Checking theoretical claims in ML
Automated mathematical claim checking with frontier LLMs.
How it works →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.
Statistical learningProposition 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…
Machine learningThe exact separation procedure never removes a dominated point. A certified frontier representation can supply a different oracle, while the weaker outer approximation remains available.
Optimization & generalizationThe proof optimizes over arbitrary critics, but a normalized policy restricts its induced critic. A two-response KL example has minimum loss where the theorem gives zero, and weak monotonicity does not ensure strict reward improvement. Two narrower corrections concern the…
Reinforcement learningOptimal 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.
Machine learningThe azimuth inverse has the wrong sine coefficient. The inverse-grid display is wrong; which mapping was used in execution remains undetermined.
Machine learningThe 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.
Optimization & generalizationIndependent 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.
Machine learningThe 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.
Representation learningThe fairness equivalence cannot divide by zero perturbation variance. The boundary-law equivalence is false; later uniform perturbation choices avoid this degeneracy.
Statistical learning