{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "guan-et-al-deliberative-alignment-2024", "title": "Guan et al.: policy-reasoning deliberative alignment", "summary": "Trained reasoning models to internalize explicit safety specifications through specification-referencing synthetic supervision followed by reinforcement learning, reporting better jailbreak resistance and less overrefusal on the evaluated tasks.", "boundary": "No universal safety, faithful reasoning or jailbreak-immunity guarantee. Human-written policies and prompt-category information remain inputs even though human-written chains of thought and answers are not required.", "area": "Reasoning supervision and alignment experiments", "topic": "Artificial intelligence", "year": 2024, "lane": "HUMAN", "result": "EMPIRICAL", "package_id": "guan-et-al-deliberative-alignment-2024", "target": { "record_id": "urn:uuid:9cadd33a-cc2d-4912-a941-f825dd61b1bc", "version": 1, "content_hash": "sha256:919d13688dd5187e1bc207a88dd63dcc7503221114b5bba6b7b9f48784226e4b" }, "humans": [ "Melody Y. Guan", "Manas Joglekar", "Eric Wallace", "Saachi Jain", "Boaz Barak", "Alec Helyar", "Rachel Dias", "Andrea Vallone", "Hongyu Ren", "Jason Wei", "Hyung Won Chung", "Sam Toyer", "Johannes Heidecke", "Alex Beutel", "Amelia Glaese" ], "authors_complete": true, "human_role_note": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261015-arxiv-2412-16339v2", "title": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" }, { "id": "cur20261015-support-guan-et-al-deliberative-alignment-2024", "title": "Guan et al.: policy-reasoning deliberative alignment — company research account", "url": "https://openai.com/index/deliberative-alignment/" } ], "points": "125000", "published_grade": { "id": "guan-et-al-deliberative-alignment-2024", "entry_id": "guan-et-al-deliberative-alignment-2024", "band": "A", "modifier": null, "target": { "record_id": "urn:uuid:9cadd33a-cc2d-4912-a941-f825dd61b1bc", "version": 1, "content_hash": "sha256:919d13688dd5187e1bc207a88dd63dcc7503221114b5bba6b7b9f48784226e4b" }, "what_changed": "Trained reasoning models to internalize explicit safety specifications through specification-referencing synthetic supervision followed by reinforcement learning, reporting better jailbreak resistance and less overrefusal on the evaluated tasks.", "scope": "Trained reasoning models to internalize explicit safety specifications through specification-referencing synthetic supervision followed by reinforcement learning, reporting better jailbreak resistance and less overrefusal on the evaluated tasks. No universal safety, faithful reasoning or jailbreak-immunity guarantee. Human-written policies and prompt-category information remain inputs even though human-written chains of thought and answers are not required.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:9cadd33a-cc2d-4912-a941-f825dd61b1bc", "version": 1, "content_hash": "sha256:919d13688dd5187e1bc207a88dd63dcc7503221114b5bba6b7b9f48784226e4b" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-16T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Melody Y. Guan", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Manas Joglekar", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Eric Wallace", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Saachi Jain", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Boaz Barak", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Alec Helyar", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Rachel Dias", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Andrea Vallone", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Hongyu Ren", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jason Wei", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Hyung Won Chung", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Sam Toyer", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Johannes Heidecke", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Alex Beutel", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Amelia Glaese", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "discoverer" ], "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" } }, { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-1", "kind": "source_reference", "source": { "citation": "Guan et al.: policy-reasoning deliberative alignment — company research account", "url": "https://openai.com/index/deliberative-alignment/" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "No universal safety, faithful reasoning or jailbreak-immunity guarantee. Human-written policies and prompt-category information remain inputs even though human-written chains of thought and answers are not required.", "domain": "Reasoning supervision and alignment experiments" }, "sources": [ { "citation": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2" }, { "citation": "Guan et al.: policy-reasoning deliberative alignment — company research account", "url": "https://openai.com/index/deliberative-alignment/" } ], "statement": "# Guan et al.: policy-reasoning deliberative alignment\n\nTrained reasoning models to internalize explicit safety specifications through specification-referencing synthetic supervision followed by reinforcement learning, reporting better jailbreak resistance and less overrefusal on the evaluated tasks.