{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:23IZB7K2H2STHS5OGU3SX45Y43","short_pith_number":"pith:23IZB7K2","schema_version":"1.0","canonical_sha256":"d6d190fd5a3ea533cbae35372bf3b8e6ddf630437eb272f01b14c5a437f010e0","source":{"kind":"arxiv","id":"2307.15217","version":2},"attestation_state":"computed","paper":{"title":"Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"RLHF, the dominant method for aligning large language models with human goals, carries fundamental limitations that incremental fixes cannot fully resolve.","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.AI","authors_text":"Anand Siththaranjan, Anca Dragan, Andi Peng, Charbel-Rapha\\\"el Segerie, Claudia Shi, David Krueger, David Lindner, Dmitrii Krasheninnikov, Dorsa Sadigh, Dylan Hadfield-Menell, Erdem B{\\i}y{\\i}k, Eric J. Michaud, Jacob Pfau, Javier Rando, J\\'er\\'emy Scheurer, Lauro Langosco, Max Nadeau, Mehul Damani, Micah Carroll, Pedro Freire, Peter Hase, Phillip Christoffersen, Rachel Freedman, Samuel Marks, Stephen Casper, Stewart Slocum, Thomas Krendl Gilbert, Tomasz Korbak, Tony Wang, Usman Anwar, Xander Davies, Xin Chen","submitted_at":"2023-07-27T22:29:25Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) is a technique for training AI systems to align with human goals. RLHF has emerged as the central method used to finetune state-of-the-art large language models (LLMs). Despite this popularity, there has been relatively little public work systematizing its flaws. In this paper, we (1) survey open problems and fundamental limitations of RLHF and related methods; (2) overview techniques to understand, improve, and complement RLHF in practice; and (3) propose auditing and disclosure standards to improve societal oversight of RLHF systems. Our work"},"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":true},"canonical_record":{"source":{"id":"2307.15217","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2023-07-27T22:29:25Z","cross_cats_sorted":["cs.CL","cs.LG"],"title_canon_sha256":"f64b8e348d873883aaed9189c75a52a859f0648b3f54a9627b55a100e94a8e16","abstract_canon_sha256":"b2d5a1ffd2f126b3687464bdd54878590cc00c30071a32d643bc8ba98db2121f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:21.419160Z","signature_b64":"IgWYAzu8E11YJKSnEBOZRO42bFde2+JBCzbLVCW8GfzIcKwkkklGwB+qS0rz/wj0+t86Er3LpC7yetHSWlyQCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d6d190fd5a3ea533cbae35372bf3b8e6ddf630437eb272f01b14c5a437f010e0","last_reissued_at":"2026-05-17T23:39:21.418541Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:21.418541Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"RLHF, the dominant method for aligning large language models with human goals, carries fundamental limitations that incremental fixes cannot fully resolve.","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.AI","authors_text":"Anand Siththaranjan, Anca Dragan, Andi Peng, Charbel-Rapha\\\"el Segerie, Claudia Shi, David Krueger, David Lindner, Dmitrii Krasheninnikov, Dorsa Sadigh, Dylan Hadfield-Menell, Erdem B{\\i}y{\\i}k, Eric J. Michaud, Jacob Pfau, Javier Rando, J\\'er\\'emy Scheurer, Lauro Langosco, Max Nadeau, Mehul Damani, Micah Carroll, Pedro Freire, Peter Hase, Phillip Christoffersen, Rachel Freedman, Samuel Marks, Stephen Casper, Stewart Slocum, Thomas Krendl Gilbert, Tomasz Korbak, Tony Wang, Usman Anwar, Xander Davies, Xin Chen","submitted_at":"2023-07-27T22:29:25Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) is a technique for training AI systems to align with human goals. RLHF has emerged as the central method used to finetune state-of-the-art large language models (LLMs). Despite this popularity, there has been relatively little public work systematizing its flaws. In this paper, we (1) survey open problems and fundamental limitations of RLHF and related methods; (2) overview techniques to understand, improve, and complement RLHF in practice; and (3) propose auditing and disclosure standards to improve societal oversight of RLHF systems. Our work"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"RLHF has emerged as the central method used to finetune state-of-the-art large language models but has fundamental limitations, emphasizing the importance of a multi-faceted approach to the development of safer AI systems.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the identified open problems represent fundamental limitations of RLHF rather than challenges that can be resolved through incremental improvements or better implementation.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"RLHF has significant open problems and fundamental limitations that require a multi-faceted approach for safer AI development.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"RLHF, the dominant method for aligning large language models with human goals, carries fundamental limitations that incremental fixes cannot fully resolve.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"006f1672e3aad5d3409d2fe0f53106294ab14aa905c8916cb46d1259040d51e5"},"source":{"id":"2307.15217","kind":"arxiv","version":2},"verdict":{"id":"1928dcfc-c59b-4bc6-a627-8307e124f306","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T22:40:12.619170Z","strongest_claim":"RLHF has emerged as the central method used to finetune state-of-the-art large language models but has fundamental limitations, emphasizing the importance of a multi-faceted approach to the development of safer AI systems.","one_line_summary":"RLHF has significant open problems and fundamental limitations that require a multi-faceted approach for safer AI development.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the identified open problems represent fundamental limitations of RLHF rather than challenges that can be resolved through incremental improvements or better implementation.","pith_extraction_headline":"RLHF, the dominant method for aligning large language models with human goals, carries fundamental limitations that incremental fixes cannot fully resolve."