{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:5MN5WHBGO7D6IB7VEYLZAYCP4E","short_pith_number":"pith:5MN5WHBG","schema_version":"1.0","canonical_sha256":"eb1bdb1c2677c7e407f5261790604fe11f6ac923d0625d17d3dc6b3be19a0d30","source":{"kind":"arxiv","id":"2506.08473","version":4},"attestation_state":"computed","paper":{"title":"AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Hailiang Dai, Jiayu Yao, Jigang Wang, Kunpeng Ning, Li Yuan, Qihui Zhang, Shuo Yang, Xiaojun Jia, Yibing Song, Yuyang Liu","submitted_at":"2025-06-10T05:59:48Z","abstract_excerpt":"Fine-tuning large language models (LLMs) improves performance but introduces critical safety vulnerabilities: even minimal harmful data can severely compromise safety measures. We observe that perturbations orthogonal to the alignment direction - defined by weight differences between aligned (safe) and unaligned models - rapidly compromise model safety. In contrast, updates along the alignment direction largely preserve it, revealing the parameter space as a \"narrow safety basin\". To address this, we propose AsFT (Anchoring Safety in Fine-Tuning) to maintain safety by explicitly constraining u"},"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":"2506.08473","kind":"arxiv","version":4},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-06-10T05:59:48Z","cross_cats_sorted":[],"title_canon_sha256":"3282fc0329ac8b465edd3ca9a6bb6c6c605208e464aa3f817440b571dd7a8921","abstract_canon_sha256":"752850bdb13c1d7a40831be7690391883c45824ad6916b6a1cc25b2add5edc3c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-11T01:10:26.770234Z","signature_b64":"uI4T+nRyl6whF6jcYHpuqgHt3M4jTos5L4LuUDvDnpbRP6Ysw/4UO/Q5dz6PrFFqNvrQ01XJhmBXNQYEPlqrCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"eb1bdb1c2677c7e407f5261790604fe11f6ac923d0625d17d3dc6b3be19a0d30","last_reissued_at":"2026-06-11T01:10:26.769367Z","signature_status":"signed_v1","first_computed_at":"2026-06-11T01:10:26.769367Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Hailiang Dai, Jiayu Yao, Jigang Wang, Kunpeng Ning, Li Yuan, Qihui Zhang, Shuo Yang, Xiaojun Jia, Yibing Song, Yuyang Liu","submitted_at":"2025-06-10T05:59:48Z","abstract_excerpt":"Fine-tuning large language models (LLMs) improves performance but introduces critical safety vulnerabilities: even minimal harmful data can severely compromise safety measures. We observe that perturbations orthogonal to the alignment direction - defined by weight differences between aligned (safe) and unaligned models - rapidly compromise model safety. In contrast, updates along the alignment direction largely preserve it, revealing the parameter space as a \"narrow safety basin\". To address this, we propose AsFT (Anchoring Safety in Fine-Tuning) to maintain safety by explicitly constraining u"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.08473","kind":"arxiv","version":4},"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/2506.08473/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":"2506.08473","created_at":"2026-06-11T01:10:26.769475+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.08473v4","created_at":"2026-06-11T01:10:26.769475+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.08473","created_at":"2026-06-11T01:10:26.769475+00:00"},{"alias_kind":"pith_short_12","alias_value":"5MN5WHBGO7D6","created_at":"2026-06-11T01:10:26.769475+00:00"},{"alias_kind":"pith_short_16","alias_value":"5MN5WHBGO7D6IB7V","created_at":"2026-06-11T01:10:26.769475+00:00"},{"alias_kind":"pith_short_8","alias_value":"5MN5WHBG","created_at":"2026-06-11T01:10:26.769475+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2510.16888","citing_title":"Uniworld-V2: Reinforce Image Editing with Diffusion Negative-aware Finetuning and MLLM Implicit Feedback","ref_index":23,"is_internal_anchor":true},{"citing_arxiv_id":"2604.12384","citing_title":"Preventing Safety Drift in Large Language Models via Coupled Weight and Activation Constraints","ref_index":50,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/5MN5WHBGO7D6IB7VEYLZAYCP4E","json":"https://pith.science/pith/5MN5WHBGO7D6IB7VEYLZAYCP4E.json","graph_json":"https://pith.science/api/pith-number/5MN5WHBGO7D6IB7VEYLZAYCP4E/graph.json","events_json":"https://pith.science/api/pith-number/5MN5WHBGO7D6IB7VEYLZAYCP4E/events.json","paper":"https://pith.science/paper/5MN5WHBG"},"agent_actions":{"view_html":"https://pith.science/pith/5MN5WHBGO7D6IB7VEYLZAYCP4E","download_json":"https://pith.science/pith/5MN5WHBGO7D6IB7VEYLZAYCP4E.json","view_paper":"https://pith.science/paper/5MN5WHBG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.08473&json=true","fetch_graph":"https://pith.science/api/pith-number/5MN5WHBGO7D6IB7VEYLZAYCP4E/graph.json","fetch_events":"https://pith.science/api/pith-number/5MN5WHBGO7D6IB7VEYLZAYCP4E/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/5MN5WHBGO7D6IB7VEYLZAYCP4E/action/timestamp_anchor","attest_storage":"https://pith.science/pith/5MN5WHBGO7D6IB7VEYLZAYCP4E/action/storage_attestation","attest_author":"https://pith.science/pith/5MN5WHBGO7D6IB7VEYLZAYCP4E/action/author_attestation","sign_citation":"https://pith.science/pith/5MN5WHBGO7D6IB7VEYLZAYCP4E/action/citation_signature","submit_replication":"https://pith.science/pith/5MN5WHBGO7D6IB7VEYLZAYCP4E/action/replication_record"}},"created_at":"2026-06-11T01:10:26.769475+00:00","updated_at":"2026-06-11T01:10:26.769475+00:00"}