{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:OPGGFYYDJVAXIBWXSYQG2TZBXT","short_pith_number":"pith:OPGGFYYD","schema_version":"1.0","canonical_sha256":"73cc62e3034d417406d796206d4f21bcc33ce8679455a13448e27df3db4bcf90","source":{"kind":"arxiv","id":"2410.03415","version":2},"attestation_state":"computed","paper":{"title":"Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Barbara Plank, Chengzhi Hu, Paul R\\\"ottger, Xinpeng Wang","submitted_at":"2024-10-04T13:25:32Z","abstract_excerpt":"Training a language model to be both helpful and harmless requires careful calibration of refusal behaviours: Models should refuse to follow malicious instructions or give harmful advice (e.g.\"how do I kill someone?\"), but they should not refuse safe requests, even if they superficially resemble unsafe ones (e.g. \"how do I kill a Python process?\"). Avoiding such false refusal, as prior work has shown, is challenging even for highly-capable language models. In this paper, we propose a simple and surgical method for mitigating false refusal in language models via single vector ablation. For a gi"},"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":"2410.03415","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-10-04T13:25:32Z","cross_cats_sorted":[],"title_canon_sha256":"48c29298b0e71f2653c460efa777a02204db5183e3e7373d6d7177399e30c6ec","abstract_canon_sha256":"fae7690b6445f3880277fa506720f0791fe49d30d89b6973ad86d906b3bd7913"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:23:23.593102Z","signature_b64":"F07A/UJbRkXDVj34zN9Wkk5u8u9WO/qnpa9ViZZCdBAnknIM25U0PwX+do1PhbmjoqJkoP/z3tAINtALYQTRCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"73cc62e3034d417406d796206d4f21bcc33ce8679455a13448e27df3db4bcf90","last_reissued_at":"2026-07-05T10:23:23.592409Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:23:23.592409Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Barbara Plank, Chengzhi Hu, Paul R\\\"ottger, Xinpeng Wang","submitted_at":"2024-10-04T13:25:32Z","abstract_excerpt":"Training a language model to be both helpful and harmless requires careful calibration of refusal behaviours: Models should refuse to follow malicious instructions or give harmful advice (e.g.\"how do I kill someone?\"), but they should not refuse safe requests, even if they superficially resemble unsafe ones (e.g. \"how do I kill a Python process?\"). Avoiding such false refusal, as prior work has shown, is challenging even for highly-capable language models. In this paper, we propose a simple and surgical method for mitigating false refusal in language models via single vector ablation. For a gi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.03415","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/2410.03415/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":"2410.03415","created_at":"2026-07-05T10:23:23.592479+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.03415v2","created_at":"2026-07-05T10:23:23.592479+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.03415","created_at":"2026-07-05T10:23:23.592479+00:00"},{"alias_kind":"pith_short_12","alias_value":"OPGGFYYDJVAX","created_at":"2026-07-05T10:23:23.592479+00:00"},{"alias_kind":"pith_short_16","alias_value":"OPGGFYYDJVAXIBWX","created_at":"2026-07-05T10:23:23.592479+00:00"},{"alias_kind":"pith_short_8","alias_value":"OPGGFYYD","created_at":"2026-07-05T10:23:23.592479+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.05842","citing_title":"Beyond Refusal: A Same-Lineage Study of Aligned and Abliterated LLMs for Vulnerability Analysis","ref_index":51,"is_internal_anchor":true},{"citing_arxiv_id":"2606.21147","citing_title":"AOR-Bench: Do Large Audio Language Models Over-Refuse Pseudo-Harmful Queries?","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2605.24154","citing_title":"Palette: A Modular, Controllable, and Efficient Framework for On-demand Authorized Safety Alignment Relaxation in LLMs","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18104","citing_title":"Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift Correction","ref_index":39,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OPGGFYYDJVAXIBWXSYQG2TZBXT","json":"https://pith.science/pith/OPGGFYYDJVAXIBWXSYQG2TZBXT.json","graph_json":"https://pith.science/api/pith-number/OPGGFYYDJVAXIBWXSYQG2TZBXT/graph.json","events_json":"https://pith.science/api/pith-number/OPGGFYYDJVAXIBWXSYQG2TZBXT/events.json","paper":"https://pith.science/paper/OPGGFYYD"},"agent_actions":{"view_html":"https://pith.science/pith/OPGGFYYDJVAXIBWXSYQG2TZBXT","download_json":"https://pith.science/pith/OPGGFYYDJVAXIBWXSYQG2TZBXT.json","view_paper":"https://pith.science/paper/OPGGFYYD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.03415&json=true","fetch_graph":"https://pith.science/api/pith-number/OPGGFYYDJVAXIBWXSYQG2TZBXT/graph.json","fetch_events":"https://pith.science/api/pith-number/OPGGFYYDJVAXIBWXSYQG2TZBXT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OPGGFYYDJVAXIBWXSYQG2TZBXT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OPGGFYYDJVAXIBWXSYQG2TZBXT/action/storage_attestation","attest_author":"https://pith.science/pith/OPGGFYYDJVAXIBWXSYQG2TZBXT/action/author_attestation","sign_citation":"https://pith.science/pith/OPGGFYYDJVAXIBWXSYQG2TZBXT/action/citation_signature","submit_replication":"https://pith.science/pith/OPGGFYYDJVAXIBWXSYQG2TZBXT/action/replication_record"}},"created_at":"2026-07-05T10:23:23.592479+00:00","updated_at":"2026-07-05T10:23:23.592479+00:00"}