{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:DJRWIYNUUJJ6H4DOUZMUHEBJWB","short_pith_number":"pith:DJRWIYNU","schema_version":"1.0","canonical_sha256":"1a636461b4a253e3f06ea659439029b05f510d6d27c044a766345258c75a8644","source":{"kind":"arxiv","id":"2406.02900","version":2},"attestation_state":"computed","paper":{"title":"Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Bradley Knox, Chelsea Finn, Harshit Sikchi, Joey Hejna, Rafael Rafailov, Ryan Park, Scott Niekum, Yaswanth Chittepu","submitted_at":"2024-06-05T03:41:37Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) has been crucial to the recent success of Large Language Models (LLMs), however, it is often a complex and brittle process. In the classical RLHF framework, a reward model is first trained to represent human preferences, which is in turn used by an online reinforcement learning (RL) algorithm to optimize the LLM. A prominent issue with such methods is reward over-optimization or reward hacking, where performance as measured by the learned proxy reward model increases, but true quality plateaus or even deteriorates. Direct Alignment Algorithms ("},"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":"2406.02900","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-06-05T03:41:37Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"e5d73989fe210f83270294d1a61ea10742618f1b9b9c8c08fdeea49a005f2b29","abstract_canon_sha256":"fb051add89d3b124e94d0c64c25b9bfb4fdc1eeb222ffa2b685d5db35a82e67e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:31:03.720365Z","signature_b64":"fG/aC6ZUIrfggUgMkoWHI4xDOoUAUcuIv8cvpml7zhe8zyPrAS55l0hOEa8b0y79Cbo0WSHAlPJxxLn66TB5AA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1a636461b4a253e3f06ea659439029b05f510d6d27c044a766345258c75a8644","last_reissued_at":"2026-07-05T09:31:03.719850Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:31:03.719850Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Bradley Knox, Chelsea Finn, Harshit Sikchi, Joey Hejna, Rafael Rafailov, Ryan Park, Scott Niekum, Yaswanth Chittepu","submitted_at":"2024-06-05T03:41:37Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) has been crucial to the recent success of Large Language Models (LLMs), however, it is often a complex and brittle process. In the classical RLHF framework, a reward model is first trained to represent human preferences, which is in turn used by an online reinforcement learning (RL) algorithm to optimize the LLM. A prominent issue with such methods is reward over-optimization or reward hacking, where performance as measured by the learned proxy reward model increases, but true quality plateaus or even deteriorates. Direct Alignment Algorithms ("},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.02900","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/2406.02900/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":"2406.02900","created_at":"2026-07-05T09:31:03.719911+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.02900v2","created_at":"2026-07-05T09:31:03.719911+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.02900","created_at":"2026-07-05T09:31:03.719911+00:00"},{"alias_kind":"pith_short_12","alias_value":"DJRWIYNUUJJ6","created_at":"2026-07-05T09:31:03.719911+00:00"},{"alias_kind":"pith_short_16","alias_value":"DJRWIYNUUJJ6H4DO","created_at":"2026-07-05T09:31:03.719911+00:00"},{"alias_kind":"pith_short_8","alias_value":"DJRWIYNU","created_at":"2026-07-05T09:31:03.719911+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":9,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.05904","citing_title":"More Convincing, Not More Correct: Self-Play Reward Hacking of Reference-Free LLM Judges","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2502.00955","citing_title":"Efficient Multi-Agent System Training with Data Influence-Oriented Tree Search","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2502.01237","citing_title":"The Differences Between Direct Alignment Algorithms are a Blur","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2503.11926","citing_title":"Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10810","citing_title":"Likelihood scoring for continuations of mathematical text: a self-supervised benchmark with tests for shortcut vulnerabilities","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2509.20265","citing_title":"Failure Modes of Maximum Entropy RLHF","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12288","citing_title":"TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching","ref_index":142,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12288","citing_title":"TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching","ref_index":142,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06977","citing_title":"$f$-Divergence Regularized RLHF: Two Tales of Sampling and Unified Analyses","ref_index":9,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/DJRWIYNUUJJ6H4DOUZMUHEBJWB","json":"https://pith.science/pith/DJRWIYNUUJJ6H4DOUZMUHEBJWB.json","graph_json":"https://pith.science/api/pith-number/DJRWIYNUUJJ6H4DOUZMUHEBJWB/graph.json","events_json":"https://pith.science/api/pith-number/DJRWIYNUUJJ6H4DOUZMUHEBJWB/events.json","paper":"https://pith.science/paper/DJRWIYNU"},"agent_actions":{"view_html":"https://pith.science/pith/DJRWIYNUUJJ6H4DOUZMUHEBJWB","download_json":"https://pith.science/pith/DJRWIYNUUJJ6H4DOUZMUHEBJWB.json","view_paper":"https://pith.science/paper/DJRWIYNU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.02900&json=true","fetch_graph":"https://pith.science/api/pith-number/DJRWIYNUUJJ6H4DOUZMUHEBJWB/graph.json","fetch_events":"https://pith.science/api/pith-number/DJRWIYNUUJJ6H4DOUZMUHEBJWB/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/DJRWIYNUUJJ6H4DOUZMUHEBJWB/action/timestamp_anchor","attest_storage":"https://pith.science/pith/DJRWIYNUUJJ6H4DOUZMUHEBJWB/action/storage_attestation","attest_author":"https://pith.science/pith/DJRWIYNUUJJ6H4DOUZMUHEBJWB/action/author_attestation","sign_citation":"https://pith.science/pith/DJRWIYNUUJJ6H4DOUZMUHEBJWB/action/citation_signature","submit_replication":"https://pith.science/pith/DJRWIYNUUJJ6H4DOUZMUHEBJWB/action/replication_record"}},"created_at":"2026-07-05T09:31:03.719911+00:00","updated_at":"2026-07-05T09:31:03.719911+00:00"}