{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:UWZSZDCNE5ZLW4MBWDPH4PRQOG","short_pith_number":"pith:UWZSZDCN","schema_version":"1.0","canonical_sha256":"a5b32c8c4d2772bb7181b0de7e3e3071932ee580b045a821d8afba7b1e7f3da2","source":{"kind":"arxiv","id":"2504.20157","version":2},"attestation_state":"computed","paper":{"title":"Toward Evaluative Thinking: Meta Policy Optimization with Evolving Reward Models","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Chanwoo Park, Dongyeop Kang, Suin Kim, Vipul Raheja, Zae Myung Kim","submitted_at":"2025-04-28T18:02:35Z","abstract_excerpt":"Reward-based alignment methods for large language models (LLMs) face two key limitations: vulnerability to reward hacking, where models exploit flaws in the reward signal; and reliance on brittle, labor-intensive prompt engineering when LLMs are used as reward models. We introduce Meta Policy Optimization (MPO), a framework that addresses these challenges by integrating a meta-reward model that dynamically refines the reward model's prompt throughout training. In MPO, the meta-reward model monitors the evolving training context and continuously adjusts the reward model's prompt to maintain hig"},"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":"2504.20157","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2025-04-28T18:02:35Z","cross_cats_sorted":[],"title_canon_sha256":"fba430344ee564341a8eeaaf413c784196d06b469a3b61aef11df8c72ee48635","abstract_canon_sha256":"cf520f99905c65841559b4016df16e4c3a7d3c2e15254ab072c7d2322d90f2f9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:04:21.896232Z","signature_b64":"M+SotEYWNLx8QdFzUk12FsYbkO/z1fhHywgHH93O0tHgq+bIfuNDiyBxQg9UwLcVAgFDb20gOAufmp/eSz5iCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a5b32c8c4d2772bb7181b0de7e3e3071932ee580b045a821d8afba7b1e7f3da2","last_reissued_at":"2026-07-05T11:04:21.895711Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:04:21.895711Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Toward Evaluative Thinking: Meta Policy Optimization with Evolving Reward Models","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Chanwoo Park, Dongyeop Kang, Suin Kim, Vipul Raheja, Zae Myung Kim","submitted_at":"2025-04-28T18:02:35Z","abstract_excerpt":"Reward-based alignment methods for large language models (LLMs) face two key limitations: vulnerability to reward hacking, where models exploit flaws in the reward signal; and reliance on brittle, labor-intensive prompt engineering when LLMs are used as reward models. We introduce Meta Policy Optimization (MPO), a framework that addresses these challenges by integrating a meta-reward model that dynamically refines the reward model's prompt throughout training. In MPO, the meta-reward model monitors the evolving training context and continuously adjusts the reward model's prompt to maintain hig"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.20157","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/2504.20157/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":"2504.20157","created_at":"2026-07-05T11:04:21.895780+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.20157v2","created_at":"2026-07-05T11:04:21.895780+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.20157","created_at":"2026-07-05T11:04:21.895780+00:00"},{"alias_kind":"pith_short_12","alias_value":"UWZSZDCNE5ZL","created_at":"2026-07-05T11:04:21.895780+00:00"},{"alias_kind":"pith_short_16","alias_value":"UWZSZDCNE5ZLW4MB","created_at":"2026-07-05T11:04:21.895780+00:00"},{"alias_kind":"pith_short_8","alias_value":"UWZSZDCN","created_at":"2026-07-05T11:04:21.895780+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2507.17746","citing_title":"Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains","ref_index":18,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UWZSZDCNE5ZLW4MBWDPH4PRQOG","json":"https://pith.science/pith/UWZSZDCNE5ZLW4MBWDPH4PRQOG.json","graph_json":"https://pith.science/api/pith-number/UWZSZDCNE5ZLW4MBWDPH4PRQOG/graph.json","events_json":"https://pith.science/api/pith-number/UWZSZDCNE5ZLW4MBWDPH4PRQOG/events.json","paper":"https://pith.science/paper/UWZSZDCN"},"agent_actions":{"view_html":"https://pith.science/pith/UWZSZDCNE5ZLW4MBWDPH4PRQOG","download_json":"https://pith.science/pith/UWZSZDCNE5ZLW4MBWDPH4PRQOG.json","view_paper":"https://pith.science/paper/UWZSZDCN","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.20157&json=true","fetch_graph":"https://pith.science/api/pith-number/UWZSZDCNE5ZLW4MBWDPH4PRQOG/graph.json","fetch_events":"https://pith.science/api/pith-number/UWZSZDCNE5ZLW4MBWDPH4PRQOG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UWZSZDCNE5ZLW4MBWDPH4PRQOG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UWZSZDCNE5ZLW4MBWDPH4PRQOG/action/storage_attestation","attest_author":"https://pith.science/pith/UWZSZDCNE5ZLW4MBWDPH4PRQOG/action/author_attestation","sign_citation":"https://pith.science/pith/UWZSZDCNE5ZLW4MBWDPH4PRQOG/action/citation_signature","submit_replication":"https://pith.science/pith/UWZSZDCNE5ZLW4MBWDPH4PRQOG/action/replication_record"}},"created_at":"2026-07-05T11:04:21.895780+00:00","updated_at":"2026-07-05T11:04:21.895780+00:00"}