{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:NTETKTWBOVTKGAHZWFZVSNI7RJ","short_pith_number":"pith:NTETKTWB","schema_version":"1.0","canonical_sha256":"6cc9354ec17566a300f9b17359351f8a628aa14d0d09da31060b1c1289498930","source":{"kind":"arxiv","id":"2403.02475","version":1},"attestation_state":"computed","paper":{"title":"Enhancing LLM Safety via Constrained Direct Preference Optimization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Xiaolin Sun, Zixuan Liu, Zizhan Zheng","submitted_at":"2024-03-04T20:39:24Z","abstract_excerpt":"The rapidly increasing capabilities of large language models (LLMs) raise an urgent need to align AI systems with diverse human preferences to simultaneously enhance their usefulness and safety, despite the often conflicting nature of these goals. To address this important problem, a promising approach is to enforce a safety constraint at the fine-tuning stage through a constrained Reinforcement Learning from Human Feedback (RLHF) framework. This approach, however, is computationally expensive and often unstable. In this work, we introduce Constrained DPO (C-DPO), a novel extension of the rece"},"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":"2403.02475","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-03-04T20:39:24Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"fde3c7c1310928b1e190cacf5a326adbd4d5d68dd78ce1f520c4818bd2e023bc","abstract_canon_sha256":"a550e16cd4ad25dabacb58709a7b9019e26acb1b81bb3b37857a99f7139a9b15"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:52:18.023542Z","signature_b64":"WbtIakVB+5xBmVGnGDewxyiGV4zr1ajJEP9Yr/F7MHi9WWESn7jt1xhdbI+eOVI9oou+yHhEmGDHr9Xm6NhgDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6cc9354ec17566a300f9b17359351f8a628aa14d0d09da31060b1c1289498930","last_reissued_at":"2026-07-05T07:52:18.023064Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:52:18.023064Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Enhancing LLM Safety via Constrained Direct Preference Optimization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Xiaolin Sun, Zixuan Liu, Zizhan Zheng","submitted_at":"2024-03-04T20:39:24Z","abstract_excerpt":"The rapidly increasing capabilities of large language models (LLMs) raise an urgent need to align AI systems with diverse human preferences to simultaneously enhance their usefulness and safety, despite the often conflicting nature of these goals. To address this important problem, a promising approach is to enforce a safety constraint at the fine-tuning stage through a constrained Reinforcement Learning from Human Feedback (RLHF) framework. This approach, however, is computationally expensive and often unstable. In this work, we introduce Constrained DPO (C-DPO), a novel extension of the rece"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2403.02475","kind":"arxiv","version":1},"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/2403.02475/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":"2403.02475","created_at":"2026-07-05T07:52:18.023124+00:00"},{"alias_kind":"arxiv_version","alias_value":"2403.02475v1","created_at":"2026-07-05T07:52:18.023124+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2403.02475","created_at":"2026-07-05T07:52:18.023124+00:00"},{"alias_kind":"pith_short_12","alias_value":"NTETKTWBOVTK","created_at":"2026-07-05T07:52:18.023124+00:00"},{"alias_kind":"pith_short_16","alias_value":"NTETKTWBOVTKGAHZ","created_at":"2026-07-05T07:52:18.023124+00:00"},{"alias_kind":"pith_short_8","alias_value":"NTETKTWB","created_at":"2026-07-05T07:52:18.023124+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2607.00481","citing_title":"Beyond the Prompt: Jailbreaking Function-Calling LLMs via Simulated Moderation Traces","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2509.03526","citing_title":"Enhancing Speech Large Language Models through Reinforced Behavior Alignment","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2407.04295","citing_title":"Jailbreak Attacks and Defenses Against Large Language Models: A Survey","ref_index":59,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24488","citing_title":"Scalable First-Order Interior Point Trust Region Algorithms for Linearly Constrained Optimization","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2604.20685","citing_title":"MGDA-Decoupled: Geometry-Aware Multi-Objective Optimisation for DPO-based LLM Alignment","ref_index":54,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NTETKTWBOVTKGAHZWFZVSNI7RJ","json":"https://pith.science/pith/NTETKTWBOVTKGAHZWFZVSNI7RJ.json","graph_json":"https://pith.science/api/pith-number/NTETKTWBOVTKGAHZWFZVSNI7RJ/graph.json","events_json":"https://pith.science/api/pith-number/NTETKTWBOVTKGAHZWFZVSNI7RJ/events.json","paper":"https://pith.science/paper/NTETKTWB"},"agent_actions":{"view_html":"https://pith.science/pith/NTETKTWBOVTKGAHZWFZVSNI7RJ","download_json":"https://pith.science/pith/NTETKTWBOVTKGAHZWFZVSNI7RJ.json","view_paper":"https://pith.science/paper/NTETKTWB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2403.02475&json=true","fetch_graph":"https://pith.science/api/pith-number/NTETKTWBOVTKGAHZWFZVSNI7RJ/graph.json","fetch_events":"https://pith.science/api/pith-number/NTETKTWBOVTKGAHZWFZVSNI7RJ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NTETKTWBOVTKGAHZWFZVSNI7RJ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NTETKTWBOVTKGAHZWFZVSNI7RJ/action/storage_attestation","attest_author":"https://pith.science/pith/NTETKTWBOVTKGAHZWFZVSNI7RJ/action/author_attestation","sign_citation":"https://pith.science/pith/NTETKTWBOVTKGAHZWFZVSNI7RJ/action/citation_signature","submit_replication":"https://pith.science/pith/NTETKTWBOVTKGAHZWFZVSNI7RJ/action/replication_record"}},"created_at":"2026-07-05T07:52:18.023124+00:00","updated_at":"2026-07-05T07:52:18.023124+00:00"}