{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:K72MCI25QXEHS4C5GWIK23MQL6","short_pith_number":"pith:K72MCI25","schema_version":"1.0","canonical_sha256":"57f4c1235d85c879705d3590ad6d905fb2ff6fb2ef04439824ec6ec28206e406","source":{"kind":"arxiv","id":"2410.19278","version":2},"attestation_state":"computed","paper":{"title":"Applying sparse autoencoders to unlearn knowledge in language models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Arthur Conmy, Eoin Farrell, Yeu-Tong Lau","submitted_at":"2024-10-25T03:21:14Z","abstract_excerpt":"We investigate whether sparse autoencoders (SAEs) can be used to remove knowledge from language models. We use the biology subset of the Weapons of Mass Destruction Proxy dataset and test on the gemma-2b-it and gemma-2-2b-it language models. We demonstrate that individual interpretable biology-related SAE features can be used to unlearn a subset of WMDP-Bio questions with minimal side-effects in domains other than biology. Our results suggest that negative scaling of feature activations is necessary and that zero ablating features is ineffective. We find that intervening using multiple SAE fea"},"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":"2410.19278","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-10-25T03:21:14Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"ee40973cd944ca113c266d461bf3a7b9b56ff13b6ad5c8725b04dd500f275a24","abstract_canon_sha256":"aafa03a7ff6f11faa2fb338fcb454f97e588f2528f0a10829fb015c31087c3c2"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:30:23.085556Z","signature_b64":"7qscgeMlH7bG7Gsw7xMKfWYLmuSHm+zYzrQgFlgQro+dgvtyRD353fNv4MK73RKxQDG9LwWdMJkrb1vdLMqCDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"57f4c1235d85c879705d3590ad6d905fb2ff6fb2ef04439824ec6ec28206e406","last_reissued_at":"2026-07-05T09:30:23.085078Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:30:23.085078Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Applying sparse autoencoders to unlearn knowledge in language models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Arthur Conmy, Eoin Farrell, Yeu-Tong Lau","submitted_at":"2024-10-25T03:21:14Z","abstract_excerpt":"We investigate whether sparse autoencoders (SAEs) can be used to remove knowledge from language models. We use the biology subset of the Weapons of Mass Destruction Proxy dataset and test on the gemma-2b-it and gemma-2-2b-it language models. We demonstrate that individual interpretable biology-related SAE features can be used to unlearn a subset of WMDP-Bio questions with minimal side-effects in domains other than biology. Our results suggest that negative scaling of feature activations is necessary and that zero ablating features is ineffective. We find that intervening using multiple SAE fea"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.19278","kind":"arxiv","version":2},"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/2410.19278/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":"2410.19278","created_at":"2026-07-05T09:30:23.085138+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.19278v2","created_at":"2026-07-05T09:30:23.085138+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.19278","created_at":"2026-07-05T09:30:23.085138+00:00"},{"alias_kind":"pith_short_12","alias_value":"K72MCI25QXEH","created_at":"2026-07-05T09:30:23.085138+00:00"},{"alias_kind":"pith_short_16","alias_value":"K72MCI25QXEHS4C5","created_at":"2026-07-05T09:30:23.085138+00:00"},{"alias_kind":"pith_short_8","alias_value":"K72MCI25","created_at":"2026-07-05T09:30:23.085138+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09559","citing_title":"Safe-RULE: Safe Reinforcement UnLEarning","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02293","citing_title":"AI as a Tool for Simulation-Based Experiments in Literary Studies","ref_index":17,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/K72MCI25QXEHS4C5GWIK23MQL6","json":"https://pith.science/pith/K72MCI25QXEHS4C5GWIK23MQL6.json","graph_json":"https://pith.science/api/pith-number/K72MCI25QXEHS4C5GWIK23MQL6/graph.json","events_json":"https://pith.science/api/pith-number/K72MCI25QXEHS4C5GWIK23MQL6/events.json","paper":"https://pith.science/paper/K72MCI25"},"agent_actions":{"view_html":"https://pith.science/pith/K72MCI25QXEHS4C5GWIK23MQL6","download_json":"https://pith.science/pith/K72MCI25QXEHS4C5GWIK23MQL6.json","view_paper":"https://pith.science/paper/K72MCI25","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.19278&json=true","fetch_graph":"https://pith.science/api/pith-number/K72MCI25QXEHS4C5GWIK23MQL6/graph.json","fetch_events":"https://pith.science/api/pith-number/K72MCI25QXEHS4C5GWIK23MQL6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/K72MCI25QXEHS4C5GWIK23MQL6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/K72MCI25QXEHS4C5GWIK23MQL6/action/storage_attestation","attest_author":"https://pith.science/pith/K72MCI25QXEHS4C5GWIK23MQL6/action/author_attestation","sign_citation":"https://pith.science/pith/K72MCI25QXEHS4C5GWIK23MQL6/action/citation_signature","submit_replication":"https://pith.science/pith/K72MCI25QXEHS4C5GWIK23MQL6/action/replication_record"}},"created_at":"2026-07-05T09:30:23.085138+00:00","updated_at":"2026-07-05T09:30:23.085138+00:00"}