{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:WMWYWDDBF3HE5NEVYEYN33D6GB","short_pith_number":"pith:WMWYWDDB","schema_version":"1.0","canonical_sha256":"b32d8b0c612ece4eb495c130ddec7e3067d4025eb277287e0fe3bdcd8f1fa47b","source":{"kind":"arxiv","id":"2504.04626","version":1},"attestation_state":"computed","paper":{"title":"Exact Unlearning of Finetuning Data via Model Merging at Scale","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Aditi Raghunathan, Amrith Setlur, Kartik Srinivas, Kevin Kuo, Virginia Smith","submitted_at":"2025-04-06T21:24:29Z","abstract_excerpt":"Approximate unlearning has gained popularity as an approach to efficiently update an LLM so that it behaves (roughly) as if it was not trained on a subset of data to begin with. However, existing methods are brittle in practice and can easily be attacked to reveal supposedly unlearned information. To alleviate issues with approximate unlearning, we instead propose SIFT-Masks (SIgn-Fixed Tuning-Masks), an exact unlearning method based on model merging. SIFT-Masks addresses two key limitations of standard model merging: (1) merging a large number of tasks can severely harm utility; and (2) metho"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2504.04626","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-04-06T21:24:29Z","cross_cats_sorted":[],"title_canon_sha256":"748ae16792a35b193649f72179352b802bdb6e3ca8fe3b5cad7a806a021ab108","abstract_canon_sha256":"f2b9fc18edf0bb0fbba251f22f5f329f95b3b2f3196e62d5c31f07ff8adf4dfe"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:45:27.295596Z","signature_b64":"6Vm6Jsa8WtuMuT7ZaE/FY0vvQsqLG/egsH7rsncufmCIfV8TZIoZO728doyn7AJ2NcvdxCoVtGmZ/brHiP2FAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b32d8b0c612ece4eb495c130ddec7e3067d4025eb277287e0fe3bdcd8f1fa47b","last_reissued_at":"2026-07-05T10:45:27.295110Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:45:27.295110Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Exact Unlearning of Finetuning Data via Model Merging at Scale","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Aditi Raghunathan, Amrith Setlur, Kartik Srinivas, Kevin Kuo, Virginia Smith","submitted_at":"2025-04-06T21:24:29Z","abstract_excerpt":"Approximate unlearning has gained popularity as an approach to efficiently update an LLM so that it behaves (roughly) as if it was not trained on a subset of data to begin with. However, existing methods are brittle in practice and can easily be attacked to reveal supposedly unlearned information. To alleviate issues with approximate unlearning, we instead propose SIFT-Masks (SIgn-Fixed Tuning-Masks), an exact unlearning method based on model merging. SIFT-Masks addresses two key limitations of standard model merging: (1) merging a large number of tasks can severely harm utility; and (2) metho"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.04626","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2504.04626/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":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2504.04626","created_at":"2026-07-05T10:45:27.295165+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.04626v1","created_at":"2026-07-05T10:45:27.295165+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.04626","created_at":"2026-07-05T10:45:27.295165+00:00"},{"alias_kind":"pith_short_12","alias_value":"WMWYWDDBF3HE","created_at":"2026-07-05T10:45:27.295165+00:00"},{"alias_kind":"pith_short_16","alias_value":"WMWYWDDBF3HE5NEV","created_at":"2026-07-05T10:45:27.295165+00:00"},{"alias_kind":"pith_short_8","alias_value":"WMWYWDDB","created_at":"2026-07-05T10:45:27.295165+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.24113","citing_title":"FedUP: One-Shot Federated Unlearning via Centroid-Guided Plug-in Filters","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2501.19202","citing_title":"Improving LLM Unlearning Robustness via Random Perturbations","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11170","citing_title":"Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data","ref_index":31,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/WMWYWDDBF3HE5NEVYEYN33D6GB","json":"https://pith.science/pith/WMWYWDDBF3HE5NEVYEYN33D6GB.json","graph_json":"https://pith.science/api/pith-number/WMWYWDDBF3HE5NEVYEYN33D6GB/graph.json","events_json":"https://pith.science/api/pith-number/WMWYWDDBF3HE5NEVYEYN33D6GB/events.json","paper":"https://pith.science/paper/WMWYWDDB"},"agent_actions":{"view_html":"https://pith.science/pith/WMWYWDDBF3HE5NEVYEYN33D6GB","download_json":"https://pith.science/pith/WMWYWDDBF3HE5NEVYEYN33D6GB.json","view_paper":"https://pith.science/paper/WMWYWDDB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.04626&json=true","fetch_graph":"https://pith.science/api/pith-number/WMWYWDDBF3HE5NEVYEYN33D6GB/graph.json","fetch_events":"https://pith.science/api/pith-number/WMWYWDDBF3HE5NEVYEYN33D6GB/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/WMWYWDDBF3HE5NEVYEYN33D6GB/action/timestamp_anchor","attest_storage":"https://pith.science/pith/WMWYWDDBF3HE5NEVYEYN33D6GB/action/storage_attestation","attest_author":"https://pith.science/pith/WMWYWDDBF3HE5NEVYEYN33D6GB/action/author_attestation","sign_citation":"https://pith.science/pith/WMWYWDDBF3HE5NEVYEYN33D6GB/action/citation_signature","submit_replication":"https://pith.science/pith/WMWYWDDBF3HE5NEVYEYN33D6GB/action/replication_record"}},"created_at":"2026-07-05T10:45:27.295165+00:00","updated_at":"2026-07-05T10:45:27.295165+00:00"}