{"paper":{"title":"Task-Driven Subspace Decomposition for Knowledge Sharing and Isolation in LoRA-based Continual Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"LoDA decomposes LoRA updates into shared general and task-specific subspaces via energy-based objectives to enable knowledge transfer without catastrophic forgetting.","cross_cats":["cs.CV"],"primary_cat":"cs.LG","authors_text":"De Cheng, Huaijie Wang, Lingfeng He, Nannan Wang, Xinbo Gao, Xi Yang","submitted_at":"2026-02-27T02:31:00Z","abstract_excerpt":"Continual Learning (CL) requires models to sequentially adapt to new tasks without forgetting old knowledge. Recently, Low-Rank Adaptation (LoRA), a representative Parameter-Efficient Fine-Tuning (PEFT) method, has gained increasing attention in CL. Several LoRA-based CL methods reduce interference across tasks by separating their update spaces, typically building the new space from the estimated null space of past tasks. However, they (i) overlook task-shared directions, which suppresses knowledge transfer, and (ii) fail to capture truly effective task-specific directions since these ``null b"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"LoDA outperforms existing CL methods by performing task-driven decomposition to build general and truly task-specific LoRA subspaces, fixing down-projections and using GAO plus closed-form recalibration for the general update.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That solving the two energy-based objectives will reliably produce effective shared and task-specific directions, and that the null-space limitations of prior methods are the primary bottleneck rather than other factors such as optimization dynamics or data distribution shifts.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LoDA introduces task-driven subspace decomposition in LoRA for continual learning to separate knowledge-sharing and isolation directions via energy objectives and closed-form recalibration, outperforming prior null-space methods.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LoDA decomposes LoRA updates into shared general and task-specific subspaces via energy-based objectives to enable knowledge transfer without catastrophic forgetting.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7c8ed58ac5108b75aff7fa0846d4f5367d92ca76946dfe310a1aae805db71049"},"source":{"id":"2603.00191","kind":"arxiv","version":4},"verdict":{"id":"d06e7625-f9bd-41a7-b0f4-9733442cc3e6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T18:57:44.672935Z","strongest_claim":"LoDA outperforms existing CL methods by performing task-driven decomposition to build general and truly task-specific LoRA subspaces, fixing down-projections and using GAO plus closed-form recalibration for the general update.","one_line_summary":"LoDA introduces task-driven subspace decomposition in LoRA for continual learning to separate knowledge-sharing and isolation directions via energy objectives and closed-form recalibration, outperforming prior null-space methods.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That solving the two energy-based objectives will reliably produce effective shared and task-specific directions, and that the null-space limitations of prior methods are the primary bottleneck rather than other factors such as optimization dynamics or data distribution shifts.","pith_extraction_headline":"LoDA decomposes LoRA updates into shared general and task-specific subspaces via energy-based objectives to enable knowledge transfer without catastrophic forgetting."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.00191/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":"9f2b9ada3c4306b0400aef4cfb5c269e4a13701c96cc272741c1871248e1619d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}