{"paper":{"title":"On the Unreasonable Effectiveness of Last-layer Retraining","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Last-layer retraining succeeds mainly because the held-out set has better group balance than the full training data.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"John C. Hill, Tyler LaBonte, Vidya Muthukumar, Xinchen Zhang","submitted_at":"2025-12-01T15:08:43Z","abstract_excerpt":"Last-layer retraining (LLR) methods -- wherein the last layer of a neural network is reinitialized and retrained on a held-out set following ERM training -- have garnered interest as an efficient approach to rectify dependence on spurious correlations and improve performance on minority groups. Surprisingly, LLR has been found to improve worst-group accuracy even when the held-out set is an imbalanced subset of the training set. We initially hypothesize that this ``unreasonable effectiveness'' of LLR is explained by its ability to mitigate neural collapse through the held-out set, resulting in"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We present strong evidence for an alternative hypothesis: that the success of LLR is primarily due to better group balance in the held-out set.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the empirical tests on neural collapse are conclusive enough to rule it out as a contributing factor, and that group balance differences are the dominant cause rather than other unmeasured effects.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Last-layer retraining succeeds primarily because the held-out set has better group balance, supported by experiments that do not back the neural collapse hypothesis.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Last-layer retraining succeeds mainly because the held-out set has better group balance than the full training data.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0b7b3eb0347cea5ea0b1bf0d3f0057551bbffbe3ec6c61c48e1da7a07544aaf7"},"source":{"id":"2512.01766","kind":"arxiv","version":2},"verdict":{"id":"f57ac31a-9834-42de-a042-3529860fa6c4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T02:39:21.315834Z","strongest_claim":"We present strong evidence for an alternative hypothesis: that the success of LLR is primarily due to better group balance in the held-out set.","one_line_summary":"Last-layer retraining succeeds primarily because the held-out set has better group balance, supported by experiments that do not back the neural collapse hypothesis.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the empirical tests on neural collapse are conclusive enough to rule it out as a contributing factor, and that group balance differences are the dominant cause rather than other unmeasured effects.","pith_extraction_headline":"Last-layer retraining succeeds mainly because the held-out set has better group balance than the full training data."},"references":{"count":4,"sample":[{"doi":"10.1073/pnas.2103091118.url:http://dx.doi.org/10.1073/","year":1998,"title":"Boosting the margin: A new explanation for the effectiveness of voting methods","work_id":"18928c48-673b-4b24-8e73-8e2efe177a6d","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"A broad-coverage challenge corpus for sentence understanding through inference","work_id":"ab1ad3fa-51ee-4b7d-ba96-8f51078051f8","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1992,"title":"Note that Waterbirds is the only dataset that has a distribution shift and MultiNLI is the only dataset which is class-balanceda priori","work_id":"2f29390d-9bd5-42f4-9409-90fa8f21e077","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"These pretrained models are used as the initialization for ERM finetuning under the cross-entropy loss","work_id":"279fb77b-4749-4b2f-93d8-9068c2720b07","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":4,"snapshot_sha256":"cce3bac278234ceef620b01009cf6196a9d54eef76e558c96f07bb8e267df06d","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"02ffadf4474dddc8cb5b9e114b7b7cc721cf2646dcbd60582f5b320aab2fcbb6"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}