{"paper":{"title":"Sparse Memory Finetuning as a Low-Forgetting Alternative to LoRA and Full Finetuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Sparse Memory Finetuning updates only the most heavily read rows in added key-value layers to gain task performance while preserving general capabilities better than LoRA or full finetuning.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Anirudh Kanchi, Garv Shah, Prakhar Gupta, Satyam Goyal","submitted_at":"2026-05-04T23:46:40Z","abstract_excerpt":"Adapting a pretrained language model to a new task often hurts the general capabilities it already had, a problem known as catastrophic forgetting. Sparse Memory Finetuning (SMF) tries to avoid this by adding key-value memory layers to the model and, on each training step, updating only the small set of memory rows that the current batch reads most heavily. We re-implement SMF on Qwen-2.5-0.5B-Instruct and compare it with LoRA and full finetuning on MedMCQA, a 4-choice medical exam task, using WikiText perplexity and TriviaQA accuracy as forgetting probes. SMF improves MedMCQA by 2.5 percentag"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"SMF improves MedMCQA by 2.5 percentage points while keeping both forgetting probes within roughly 1 point of the base model, whereas LoRA and full finetuning achieve larger gains but with clear drift on both.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That selectively updating only the most heavily read memory rows during training is sufficient to acquire new task knowledge without unintended interference in unrelated general capabilities.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SMF improves MedMCQA accuracy by 2.5 points while keeping WikiText perplexity and TriviaQA accuracy within 1 point of the base model, outperforming LoRA and full finetuning on forgetting metrics.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Sparse Memory Finetuning updates only the most heavily read rows in added key-value layers to gain task performance while preserving general capabilities better than LoRA or full finetuning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f936b72255b9524f1282e03c8b02357a9ae1149b2ebc7e71d4028d621d89a9c2"},"source":{"id":"2605.03229","kind":"arxiv","version":2},"verdict":{"id":"919f7bab-02fa-49fa-9ecc-267ca47dbda6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T17:47:39.561915Z","strongest_claim":"SMF improves MedMCQA by 2.5 percentage points while keeping both forgetting probes within roughly 1 point of the base model, whereas LoRA and full finetuning achieve larger gains but with clear drift on both.","one_line_summary":"SMF improves MedMCQA accuracy by 2.5 points while keeping WikiText perplexity and TriviaQA accuracy within 1 point of the base model, outperforming LoRA and full finetuning on forgetting metrics.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That selectively updating only the most heavily read memory rows during training is sufficient to acquire new task knowledge without unintended interference in unrelated general capabilities.","pith_extraction_headline":"Sparse Memory Finetuning updates only the most heavily read rows in added key-value layers to gain task performance while preserving general capabilities better than LoRA or full finetuning."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.03229/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T14:35:19.247524Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T01:31:21.794774Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T15:34:04.280909Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"6dc37e4ce03a15165ed85bcb415ba044c98eea186f08319dd73876c48ed8e5a9"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"58dce1ef058b469a5bb57b12b9a65cce0e644b5409789449ba5ae62c061a9f55"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}