{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:BHEEIWWARMYM576EHCA63VHSK6","short_pith_number":"pith:BHEEIWWA","schema_version":"1.0","canonical_sha256":"09c8445ac08b30ceffc43881edd4f257b658acca5c3f45d2f9ae4004fb062ea5","source":{"kind":"arxiv","id":"2306.14111","version":2},"attestation_state":"computed","paper":{"title":"Is RLHF More Difficult than Standard RL?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Chi Jin, Qinghua Liu, Yuanhao Wang","submitted_at":"2023-06-25T03:18:15Z","abstract_excerpt":"Reinforcement learning from Human Feedback (RLHF) learns from preference signals, while standard Reinforcement Learning (RL) directly learns from reward signals. Preferences arguably contain less information than rewards, which makes preference-based RL seemingly more difficult. This paper theoretically proves that, for a wide range of preference models, we can solve preference-based RL directly using existing algorithms and techniques for reward-based RL, with small or no extra costs. Specifically, (1) for preferences that are drawn from reward-based probabilistic models, we reduce the proble"},"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":"2306.14111","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2023-06-25T03:18:15Z","cross_cats_sorted":["cs.AI","stat.ML"],"title_canon_sha256":"28195d3291c25f21e48733d9edaf4a797d445b9fb05f979b59fe792a95fe8813","abstract_canon_sha256":"9f77f3bed07806e97f39a6837375d2af903d0c855ac6fc13d342179de84504b4"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:08:59.356616Z","signature_b64":"B0tk8GTpKsWWpGj3CNt9u58o6/t5wsuj6+neqerq3BlK7PEiMG/ib5oEI2+8GqZJqFtAzt/Dfieh5mIG4DriCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"09c8445ac08b30ceffc43881edd4f257b658acca5c3f45d2f9ae4004fb062ea5","last_reissued_at":"2026-07-05T07:08:59.356127Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:08:59.356127Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Is RLHF More Difficult than Standard RL?","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Chi Jin, Qinghua Liu, Yuanhao Wang","submitted_at":"2023-06-25T03:18:15Z","abstract_excerpt":"Reinforcement learning from Human Feedback (RLHF) learns from preference signals, while standard Reinforcement Learning (RL) directly learns from reward signals. Preferences arguably contain less information than rewards, which makes preference-based RL seemingly more difficult. This paper theoretically proves that, for a wide range of preference models, we can solve preference-based RL directly using existing algorithms and techniques for reward-based RL, with small or no extra costs. Specifically, (1) for preferences that are drawn from reward-based probabilistic models, we reduce the proble"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2306.14111","kind":"arxiv","version":2},"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/2306.14111/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":"2306.14111","created_at":"2026-07-05T07:08:59.356184+00:00"},{"alias_kind":"arxiv_version","alias_value":"2306.14111v2","created_at":"2026-07-05T07:08:59.356184+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2306.14111","created_at":"2026-07-05T07:08:59.356184+00:00"},{"alias_kind":"pith_short_12","alias_value":"BHEEIWWARMYM","created_at":"2026-07-05T07:08:59.356184+00:00"},{"alias_kind":"pith_short_16","alias_value":"BHEEIWWARMYM576E","created_at":"2026-07-05T07:08:59.356184+00:00"},{"alias_kind":"pith_short_8","alias_value":"BHEEIWWA","created_at":"2026-07-05T07:08:59.356184+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.27082","citing_title":"Finding Stationary Points by Comparisons","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06053","citing_title":"Online KL-Regularized Reinforcement Learning with Function Approximation under Misspecification","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.23244","citing_title":"Convex Optimization for Alignment and Preference Learning on a Single GPU","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2510.08539","citing_title":"On the optimization dynamics of RLVR: Gradient gap and step size thresholds","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02349","citing_title":"OPRIDE: Offline Preference-based Reinforcement Learning via In-Dataset Exploration","ref_index":47,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BHEEIWWARMYM576EHCA63VHSK6","json":"https://pith.science/pith/BHEEIWWARMYM576EHCA63VHSK6.json","graph_json":"https://pith.science/api/pith-number/BHEEIWWARMYM576EHCA63VHSK6/graph.json","events_json":"https://pith.science/api/pith-number/BHEEIWWARMYM576EHCA63VHSK6/events.json","paper":"https://pith.science/paper/BHEEIWWA"},"agent_actions":{"view_html":"https://pith.science/pith/BHEEIWWARMYM576EHCA63VHSK6","download_json":"https://pith.science/pith/BHEEIWWARMYM576EHCA63VHSK6.json","view_paper":"https://pith.science/paper/BHEEIWWA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2306.14111&json=true","fetch_graph":"https://pith.science/api/pith-number/BHEEIWWARMYM576EHCA63VHSK6/graph.json","fetch_events":"https://pith.science/api/pith-number/BHEEIWWARMYM576EHCA63VHSK6/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BHEEIWWARMYM576EHCA63VHSK6/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BHEEIWWARMYM576EHCA63VHSK6/action/storage_attestation","attest_author":"https://pith.science/pith/BHEEIWWARMYM576EHCA63VHSK6/action/author_attestation","sign_citation":"https://pith.science/pith/BHEEIWWARMYM576EHCA63VHSK6/action/citation_signature","submit_replication":"https://pith.science/pith/BHEEIWWARMYM576EHCA63VHSK6/action/replication_record"}},"created_at":"2026-07-05T07:08:59.356184+00:00","updated_at":"2026-07-05T07:08:59.356184+00:00"}