{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:IHFZODKQQJYE2Z27EINNJ4KXUM","short_pith_number":"pith:IHFZODKQ","schema_version":"1.0","canonical_sha256":"41cb970d5082704d675f221ad4f157a302dc40d48b6e687c6c7bf2816b10746f","source":{"kind":"arxiv","id":"2405.19544","version":3},"attestation_state":"computed","paper":{"title":"One-Shot Safety Alignment for Large Language Models via Optimal Dualization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.LG","math.OC","stat.ML"],"primary_cat":"cs.AI","authors_text":"Dongsheng Ding, Edgar Dobriban, Hamed Hassani, Osbert Bastani, Shuo Li, Xinmeng Huang","submitted_at":"2024-05-29T22:12:52Z","abstract_excerpt":"The growing safety concerns surrounding large language models raise an urgent need to align them with diverse human preferences to simultaneously enhance their helpfulness and safety. A promising approach is to enforce safety constraints through Reinforcement Learning from Human Feedback (RLHF). For such constrained RLHF, typical Lagrangian-based primal-dual policy optimization methods are computationally expensive and often unstable. This paper presents a perspective of dualization that reduces constrained alignment to an equivalent unconstrained alignment problem. We do so by pre-optimizing "},"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":"2405.19544","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2024-05-29T22:12:52Z","cross_cats_sorted":["cs.CL","cs.LG","math.OC","stat.ML"],"title_canon_sha256":"d46dd572ec3f5618ebea3d7f6c29db0491f1792d4005755e0fe5010ab4eb2c28","abstract_canon_sha256":"dbd995990b49a402f0ffa513993710bbff46f570d349db22e3c14ca02f001b3d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:39:06.771618Z","signature_b64":"jc+1E7Itk9X2cN9lKyQsRr6f9KuUxVZ6AMGn7OyZK46umgg7vnMqSRKb0u1ZNpaEMYl+9QVhC0kPkWXaU2LkAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"41cb970d5082704d675f221ad4f157a302dc40d48b6e687c6c7bf2816b10746f","last_reissued_at":"2026-07-05T09:39:06.771131Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:39:06.771131Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"One-Shot Safety Alignment for Large Language Models via Optimal Dualization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL","cs.LG","math.OC","stat.ML"],"primary_cat":"cs.AI","authors_text":"Dongsheng Ding, Edgar Dobriban, Hamed Hassani, Osbert Bastani, Shuo Li, Xinmeng Huang","submitted_at":"2024-05-29T22:12:52Z","abstract_excerpt":"The growing safety concerns surrounding large language models raise an urgent need to align them with diverse human preferences to simultaneously enhance their helpfulness and safety. A promising approach is to enforce safety constraints through Reinforcement Learning from Human Feedback (RLHF). For such constrained RLHF, typical Lagrangian-based primal-dual policy optimization methods are computationally expensive and often unstable. This paper presents a perspective of dualization that reduces constrained alignment to an equivalent unconstrained alignment problem. We do so by pre-optimizing "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2405.19544","kind":"arxiv","version":3},"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/2405.19544/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":"2405.19544","created_at":"2026-07-05T09:39:06.771189+00:00"},{"alias_kind":"arxiv_version","alias_value":"2405.19544v3","created_at":"2026-07-05T09:39:06.771189+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2405.19544","created_at":"2026-07-05T09:39:06.771189+00:00"},{"alias_kind":"pith_short_12","alias_value":"IHFZODKQQJYE","created_at":"2026-07-05T09:39:06.771189+00:00"},{"alias_kind":"pith_short_16","alias_value":"IHFZODKQQJYE2Z27","created_at":"2026-07-05T09:39:06.771189+00:00"},{"alias_kind":"pith_short_8","alias_value":"IHFZODKQ","created_at":"2026-07-05T09:39:06.771189+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2505.23729","citing_title":"Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time","ref_index":13,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/IHFZODKQQJYE2Z27EINNJ4KXUM","json":"https://pith.science/pith/IHFZODKQQJYE2Z27EINNJ4KXUM.json","graph_json":"https://pith.science/api/pith-number/IHFZODKQQJYE2Z27EINNJ4KXUM/graph.json","events_json":"https://pith.science/api/pith-number/IHFZODKQQJYE2Z27EINNJ4KXUM/events.json","paper":"https://pith.science/paper/IHFZODKQ"},"agent_actions":{"view_html":"https://pith.science/pith/IHFZODKQQJYE2Z27EINNJ4KXUM","download_json":"https://pith.science/pith/IHFZODKQQJYE2Z27EINNJ4KXUM.json","view_paper":"https://pith.science/paper/IHFZODKQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2405.19544&json=true","fetch_graph":"https://pith.science/api/pith-number/IHFZODKQQJYE2Z27EINNJ4KXUM/graph.json","fetch_events":"https://pith.science/api/pith-number/IHFZODKQQJYE2Z27EINNJ4KXUM/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/IHFZODKQQJYE2Z27EINNJ4KXUM/action/timestamp_anchor","attest_storage":"https://pith.science/pith/IHFZODKQQJYE2Z27EINNJ4KXUM/action/storage_attestation","attest_author":"https://pith.science/pith/IHFZODKQQJYE2Z27EINNJ4KXUM/action/author_attestation","sign_citation":"https://pith.science/pith/IHFZODKQQJYE2Z27EINNJ4KXUM/action/citation_signature","submit_replication":"https://pith.science/pith/IHFZODKQQJYE2Z27EINNJ4KXUM/action/replication_record"}},"created_at":"2026-07-05T09:39:06.771189+00:00","updated_at":"2026-07-05T09:39:06.771189+00:00"}