{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:L72V4I34P6POZ6CZ4QZ6UFMOGW","short_pith_number":"pith:L72V4I34","schema_version":"1.0","canonical_sha256":"5ff55e237c7f9eecf859e433ea158e358bab18905bc2da16d838cbea48f2d927","source":{"kind":"arxiv","id":"2404.08495","version":3},"attestation_state":"computed","paper":{"title":"Dataset Reset Policy Optimization for RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Dipendra Misra, Jason D. Lee, Jonathan D. Chang, Kiant\\'e Brantley, Owen Oertell, Wenhao Zhan, Wen Sun","submitted_at":"2024-04-12T14:25:49Z","abstract_excerpt":"Reinforcement Learning (RL) from Human Preference-based feedback is a popular paradigm for fine-tuning generative models, which has produced impressive models such as GPT-4 and Claude3 Opus. This framework often consists of two steps: learning a reward model from an offline preference dataset followed by running online RL to optimize the learned reward model. In this work, leveraging the idea of reset, we propose a new RLHF algorithm with provable guarantees. Motivated by the fact that offline preference dataset provides informative states (i.e., data that is preferred by the labelers), our ne"},"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":"2404.08495","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-12T14:25:49Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"79d9dbeb2f642b0a817b8df4cb7ab1b2ec56e54d840f15c195d5b190e4dd8a48","abstract_canon_sha256":"2d8e0fe7c152ea10edda9d5e82c87427d50b60f0f76fd3b9e22fffc15ead9f52"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:08:41.512409Z","signature_b64":"W5Jp4buahnLZ+baDgMrxSO+ECvnuV26ydgNJURHCvG78Vp2BUpkkKRm8eKGPDgM5r0U8v/ecrDk9myygliNIAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5ff55e237c7f9eecf859e433ea158e358bab18905bc2da16d838cbea48f2d927","last_reissued_at":"2026-07-05T08:08:41.512017Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:08:41.512017Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Dataset Reset Policy Optimization for RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Dipendra Misra, Jason D. Lee, Jonathan D. Chang, Kiant\\'e Brantley, Owen Oertell, Wenhao Zhan, Wen Sun","submitted_at":"2024-04-12T14:25:49Z","abstract_excerpt":"Reinforcement Learning (RL) from Human Preference-based feedback is a popular paradigm for fine-tuning generative models, which has produced impressive models such as GPT-4 and Claude3 Opus. This framework often consists of two steps: learning a reward model from an offline preference dataset followed by running online RL to optimize the learned reward model. In this work, leveraging the idea of reset, we propose a new RLHF algorithm with provable guarantees. Motivated by the fact that offline preference dataset provides informative states (i.e., data that is preferred by the labelers), our ne"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2404.08495","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/2404.08495/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":"2404.08495","created_at":"2026-07-05T08:08:41.512070+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.08495v3","created_at":"2026-07-05T08:08:41.512070+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.08495","created_at":"2026-07-05T08:08:41.512070+00:00"},{"alias_kind":"pith_short_12","alias_value":"L72V4I34P6PO","created_at":"2026-07-05T08:08:41.512070+00:00"},{"alias_kind":"pith_short_16","alias_value":"L72V4I34P6POZ6CZ","created_at":"2026-07-05T08:08:41.512070+00:00"},{"alias_kind":"pith_short_8","alias_value":"L72V4I34","created_at":"2026-07-05T08:08:41.512070+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09711","citing_title":"Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization","ref_index":201,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05152","citing_title":"Reinforcement Learning from Rich Feedback with Distributional DAgger","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2605.25507","citing_title":"Credit Assignment with Resets in Language Model Reasoning","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09214","citing_title":"Fast Rates for Offline Contextual Bandits with Forward-KL Regularization under Single-Policy Concentrability","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06987","citing_title":"Response Time Enhances Alignment with Heterogeneous Preferences","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2604.13602","citing_title":"Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges","ref_index":98,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/L72V4I34P6POZ6CZ4QZ6UFMOGW","json":"https://pith.science/pith/L72V4I34P6POZ6CZ4QZ6UFMOGW.json","graph_json":"https://pith.science/api/pith-number/L72V4I34P6POZ6CZ4QZ6UFMOGW/graph.json","events_json":"https://pith.science/api/pith-number/L72V4I34P6POZ6CZ4QZ6UFMOGW/events.json","paper":"https://pith.science/paper/L72V4I34"},"agent_actions":{"view_html":"https://pith.science/pith/L72V4I34P6POZ6CZ4QZ6UFMOGW","download_json":"https://pith.science/pith/L72V4I34P6POZ6CZ4QZ6UFMOGW.json","view_paper":"https://pith.science/paper/L72V4I34","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.08495&json=true","fetch_graph":"https://pith.science/api/pith-number/L72V4I34P6POZ6CZ4QZ6UFMOGW/graph.json","fetch_events":"https://pith.science/api/pith-number/L72V4I34P6POZ6CZ4QZ6UFMOGW/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/L72V4I34P6POZ6CZ4QZ6UFMOGW/action/timestamp_anchor","attest_storage":"https://pith.science/pith/L72V4I34P6POZ6CZ4QZ6UFMOGW/action/storage_attestation","attest_author":"https://pith.science/pith/L72V4I34P6POZ6CZ4QZ6UFMOGW/action/author_attestation","sign_citation":"https://pith.science/pith/L72V4I34P6POZ6CZ4QZ6UFMOGW/action/citation_signature","submit_replication":"https://pith.science/pith/L72V4I34P6POZ6CZ4QZ6UFMOGW/action/replication_record"}},"created_at":"2026-07-05T08:08:41.512070+00:00","updated_at":"2026-07-05T08:08:41.512070+00:00"}