3896d318432",
    "citationDate": "2026-09-09",
    "issueTitle": "Misplaced constant block in the SDP matrix",
    "addressingSources": [
      {
        "label": "Author code-bounds bundle: bounds_sdp.py, lines 23–40",
        "url": "https://worksheets.codalab.org/worksheets/0xa21e794020bb474d8804ec7bc0543f52"
      }
    ],
    "resolutionType": "implementation",
    "slug": "addressed-io-1801-09344",
    "status": "addressed",
    "plainExplanation": "The printed SDP matrix places the constant term in the wrong block, encoding a different quadratic form and permitting an understated adversarial-score bound. The official code-bounds implementation constructs the correct block placement. The paper's displayed matrix issue therefore does not establish a failure of the implemented certificates.",
    "addressingRelationship": "separate-code-source"
  }
]