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"5803d0de03d454fe4fec6249f84c86bea3f67e278651388713b656953744c67d"},"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":"2307.15217","created_at":"2026-05-17T23:39:21.418626+00:00"},{"alias_kind":"arxiv_version","alias_value":"2307.15217v2","created_at":"2026-05-17T23:39:21.418626+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2307.15217","created_at":"2026-05-17T23:39:21.418626+00:00"},{"alias_kind":"pith_short_12","alias_value":"23IZB7K2H2ST","created_at":"2026-05-18T12:33:33.725879+00:00"},{"alias_kind":"pith_short_16","alias_value":"23IZB7K2H2STHS5O","created_at":"2026-05-18T12:33:33.725879+00:00"},{"alias_kind":"pith_short_8","alias_value":"23IZB7K2","created_at":"2026-05-18T12:33:33.725879+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":45,"internal_anchor_count":45,"sample":[{"citing_arxiv_id":"2310.15288","citing_title":"Active teacher selection for reward learning","ref_index":2,"is_internal_anchor":true},{"citing_arxiv_id":"2408.12935","citing_title":"AI Safety Landscape for Large Language Models: Taxonomy, State-of-the-art, and Future Directions","ref_index":102,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21984","citing_title":"Echo: Learning from Experience Data via User-Driven Refinement","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2508.04149","citing_title":"Difficulty-Based Preference Data Selection by DPO Implicit Reward Gap","ref_index":30,"is_internal_anchor":true},{"citing_arxiv_id":"2601.21484","citing_title":"ETS: Energy-Guided Test-Time Scaling for Training-Free RL Alignment","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2605.20654","citing_title":"REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2604.28010","citing_title":"Learning from Disagreement: Clinician Overrides as Implicit Preference Signals for Clinical AI in Value-Based Care","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16339","citing_title":"Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16198","citing_title":"Formal Methods Meet LLMs: Auditing, Monitoring, and Intervention for Compliance of Advanced AI Systems","ref_index":26,"is_internal_anchor":true},{"citing_arxiv_id":"2605.17458","citing_title":"ClaHF: A Human Feedback-inspired Reinforcement Learning Framework for Improving Classification Tasks","ref_index":52,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19516","citing_title":"Base Models Look Human To AI Detectors","ref_index":18,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16872","citing_title":"Some[Body] Must Receive That Pain for Agent Accountability","ref_index":55,"is_internal_anchor":true},{"citing_arxiv_id":"2605.15207","citing_title":"TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination","ref_index":63,"is_internal_anchor":true},{"citing_arxiv_id":"2506.12382","citing_title":"Exploring the Secondary Risks of Large Language Models","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2506.21834","citing_title":"PrefPaint: Enhancing Medical Image Inpainting through Expert Human Feedback","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2507.04005","citing_title":"Exploring a Gamified Personality Assessment Method through Interaction with LLM Agents Embodying Different Personalities","ref_index":22,"is_internal_anchor":true},{"citing_arxiv_id":"2304.06767","citing_title":"RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment","ref_index":126,"is_internal_anchor":true},{"citing_arxiv_id":"2410.18451","citing_title":"Skywork-Reward: Bag of Tricks for Reward Modeling in LLMs","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2406.10162","citing_title":"Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models","ref_index":264,"is_internal_anchor":true},{"citing_arxiv_id":"2409.12917","citing_title":"Training Language Models to Self-Correct via Reinforcement Learning","ref_index":141,"is_internal_anchor":true},{"citing_arxiv_id":"2512.21110","citing_title":"Beyond Context: Large Language Models' Failure to Grasp Users' Intent","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2402.05070","citing_title":"A Roadmap to Pluralistic Alignment","ref_index":154,"is_internal_anchor":true},{"citing_arxiv_id":"2601.21484","citing_title":"ETS: Energy-Guided Test-Time Scaling for Training-Free RL Alignment","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2602.19837","citing_title":"Meta-Learning and Meta-Reinforcement Learning -- Tracing the Path towards DeepMind's Adaptive Agent","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13875","citing_title":"Common-agency Games for Multi-Objective Test-Time Alignment","ref_index":184,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/23IZB7K2H2STHS5OGU3SX45Y43","json":"https://pith.science/pith/23IZB7K2H2STHS5OGU3SX45Y43.json","graph_json":"https://pith.science/api/pith-number/23IZB7K2H2STHS5OGU3SX45Y43/graph.json","events_json":"https://pith.science/api/pith-number/23IZB7K2H2STHS5OGU3SX45Y43/events.json","paper":"https://pith.science/paper/23IZB7K2"},"agent_actions":{"view_html":"https://pith.science/pith/23IZB7K2H2STHS5OGU3SX45Y43","download_json":"https://pith.science/pith/23IZB7K2H2STHS5OGU3SX45Y43.json","view_paper":"https://pith.science/paper/23IZB7K2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2307.15217&json=true","fetch_graph":"https://pith.science/api/pith-number/23IZB7K2H2STHS5OGU3SX45Y43/graph.json","fetch_events":"https://pith.science/api/pith-number/23IZB7K2H2STHS5OGU3SX45Y43/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/23IZB7K2H2STHS5OGU3SX45Y43/action/timestamp_anchor","attest_storage":"https://pith.science/pith/23IZB7K2H2STHS5OGU3SX45Y43/action/storage_attestation","attest_author":"https://pith.science/pith/23IZB7K2H2STHS5OGU3SX45Y43/action/author_attestation","sign_citation":"https://pith.science/pith/23IZB7K2H2STHS5OGU3SX45Y43/action/citation_signature","submit_replication":"https://pith.science/pith/23IZB7K2H2STHS5OGU3SX45Y43/action/replication_record"}},"created_at":"2026-05-17T23:39:21.418626+00:00","updated_at":"2026-05-17T23:39:21.418626+00:00"}