{"paper":{"title":"LoRA-DA: Data-Aware Initialization for Low-Rank Adaptation via Asymptotic Analysis","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Minimizing the expected parameter discrepancy between fine-tuned and target models yields an optimal data-aware initialization for LoRA.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chang Chu, Qi Li, Qingyue Zhang, Shao-Lun Huang, Tianren Peng, Xiangyang Luo, Zhihao Jiang","submitted_at":"2025-10-28T15:55:36Z","abstract_excerpt":"LoRA has become a widely adopted method for PEFT, and its initialization methods have attracted increasing attention. However, existing methods have notable limitations: many methods do not incorporate target-domain data, while gradient-based methods exploit data only at a shallow level by relying on one-step gradient decomposition. In this paper, we establish a theoretical framework for data-aware LoRA initialization. Starting from minimizing the expectation of the parameter discrepancy between the fine-tuned and target models, we derive an optimization problem with two components: a bias ter"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Solving this problem yields an optimal initialization strategy for LoRA, based on which we develop an efficient algorithm, LoRA-DA. Empirical results across multiple benchmarks demonstrate that LoRA-DA consistently improves final accuracy over existing initialization methods.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The bias term is approximated using a Fisher-gradient formulation to preserve anisotropy while the variance term uses the Fisher information to capture sampling uncertainty; this approximation must hold for the derived initialization to be optimal in the target fine-tuning regime.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LoRA-DA derives an optimal data-aware LoRA initialization by solving an optimization problem from asymptotic analysis of parameter discrepancy using Fisher-gradient bias and Fisher-information variance terms.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Minimizing the expected parameter discrepancy between fine-tuned and target models yields an optimal data-aware initialization for LoRA.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f5d2579ad610b677589cf8e06b4f1aed8c7c49a51d09d0b2e63ae1edf92b9883"},"source":{"id":"2510.24561","kind":"arxiv","version":3},"verdict":{"id":"dfbdd8dc-eae1-4469-959a-c0a7071af2a3","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T02:58:14.357911Z","strongest_claim":"Solving this problem yields an optimal initialization strategy for LoRA, based on which we develop an efficient algorithm, LoRA-DA. Empirical results across multiple benchmarks demonstrate that LoRA-DA consistently improves final accuracy over existing initialization methods.","one_line_summary":"LoRA-DA derives an optimal data-aware LoRA initialization by solving an optimization problem from asymptotic analysis of parameter discrepancy using Fisher-gradient bias and Fisher-information variance terms.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The bias term is approximated using a Fisher-gradient formulation to preserve anisotropy while the variance term uses the Fisher information to capture sampling uncertainty; this approximation must hold for the derived initialization to be optimal in the target fine-tuning regime.","pith_extraction_headline":"Minimizing the expected parameter discrepancy between fine-tuned and target models yields an optimal data-aware initialization for LoRA."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2510.24561/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"7f21efc0f643fe16f221e875a524f596f1cdad16260413e46ef1097ae2074a7a"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}