\n\n## Scientific scope\n\nNo universal safety, faithful reasoning or jailbreak-immunity guarantee. Human-written policies and prompt-category information remain inputs even though human-written chains of thought and answers are not required.\n\n## Reading status\n\nThis is a source-linked WRS catalogue note. A full explanatory guide and independent scientific reconstruction have not been completed for this entry. Read the identified original work for its complete narrative and arguments.\n\n## People and systems\n\nMelody Y. Guan; Manas Joglekar; Eric Wallace; Saachi Jain; Boaz Barak; Alec Helyar; Rachel Dias; Andrea Vallone; Hongyu Ren; Jason Wei; Hyung Won Chung; Sam Toyer; Johannes Heidecke; Alex Beutel; Amelia Glaese\n\nSource-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.\n\n\n\n## Original sources\n\n- [Deliberative Alignment: Reasoning Enables Safer Language Models](https://arxiv.org/abs/2412.16339v2)\n- [Guan et al.: policy-reasoning deliberative alignment — company research account](https://openai.com/index/deliberative-alignment/)\n", "summary": "Trained reasoning models to internalize explicit safety specifications through specification-referencing synthetic supervision followed by reinforcement learning, reporting better jailbreak resistance and less overrefusal on the evaluated tasks.", "title": "Guan et al.: policy-reasoning deliberative alignment" }, "record_id": "urn:uuid:9cadd33a-cc2d-4912-a941-f825dd61b1bc", "version": 1, "source_ref": { "record_id": "urn:uuid:9cadd33a-cc2d-4912-a941-f825dd61b1bc", "version": 1, "content_hash": "sha256:919d13688dd5187e1bc207a88dd63dcc7503221114b5bba6b7b9f48784226e4b" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:0ce5570f06dacb37a54eda7ec3d56fe966497ade15af05df22c127b0b252e0da" }, "target": { "record_id": "urn:uuid:9cadd33a-cc2d-4912-a941-f825dd61b1bc", "version": 1, "content_hash": "sha256:919d13688dd5187e1bc207a88dd63dcc7503221114b5bba6b7b9f48784226e4b" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "guan-et-al-deliberative-alignment-2024", "entry_id": "guan-et-al-deliberative-alignment-2024", "revision": 1, "title": "Guan et al.: policy-reasoning deliberative alignment", "context": "Reasoning supervision and alignment experiments", "track": "Human-designed research method or resource", "band": "A", "modifier": null, "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-16", "evidence_cutoff": "2026-10-16", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "Trained reasoning models to internalize explicit safety specifications through specification-referencing synthetic supervision followed by reinforcement learning, reporting better jailbreak resistance and less overrefusal on the evaluated tasks.", "scope": "Trained reasoning models to internalize explicit safety specifications through specification-referencing synthetic supervision followed by reinforcement learning, reporting better jailbreak resistance and less overrefusal on the evaluated tasks. No universal safety, faithful reasoning or jailbreak-immunity guarantee. Human-written policies and prompt-category information remain inputs even though human-written chains of thought and answers are not required.", "evidence_summary": "The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. No universal safety, faithful reasoning or jailbreak-immunity guarantee. Human-written policies and prompt-category information remain inputs even though human-written chains of thought and answers are not required.", "references": [ { "title": "Deliberative Alignment: Reasoning Enables Safer Language Models", "url": "https://arxiv.org/abs/2412.16339v2", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." }, { "title": "Guan et al.: policy-reasoning deliberative alignment — company research account", "url": "https://openai.com/index/deliberative-alignment/", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." } ], "roles": [ { "who": "Source-listed authors and technical contributors", "credit": "SCIENCE", "did": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities." } ], "target": { "record_id": "urn:uuid:9cadd33a-cc2d-4912-a941-f825dd61b1bc", "version": 1, "content_hash": "sha256:919d13688dd5187e1bc207a88dd63dcc7503221114b5bba6b7b9f48784226e4b" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "125000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": null, "reading_status": "SOURCE_LINKED_NOTE" }