{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:H6X43XHDWAIJSECPR5PJ4LJMFJ","short_pith_number":"pith:H6X43XHD","schema_version":"1.0","canonical_sha256":"3fafcddce3b01099104f8f5e9e2d2c2a5c6bd7fdd6202b3fda5e2990f4642373","source":{"kind":"arxiv","id":"2310.04373","version":2},"attestation_state":"computed","paper":{"title":"Confronting Reward Model Overoptimization with Constrained RLHF","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Aaditya K. Singh, Anca D. Dragan, DJ Strouse, Ruslan Salakhutdinov, Stephen McAleer, Ted Moskovitz, Tuomas Sandholm","submitted_at":"2023-10-06T16:59:17Z","abstract_excerpt":"Large language models are typically aligned with human preferences by optimizing $\\textit{reward models}$ (RMs) fitted to human feedback. However, human preferences are multi-faceted, and it is increasingly common to derive reward from a composition of simpler reward models which each capture a different aspect of language quality. This itself presents a challenge, as it is difficult to appropriately weight these component RMs when combining them. Compounding this difficulty, because any RM is only a proxy for human evaluation, this process is vulnerable to $\\textit{overoptimization}$, wherein"},"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":"2310.04373","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2023-10-06T16:59:17Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"a2efabab29e20192724b2d3668dab8b7c859d1be77b79f786bef31770ef250fe","abstract_canon_sha256":"46631051abb021ca270527031b82525a0b959a9550ed0472937f866b725e7fc7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T06:59:00.578663Z","signature_b64":"+dRKoUlJtn7MZmue2ZbkIXpQyM2l+twMQ+IPoPnLqxcvknbbl5/YFugqU23591C/VzsKOL7gv3J0cNDJd6DLDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3fafcddce3b01099104f8f5e9e2d2c2a5c6bd7fdd6202b3fda5e2990f4642373","last_reissued_at":"2026-07-05T06:59:00.578183Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T06:59:00.578183Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Confronting Reward Model Overoptimization with Constrained RLHF","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Aaditya K. Singh, Anca D. Dragan, DJ Strouse, Ruslan Salakhutdinov, Stephen McAleer, Ted Moskovitz, Tuomas Sandholm","submitted_at":"2023-10-06T16:59:17Z","abstract_excerpt":"Large language models are typically aligned with human preferences by optimizing $\\textit{reward models}$ (RMs) fitted to human feedback. However, human preferences are multi-faceted, and it is increasingly common to derive reward from a composition of simpler reward models which each capture a different aspect of language quality. This itself presents a challenge, as it is difficult to appropriately weight these component RMs when combining them. Compounding this difficulty, because any RM is only a proxy for human evaluation, this process is vulnerable to $\\textit{overoptimization}$, wherein"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2310.04373","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/2310.04373/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":"2310.04373","created_at":"2026-07-05T06:59:00.578244+00:00"},{"alias_kind":"arxiv_version","alias_value":"2310.04373v2","created_at":"2026-07-05T06:59:00.578244+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2310.04373","created_at":"2026-07-05T06:59:00.578244+00:00"},{"alias_kind":"pith_short_12","alias_value":"H6X43XHDWAIJ","created_at":"2026-07-05T06:59:00.578244+00:00"},{"alias_kind":"pith_short_16","alias_value":"H6X43XHDWAIJSECP","created_at":"2026-07-05T06:59:00.578244+00:00"},{"alias_kind":"pith_short_8","alias_value":"H6X43XHD","created_at":"2026-07-05T06:59:00.578244+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.23597","citing_title":"Against Proxy Optimization","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20994","citing_title":"Towards Context-Invariant Safety Alignment for Large Language Models","ref_index":90,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11679","citing_title":"Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion","ref_index":71,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02766","citing_title":"Random Is Hard to Beat: Active Selection in online DPO with Modern LLMs","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11679","citing_title":"Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion","ref_index":71,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14265","citing_title":"Reinforcement Learning via Value Gradient Flow","ref_index":45,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/H6X43XHDWAIJSECPR5PJ4LJMFJ","json":"https://pith.science/pith/H6X43XHDWAIJSECPR5PJ4LJMFJ.json","graph_json":"https://pith.science/api/pith-number/H6X43XHDWAIJSECPR5PJ4LJMFJ/graph.json","events_json":"https://pith.science/api/pith-number/H6X43XHDWAIJSECPR5PJ4LJMFJ/events.json","paper":"https://pith.science/paper/H6X43XHD"},"agent_actions":{"view_html":"https://pith.science/pith/H6X43XHDWAIJSECPR5PJ4LJMFJ","download_json":"https://pith.science/pith/H6X43XHDWAIJSECPR5PJ4LJMFJ.json","view_paper":"https://pith.science/paper/H6X43XHD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2310.04373&json=true","fetch_graph":"https://pith.science/api/pith-number/H6X43XHDWAIJSECPR5PJ4LJMFJ/graph.json","fetch_events":"https://pith.science/api/pith-number/H6X43XHDWAIJSECPR5PJ4LJMFJ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/H6X43XHDWAIJSECPR5PJ4LJMFJ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/H6X43XHDWAIJSECPR5PJ4LJMFJ/action/storage_attestation","attest_author":"https://pith.science/pith/H6X43XHDWAIJSECPR5PJ4LJMFJ/action/author_attestation","sign_citation":"https://pith.science/pith/H6X43XHDWAIJSECPR5PJ4LJMFJ/action/citation_signature","submit_replication":"https://pith.science/pith/H6X43XHDWAIJSECPR5PJ4LJMFJ/action/replication_record"}},"created_at":"2026-07-05T06:59:00.578244+00:00","updated_at":"2026-07-05T06:59:00.578244+00:00"}