{"paper":{"title":"Learning in the Fisher Subspace: A Guided Initialization for LoRA Fine-Tuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Using Fisher curvature from downstream data to initialize LoRA subspaces improves fine-tuning performance over weight-only methods.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Hung-Yu Kao, Ying-Jia Lin, Zhi-Quan Feng","submitted_at":"2026-05-01T19:20:25Z","abstract_excerpt":"LoRA adapts large language models (LLMs) by restricting updates to low-rank subspaces of pre-trained weights. While this substantially reduces training cost, the effectiveness of adaptation critically depends on which subspace is chosen at initialization: a poor initialization that allocates capacity to task-irrelevant directions can severely hinder downstream performance. Existing initialization strategies primarily rely on the intrinsic properties of pre-trained weights, implicitly assuming that weight geometry alone reflects task relevance. However, such criteria overlook how the model inte"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Empirical results across diverse tasks and modalities demonstrate that data-aware initialization consistently and significantly improves downstream performance over existing approaches.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the curvature information induced by the downstream data distribution accurately identifies parameter directions whose perturbations most influence model predictions on the target objective.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Fisher information from target data provides a better criterion than weight geometry for choosing LoRA subspaces, yielding consistent performance gains on downstream tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Using Fisher curvature from downstream data to initialize LoRA subspaces improves fine-tuning performance over weight-only methods.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"4fe263de3d7f2221f758ef3207f7a5448cd5070aafd6700689665855ab8d5c6e"},"source":{"id":"2605.01046","kind":"arxiv","version":3},"verdict":{"id":"95a153b7-20d6-4c4b-97f8-c6e2655c7e5d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-09T19:30:10.890636Z","strongest_claim":"Empirical results across diverse tasks and modalities demonstrate that data-aware initialization consistently and significantly improves downstream performance over existing approaches.","one_line_summary":"Fisher information from target data provides a better criterion than weight geometry for choosing LoRA subspaces, yielding consistent performance gains on downstream tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the curvature information induced by the downstream data distribution accurately identifies parameter directions whose perturbations most influence model predictions on the target objective.","pith_extraction_headline":"Using Fisher curvature from downstream data to initialize LoRA subspaces improves fine-tuning performance over weight-only methods."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.01046/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T18:39:26.354215Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T17:40:54.459980Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"a18f734de0fb534ad988031bd386a8993ac4714d14d0fd0bf72e76c0f01ca21b"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}