ProCreations/repro-when-more-data-doesn-t-help-limits-of-adaptation-in-multitask-learning

Reproduction - When More Data Doesn't Help: Limits of Adaptation in Multitask Learning

Clean-room CPU audit of all five anchored challenge claims and all three public challenge claims. The package checks the exact Section-5 rate algebra, constraint regimes, oracle fair-source ERM, the specialized pooling bound, a destructive pooling counterexample, and exact finite mixture likelihoods and KL divergences. All 15 predeclared gates pass in two warning-clean, byte-identical runs.

Two anchored claims are falsified as written because the live catalog reverses the manuscript's task/sample notation. The corrected Theorems 5.1 and 5.2 are audited separately; no nearby statement is substituted for the literal claim.

Paper: When More Data Doesn't Help: Limits of Adaptation in Multitask Learning Authors: Steve Hanneke; Mingyue Xu OpenReview: KFIW6LOeT1 arXiv: 2601.20774v1 Source PDF: https://arxiv.org/pdf/2601.20774v1 Source PDF SHA-256: 1bafab48a63861ee951106a3f6015587ab7261cea56210937c22132f6e80a9f1 Anchored claim-catalog ETag: "4093a21c968f9e7e6a33f495fe0e59bb37fe912d" Anchored claim-catalog SHA-256: eb3f2d878646ca5c40da121741a613c1f9e1e10c14845f373db107e74a4ef439 Public claim-catalog ETag: "f8b74517cd33712f80399eea913efac0bb41078f" Public claim-catalog SHA-256: af5ab2d62f786ae36861957cbd08b4188f6d4c86e67152becc661a9c5bbb9d57

Scope: finite simulation and exact small-instance enumeration support the mechanisms and expose literal/source mismatches; they do not replace the manuscript's general information-theoretic proofs.