{"paper":{"title":"RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reinforcement learning from AI feedback matches human feedback performance for aligning large language models.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Abhinav Rastogi, Colton Bishop, Ethan Hall, Harrison Lee, Hassan Mansoor, Johan Ferret, Kellie Lu, Samrat Phatale, Sushant Prakash, Thomas Mesnard, Victor Carbune","submitted_at":"2023-09-01T05:53:33Z","abstract_excerpt":"Reinforcement learning from human feedback (RLHF) has proven effective in aligning large language models (LLMs) with human preferences, but gathering high-quality preference labels is expensive. RL from AI Feedback (RLAIF), introduced in Bai et al., offers a promising alternative that trains the reward model (RM) on preferences generated by an off-the-shelf LLM. Across the tasks of summarization, helpful dialogue generation, and harmless dialogue generation, we show that RLAIF achieves comparable performance to RLHF. Furthermore, we take a step towards \"self-improvement\" by demonstrating that "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across the tasks of summarization, helpful dialogue generation, and harmless dialogue generation, we show that RLAIF achieves comparable performance to RLHF. ... we introduce direct-RLAIF (d-RLAIF) ... which achieves superior performance to canonical RLAIF.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the preferences generated by an off-the-shelf LLM are high-quality enough to serve as a substitute for human preferences in training the reward model.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"RLAIF matches RLHF on summarization and dialogue tasks, with a direct-RLAIF variant achieving superior results by using LLM rewards directly during training.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning from AI feedback matches human feedback performance for aligning large language models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"4ddbd0027246b56fe7179d811e1ce883159e5134259fd24be22f93395f94c2b3"},"source":{"id":"2309.00267","kind":"arxiv","version":3},"verdict":{"id":"a3f1a224-083d-4a40-8281-981bb014d035","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T21:29:18.481786Z","strongest_claim":"Across the tasks of summarization, helpful dialogue generation, and harmless dialogue generation, we show that RLAIF achieves comparable performance to RLHF. ... we introduce direct-RLAIF (d-RLAIF) ... which achieves superior performance to canonical RLAIF.","one_line_summary":"RLAIF matches RLHF on summarization and dialogue tasks, with a direct-RLAIF variant achieving superior results by using LLM rewards directly during training.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the preferences generated by an off-the-shelf LLM are high-quality enough to serve as a substitute for human preferences in training the reward model.","pith_extraction_headline":"Reinforcement learning from AI feedback matches human feedback performance for aligning large language models."},"references":{"count":98,"sample":[{"doi":"","year":2022,"title":"E., Fort, S., Lanham, T., Telleen-Lawton, T., Conerly, T., Henighan, T., Hume, T., Bowman, S","work_id":"686a699b-d7b5-4f10-ba6c-e3df30418d80","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1901,"title":"D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al","work_id":"fb006848-e99b-481f-9340-c35dbff67d47","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"F., Leike, J., Brown, T., Martic, M., Legg, S., and Amodei, D","work_id":"463eef82-ca77-4ce1-8588-f7c3f7abe5a4","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"RAFT : Reward ranked finetuning for generative foundation model alignment","work_id":"aa80a8cf-c1b9-4e19-bb61-6c9aea78c183","ref_index":8,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Understanding dataset difficulty with V -usable information","work_id":"8b66d63b-71d0-445b-8db1-ceaecd3c2e1f","ref_index":9,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":98,"snapshot_sha256":"280efcd0accaebc769d975a017e5a7572be47b99c701fdd06e17e462155ec766","internal_anchors":15},"formal_canon":{"evidence_count":2,"snapshot_sha256":"4ea79e99becc6383a8dc457d14fef873224bf8b76f834a6ee459b0fefd95c6cb"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}