{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:OGLUKIAHTLDA3WW5KB7BF7HS3D","short_pith_number":"pith:OGLUKIAH","schema_version":"1.0","canonical_sha256":"71974520079ac60ddadd507e12fcf2d8f0a0e9d09ff379257ef8c32e25004dc1","source":{"kind":"arxiv","id":"2603.07445","version":2},"attestation_state":"computed","paper":{"title":"Few Tokens, Big Leverage: Preserving Safety Alignment by Constraining Safety Tokens during Fine-tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"An Wang, Guoli Wang, Haonan Shi, Tu Ouyang","submitted_at":"2026-03-08T03:42:55Z","abstract_excerpt":"Large language models (LLMs) often require fine-tuning (FT) to perform well on downstream tasks, but FT can induce safety-alignment drift even when the training dataset contains only benign data. Prior work shows that introducing a small fraction of harmful data can substantially compromise LLM refusal behavior, causing LLMs to comply with harmful requests. Existing defense methods often rely on model-wide interventions, such as restricting which parameters are updated or injecting additional safety data, which can limit generality and degrade downstream task performance. To address these limi"},"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":"2603.07445","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-03-08T03:42:55Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"a10f9af0509997dc293432f4c36a1b1fab03222629d0056c181857d1aac8d9b0","abstract_canon_sha256":"e3c079ae2600bf275b08d69cd35186b06c6e0866a8fc8437aeb3883148e4eabc"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-04T01:09:48.023701Z","signature_b64":"YZRckPpU3mI+3K724pZMx6IiJ8q3HWEkEUO6qvdqTSqm4MwBDR2tX2vrm6XQipX8fpPVuX4+SyfjUhiXOBViAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"71974520079ac60ddadd507e12fcf2d8f0a0e9d09ff379257ef8c32e25004dc1","last_reissued_at":"2026-06-04T01:09:48.022479Z","signature_status":"signed_v1","first_computed_at":"2026-06-04T01:09:48.022479Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Few Tokens, Big Leverage: Preserving Safety Alignment by Constraining Safety Tokens during Fine-tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"An Wang, Guoli Wang, Haonan Shi, Tu Ouyang","submitted_at":"2026-03-08T03:42:55Z","abstract_excerpt":"Large language models (LLMs) often require fine-tuning (FT) to perform well on downstream tasks, but FT can induce safety-alignment drift even when the training dataset contains only benign data. Prior work shows that introducing a small fraction of harmful data can substantially compromise LLM refusal behavior, causing LLMs to comply with harmful requests. Existing defense methods often rely on model-wide interventions, such as restricting which parameters are updated or injecting additional safety data, which can limit generality and degrade downstream task performance. To address these limi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.07445","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/2603.07445/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":"2603.07445","created_at":"2026-06-04T01:09:48.022624+00:00"},{"alias_kind":"arxiv_version","alias_value":"2603.07445v2","created_at":"2026-06-04T01:09:48.022624+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.07445","created_at":"2026-06-04T01:09:48.022624+00:00"},{"alias_kind":"pith_short_12","alias_value":"OGLUKIAHTLDA","created_at":"2026-06-04T01:09:48.022624+00:00"},{"alias_kind":"pith_short_16","alias_value":"OGLUKIAHTLDA3WW5","created_at":"2026-06-04T01:09:48.022624+00:00"},{"alias_kind":"pith_short_8","alias_value":"OGLUKIAH","created_at":"2026-06-04T01:09:48.022624+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":3,"sample":[{"citing_arxiv_id":"2605.17413","citing_title":"Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications","ref_index":39,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02971","citing_title":"Multilingual Safety Alignment via Self-Distillation","ref_index":63,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02971","citing_title":"Multilingual Safety Alignment via Self-Distillation","ref_index":64,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OGLUKIAHTLDA3WW5KB7BF7HS3D","json":"https://pith.science/pith/OGLUKIAHTLDA3WW5KB7BF7HS3D.json","graph_json":"https://pith.science/api/pith-number/OGLUKIAHTLDA3WW5KB7BF7HS3D/graph.json","events_json":"https://pith.science/api/pith-number/OGLUKIAHTLDA3WW5KB7BF7HS3D/events.json","paper":"https://pith.science/paper/OGLUKIAH"},"agent_actions":{"view_html":"https://pith.science/pith/OGLUKIAHTLDA3WW5KB7BF7HS3D","download_json":"https://pith.science/pith/OGLUKIAHTLDA3WW5KB7BF7HS3D.json","view_paper":"https://pith.science/paper/OGLUKIAH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2603.07445&json=true","fetch_graph":"https://pith.science/api/pith-number/OGLUKIAHTLDA3WW5KB7BF7HS3D/graph.json","fetch_events":"https://pith.science/api/pith-number/OGLUKIAHTLDA3WW5KB7BF7HS3D/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OGLUKIAHTLDA3WW5KB7BF7HS3D/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OGLUKIAHTLDA3WW5KB7BF7HS3D/action/storage_attestation","attest_author":"https://pith.science/pith/OGLUKIAHTLDA3WW5KB7BF7HS3D/action/author_attestation","sign_citation":"https://pith.science/pith/OGLUKIAHTLDA3WW5KB7BF7HS3D/action/citation_signature","submit_replication":"https://pith.science/pith/OGLUKIAHTLDA3WW5KB7BF7HS3D/action/replication_record"}},"created_at":"2026-06-04T01:09:48.022624+00:00","updated_at":"2026-06-04T01:09:48.022624+00:00"}