{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:5FWM6ENK2CNBXOSQUVM46TRFTD","short_pith_number":"pith:5FWM6ENK","schema_version":"1.0","canonical_sha256":"e96ccf11aad09a1bba50a559cf4e2598e6902502aee109858379819890b025a3","source":{"kind":"arxiv","id":"2402.13210","version":2},"attestation_state":"computed","paper":{"title":"Bayesian Reward Models for LLM Alignment","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Adam X. Yang, Haitham Bou-Ammar, Jun Wang, Laurence Aitchison, Maxime Robeyns, Thomas Coste, Zhengyan Shi","submitted_at":"2024-02-20T18:20:59Z","abstract_excerpt":"To ensure that large language model (LLM) responses are helpful and non-toxic, a reward model trained on human preference data is usually used. LLM responses with high rewards are then selected through best-of-$n$ (BoN) sampling or the LLM is further optimized to produce responses with high rewards through reinforcement learning from human feedback (RLHF). However, these processes are susceptible to reward overoptimization or `hacking', where responses receive high rewards due to imperfections in the reward model rather than true preference, particularly as prompts or responses deviate from th"},"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":"2402.13210","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-02-20T18:20:59Z","cross_cats_sorted":[],"title_canon_sha256":"5ff4f7778d156ed353f1692dc075df8f93cda31b80614d20bd1d2ab4f42adcd6","abstract_canon_sha256":"2c2e44973de72317414e68684976a097cf151b7a7198018ff27c73f72b26895a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:39:31.280457Z","signature_b64":"PA0qrfUDgsm9t6FQ40OKCMAP+001axEtqy4TR0zU08aCU6nnJC1XZILImbszagF3gjALzgpzP8xx96dcM3+rCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e96ccf11aad09a1bba50a559cf4e2598e6902502aee109858379819890b025a3","last_reissued_at":"2026-07-05T08:39:31.279957Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:39:31.279957Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Bayesian Reward Models for LLM Alignment","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Adam X. Yang, Haitham Bou-Ammar, Jun Wang, Laurence Aitchison, Maxime Robeyns, Thomas Coste, Zhengyan Shi","submitted_at":"2024-02-20T18:20:59Z","abstract_excerpt":"To ensure that large language model (LLM) responses are helpful and non-toxic, a reward model trained on human preference data is usually used. LLM responses with high rewards are then selected through best-of-$n$ (BoN) sampling or the LLM is further optimized to produce responses with high rewards through reinforcement learning from human feedback (RLHF). However, these processes are susceptible to reward overoptimization or `hacking', where responses receive high rewards due to imperfections in the reward model rather than true preference, particularly as prompts or responses deviate from th"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.13210","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/2402.13210/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":"2402.13210","created_at":"2026-07-05T08:39:31.280018+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.13210v2","created_at":"2026-07-05T08:39:31.280018+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.13210","created_at":"2026-07-05T08:39:31.280018+00:00"},{"alias_kind":"pith_short_12","alias_value":"5FWM6ENK2CNB","created_at":"2026-07-05T08:39:31.280018+00:00"},{"alias_kind":"pith_short_16","alias_value":"5FWM6ENK2CNBXOSQ","created_at":"2026-07-05T08:39:31.280018+00:00"},{"alias_kind":"pith_short_8","alias_value":"5FWM6ENK","created_at":"2026-07-05T08:39:31.280018+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09078","citing_title":"The Hidden Bias of Process Reward Models:PRISM for Rewarding the Right Reasoning","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2606.29184","citing_title":"BaRA: Bayesian Adaptive Rank Allocation for Parameter-Efficient Fine-Tuning","ref_index":12,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/5FWM6ENK2CNBXOSQUVM46TRFTD","json":"https://pith.science/pith/5FWM6ENK2CNBXOSQUVM46TRFTD.json","graph_json":"https://pith.science/api/pith-number/5FWM6ENK2CNBXOSQUVM46TRFTD/graph.json","events_json":"https://pith.science/api/pith-number/5FWM6ENK2CNBXOSQUVM46TRFTD/events.json","paper":"https://pith.science/paper/5FWM6ENK"},"agent_actions":{"view_html":"https://pith.science/pith/5FWM6ENK2CNBXOSQUVM46TRFTD","download_json":"https://pith.science/pith/5FWM6ENK2CNBXOSQUVM46TRFTD.json","view_paper":"https://pith.science/paper/5FWM6ENK","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.13210&json=true","fetch_graph":"https://pith.science/api/pith-number/5FWM6ENK2CNBXOSQUVM46TRFTD/graph.json","fetch_events":"https://pith.science/api/pith-number/5FWM6ENK2CNBXOSQUVM46TRFTD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/5FWM6ENK2CNBXOSQUVM46TRFTD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/5FWM6ENK2CNBXOSQUVM46TRFTD/action/storage_attestation","attest_author":"https://pith.science/pith/5FWM6ENK2CNBXOSQUVM46TRFTD/action/author_attestation","sign_citation":"https://pith.science/pith/5FWM6ENK2CNBXOSQUVM46TRFTD/action/citation_signature","submit_replication":"https://pith.science/pith/5FWM6ENK2CNBXOSQUVM46TRFTD/action/replication_record"}},"created_at":"2026-07-05T08:39:31.280018+00:00","updated_at":"2026-07-05T08:39:31.280018+00:00"}