{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:4THQND7G6V7WMGNWUZDKT4K7O7","short_pith_number":"pith:4THQND7G","schema_version":"1.0","canonical_sha256":"e4cf068fe6f57f6619b6a646a9f15f77eb2ba8f2cfec5ee3ff75a3e3df223703","source":{"kind":"arxiv","id":"2309.16155","version":1},"attestation_state":"computed","paper":{"title":"The Trickle-down Impact of Reward (In-)consistency on RLHF","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Baolin Peng, Daniel Khashabi, Dong Yu, Haitao Mi, Lifeng Jin, Linfeng Song, Lingfeng Shen, Sihao Chen","submitted_at":"2023-09-28T04:05:13Z","abstract_excerpt":"Standard practice within Reinforcement Learning from Human Feedback (RLHF) involves optimizing against a Reward Model (RM), which itself is trained to reflect human preferences for desirable generations. A notable subject that is understudied is the (in-)consistency of RMs -- whether they can recognize the semantic changes to different prompts and appropriately adapt their reward assignments -- and their impact on the downstream RLHF model.\n  In this paper, we visit a series of research questions relevant to RM inconsistency: (1) How can we measure the consistency of reward models? (2) How con"},"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":"2309.16155","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2023-09-28T04:05:13Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"7b56e817348d85d3f7fdb35bb2bfcae0bda1a16abec94596a4da8c0bc24df7f5","abstract_canon_sha256":"c3752e9faa924b6fc19bfd4fa4a6e0a68da646947fd86f6d060e8a7e1c86df16"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T06:55:12.142844Z","signature_b64":"qBcZdo2PBeYaF3udMfRVrDGzw2wu2ohCDqRZvoFX9vQfFCEpxF0LjTaWxB73LHLej/PztZ8TbTijg54c+0W0BA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e4cf068fe6f57f6619b6a646a9f15f77eb2ba8f2cfec5ee3ff75a3e3df223703","last_reissued_at":"2026-07-05T06:55:12.142394Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T06:55:12.142394Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Trickle-down Impact of Reward (In-)consistency on RLHF","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Baolin Peng, Daniel Khashabi, Dong Yu, Haitao Mi, Lifeng Jin, Linfeng Song, Lingfeng Shen, Sihao Chen","submitted_at":"2023-09-28T04:05:13Z","abstract_excerpt":"Standard practice within Reinforcement Learning from Human Feedback (RLHF) involves optimizing against a Reward Model (RM), which itself is trained to reflect human preferences for desirable generations. A notable subject that is understudied is the (in-)consistency of RMs -- whether they can recognize the semantic changes to different prompts and appropriately adapt their reward assignments -- and their impact on the downstream RLHF model.\n  In this paper, we visit a series of research questions relevant to RM inconsistency: (1) How can we measure the consistency of reward models? (2) How con"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2309.16155","kind":"arxiv","version":1},"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/2309.16155/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":"2309.16155","created_at":"2026-07-05T06:55:12.142449+00:00"},{"alias_kind":"arxiv_version","alias_value":"2309.16155v1","created_at":"2026-07-05T06:55:12.142449+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2309.16155","created_at":"2026-07-05T06:55:12.142449+00:00"},{"alias_kind":"pith_short_12","alias_value":"4THQND7G6V7W","created_at":"2026-07-05T06:55:12.142449+00:00"},{"alias_kind":"pith_short_16","alias_value":"4THQND7G6V7WMGNW","created_at":"2026-07-05T06:55:12.142449+00:00"},{"alias_kind":"pith_short_8","alias_value":"4THQND7G","created_at":"2026-07-05T06:55:12.142449+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.16339","citing_title":"Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders","ref_index":30,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4THQND7G6V7WMGNWUZDKT4K7O7","json":"https://pith.science/pith/4THQND7G6V7WMGNWUZDKT4K7O7.json","graph_json":"https://pith.science/api/pith-number/4THQND7G6V7WMGNWUZDKT4K7O7/graph.json","events_json":"https://pith.science/api/pith-number/4THQND7G6V7WMGNWUZDKT4K7O7/events.json","paper":"https://pith.science/paper/4THQND7G"},"agent_actions":{"view_html":"https://pith.science/pith/4THQND7G6V7WMGNWUZDKT4K7O7","download_json":"https://pith.science/pith/4THQND7G6V7WMGNWUZDKT4K7O7.json","view_paper":"https://pith.science/paper/4THQND7G","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2309.16155&json=true","fetch_graph":"https://pith.science/api/pith-number/4THQND7G6V7WMGNWUZDKT4K7O7/graph.json","fetch_events":"https://pith.science/api/pith-number/4THQND7G6V7WMGNWUZDKT4K7O7/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4THQND7G6V7WMGNWUZDKT4K7O7/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4THQND7G6V7WMGNWUZDKT4K7O7/action/storage_attestation","attest_author":"https://pith.science/pith/4THQND7G6V7WMGNWUZDKT4K7O7/action/author_attestation","sign_citation":"https://pith.science/pith/4THQND7G6V7WMGNWUZDKT4K7O7/action/citation_signature","submit_replication":"https://pith.science/pith/4THQND7G6V7WMGNWUZDKT4K7O7/action/replication_record"}},"created_at":"2026-07-05T06:55:12.142449+00:00","updated_at":"2026-07-05T06:55:12.142449+00:00"}