{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:3M75WMYOFCPZWBFN26H7JWVAGS","short_pith_number":"pith:3M75WMYO","schema_version":"1.0","canonical_sha256":"db3fdb330e289f9b04add78ff4daa034929997a973031a6013ed9ca1f92c4340","source":{"kind":"arxiv","id":"2410.14872","version":2},"attestation_state":"computed","paper":{"title":"How to Evaluate Reward Models for RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Anastasios N. Angelopoulos, Banghua Zhu, Connor Chen, Evan Frick, Ion Stoica, Jiantao Jiao, Joseph E. Gonzalez, Tianle Li, Wei-Lin Chiang","submitted_at":"2024-10-18T21:38:21Z","abstract_excerpt":"We introduce a new benchmark for reward models that quantifies their ability to produce strong language models through RLHF (Reinforcement Learning from Human Feedback). The gold-standard approach is to run a full RLHF training pipeline and directly probe downstream LLM performance. However, this process is prohibitively expensive. To address this, we build a predictive model of downstream LLM performance by evaluating the reward model on proxy tasks. These proxy tasks consist of a large-scale human preference and a verifiable correctness preference dataset, in which we measure 12 metrics acro"},"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":"2410.14872","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-10-18T21:38:21Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"b737530287134d059cab2328ba8c6d6f04742cd2c73b92b2f3def4e33c1d9f73","abstract_canon_sha256":"611f09584a763fd3d4aa47df4740f36a2158dba020feb4797fb9c7ef8225bdd9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:24:18.805808Z","signature_b64":"+i7eTlT6LPaShlp0baDwov8qy34tSmxKcxJBllabSra62l4ZoHxW/trqp2xAuEmeYll5f6iMDlO8vDhp3kh8CQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"db3fdb330e289f9b04add78ff4daa034929997a973031a6013ed9ca1f92c4340","last_reissued_at":"2026-07-05T09:24:18.805292Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:24:18.805292Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"How to Evaluate Reward Models for RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Anastasios N. Angelopoulos, Banghua Zhu, Connor Chen, Evan Frick, Ion Stoica, Jiantao Jiao, Joseph E. Gonzalez, Tianle Li, Wei-Lin Chiang","submitted_at":"2024-10-18T21:38:21Z","abstract_excerpt":"We introduce a new benchmark for reward models that quantifies their ability to produce strong language models through RLHF (Reinforcement Learning from Human Feedback). The gold-standard approach is to run a full RLHF training pipeline and directly probe downstream LLM performance. However, this process is prohibitively expensive. To address this, we build a predictive model of downstream LLM performance by evaluating the reward model on proxy tasks. These proxy tasks consist of a large-scale human preference and a verifiable correctness preference dataset, in which we measure 12 metrics acro"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.14872","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/2410.14872/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":"2410.14872","created_at":"2026-07-05T09:24:18.805360+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.14872v2","created_at":"2026-07-05T09:24:18.805360+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.14872","created_at":"2026-07-05T09:24:18.805360+00:00"},{"alias_kind":"pith_short_12","alias_value":"3M75WMYOFCPZ","created_at":"2026-07-05T09:24:18.805360+00:00"},{"alias_kind":"pith_short_16","alias_value":"3M75WMYOFCPZWBFN","created_at":"2026-07-05T09:24:18.805360+00:00"},{"alias_kind":"pith_short_8","alias_value":"3M75WMYO","created_at":"2026-07-05T09:24:18.805360+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2412.15115","citing_title":"Qwen2.5 Technical Report","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2506.01937","citing_title":"RewardBench 2: Advancing Reward Model Evaluation","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2604.12312","citing_title":"CompliBench: Benchmarking LLM Judges for Compliance Violation Detection in Dialogue Systems","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3M75WMYOFCPZWBFN26H7JWVAGS","json":"https://pith.science/pith/3M75WMYOFCPZWBFN26H7JWVAGS.json","graph_json":"https://pith.science/api/pith-number/3M75WMYOFCPZWBFN26H7JWVAGS/graph.json","events_json":"https://pith.science/api/pith-number/3M75WMYOFCPZWBFN26H7JWVAGS/events.json","paper":"https://pith.science/paper/3M75WMYO"},"agent_actions":{"view_html":"https://pith.science/pith/3M75WMYOFCPZWBFN26H7JWVAGS","download_json":"https://pith.science/pith/3M75WMYOFCPZWBFN26H7JWVAGS.json","view_paper":"https://pith.science/paper/3M75WMYO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.14872&json=true","fetch_graph":"https://pith.science/api/pith-number/3M75WMYOFCPZWBFN26H7JWVAGS/graph.json","fetch_events":"https://pith.science/api/pith-number/3M75WMYOFCPZWBFN26H7JWVAGS/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3M75WMYOFCPZWBFN26H7JWVAGS/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3M75WMYOFCPZWBFN26H7JWVAGS/action/storage_attestation","attest_author":"https://pith.science/pith/3M75WMYOFCPZWBFN26H7JWVAGS/action/author_attestation","sign_citation":"https://pith.science/pith/3M75WMYOFCPZWBFN26H7JWVAGS/action/citation_signature","submit_replication":"https://pith.science/pith/3M75WMYOFCPZWBFN26H7JWVAGS/action/replication_record"}},"created_at":"2026-07-05T09:24:18.805360+00:00","updated_at":"2026-07-05T09:24:18.805360+00:00"}