{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:4RLN2RG5WEMP4F5E727M3HJJ3J","short_pith_number":"pith:4RLN2RG5","schema_version":"1.0","canonical_sha256":"e456dd44ddb118fe17a4febecd9d29da6bb7c51eab1b629216a2388116d43365","source":{"kind":"arxiv","id":"2407.09121","version":2},"attestation_state":"computed","paper":{"title":"Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Jen-tse Huang, Jiahao Xu, Pinjia He, Tian Liang, Wenxiang Jiao, Wenxuan Wang, Youliang Yuan, Zhaopeng Tu","submitted_at":"2024-07-12T09:36:33Z","abstract_excerpt":"This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) by identifying and tackling a refusal position bias within safety tuning data, which compromises the models' ability to appropriately refuse generating unsafe content. We introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position, significantly enhancing their safety capabilities. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation (MLE) with Harmful Response Prefix, which trains mod"},"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":"2407.09121","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-07-12T09:36:33Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"b7a38144a197c386e15e5df4decbf4de6322cc3154daeac4c10ac92d3cc99532","abstract_canon_sha256":"f8191bfd0cbba752820303714ce47e1b5a5b29acd4f0a05a20392e03dfbd47ed"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:07:59.288068Z","signature_b64":"RgljBMvfQ/Wgud97Gq6Q/YoO0wqHNcqBIvNKTBFCtv6seotoS1Km6M26jWvjFJNsB3kss7H8OZ+rgiHbcToICg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e456dd44ddb118fe17a4febecd9d29da6bb7c51eab1b629216a2388116d43365","last_reissued_at":"2026-07-05T11:07:59.287516Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:07:59.287516Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Jen-tse Huang, Jiahao Xu, Pinjia He, Tian Liang, Wenxiang Jiao, Wenxuan Wang, Youliang Yuan, Zhaopeng Tu","submitted_at":"2024-07-12T09:36:33Z","abstract_excerpt":"This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) by identifying and tackling a refusal position bias within safety tuning data, which compromises the models' ability to appropriately refuse generating unsafe content. We introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position, significantly enhancing their safety capabilities. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation (MLE) with Harmful Response Prefix, which trains mod"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2407.09121","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/2407.09121/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":"2407.09121","created_at":"2026-07-05T11:07:59.287577+00:00"},{"alias_kind":"arxiv_version","alias_value":"2407.09121v2","created_at":"2026-07-05T11:07:59.287577+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2407.09121","created_at":"2026-07-05T11:07:59.287577+00:00"},{"alias_kind":"pith_short_12","alias_value":"4RLN2RG5WEMP","created_at":"2026-07-05T11:07:59.287577+00:00"},{"alias_kind":"pith_short_16","alias_value":"4RLN2RG5WEMP4F5E","created_at":"2026-07-05T11:07:59.287577+00:00"},{"alias_kind":"pith_short_8","alias_value":"4RLN2RG5","created_at":"2026-07-05T11:07:59.287577+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2505.14226","citing_title":"Phonetic Perturbations Reveal Tokenizer-Rooted Safety Gaps in LLMs","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2604.04992","citing_title":"FreakOut-LLM: The Effect of Emotional Stimuli on Safety Alignment","ref_index":2,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4RLN2RG5WEMP4F5E727M3HJJ3J","json":"https://pith.science/pith/4RLN2RG5WEMP4F5E727M3HJJ3J.json","graph_json":"https://pith.science/api/pith-number/4RLN2RG5WEMP4F5E727M3HJJ3J/graph.json","events_json":"https://pith.science/api/pith-number/4RLN2RG5WEMP4F5E727M3HJJ3J/events.json","paper":"https://pith.science/paper/4RLN2RG5"},"agent_actions":{"view_html":"https://pith.science/pith/4RLN2RG5WEMP4F5E727M3HJJ3J","download_json":"https://pith.science/pith/4RLN2RG5WEMP4F5E727M3HJJ3J.json","view_paper":"https://pith.science/paper/4RLN2RG5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2407.09121&json=true","fetch_graph":"https://pith.science/api/pith-number/4RLN2RG5WEMP4F5E727M3HJJ3J/graph.json","fetch_events":"https://pith.science/api/pith-number/4RLN2RG5WEMP4F5E727M3HJJ3J/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4RLN2RG5WEMP4F5E727M3HJJ3J/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4RLN2RG5WEMP4F5E727M3HJJ3J/action/storage_attestation","attest_author":"https://pith.science/pith/4RLN2RG5WEMP4F5E727M3HJJ3J/action/author_attestation","sign_citation":"https://pith.science/pith/4RLN2RG5WEMP4F5E727M3HJJ3J/action/citation_signature","submit_replication":"https://pith.science/pith/4RLN2RG5WEMP4F5E727M3HJJ3J/action/replication_record"}},"created_at":"2026-07-05T11:07:59.287577+00:00","updated_at":"2026-07-05T11:07:59.287577+00:00"}