{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:6CUDVGOGXH6V6JM7WGJJUMKLT4","short_pith_number":"pith:6CUDVGOG","schema_version":"1.0","canonical_sha256":"f0a83a99c6b9fd5f259fb1929a314b9f26dd68d4c5233d0ae4fbd6ecfc108be7","source":{"kind":"arxiv","id":"2412.13862","version":1},"attestation_state":"computed","paper":{"title":"Energy-Based Preference Model Offers Better Offline Alignment than the Bradley-Terry Preference Model","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Hanshan Zhang, Hongfei Jiang, Junwei Bao, Yang Song, Yuzhong Hong","submitted_at":"2024-12-18T13:55:42Z","abstract_excerpt":"Since the debut of DPO, it has been shown that aligning a target LLM with human preferences via the KL-constrained RLHF loss is mathematically equivalent to a special kind of reward modeling task. Concretely, the task requires: 1) using the target LLM to parameterize the reward model, and 2) tuning the reward model so that it has a 1:1 linear relationship with the true reward. However, we identify a significant issue: the DPO loss might have multiple minimizers, of which only one satisfies the required linearity condition. The problem arises from a well-known issue of the underlying Bradley-Te"},"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":"2412.13862","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2024-12-18T13:55:42Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"0a3130b756b60c9516f46b3a2ab93c74afd0c8b792553da2cb6f5595f9a99f6f","abstract_canon_sha256":"e5b5feaac69c3794db4cdd7bdcc1eabc01c86163f9ed6b35736e54805984b871"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:51:09.916006Z","signature_b64":"O0UTpqsqe7BsXAvjlXCBXF6muodMInoApPCY5K7SvdctfNFmIKvtKdwYIbbTyWGnvY/ULMoTfhevAmIZ6t/eAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f0a83a99c6b9fd5f259fb1929a314b9f26dd68d4c5233d0ae4fbd6ecfc108be7","last_reissued_at":"2026-07-05T09:51:09.915528Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:51:09.915528Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Energy-Based Preference Model Offers Better Offline Alignment than the Bradley-Terry Preference Model","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Hanshan Zhang, Hongfei Jiang, Junwei Bao, Yang Song, Yuzhong Hong","submitted_at":"2024-12-18T13:55:42Z","abstract_excerpt":"Since the debut of DPO, it has been shown that aligning a target LLM with human preferences via the KL-constrained RLHF loss is mathematically equivalent to a special kind of reward modeling task. Concretely, the task requires: 1) using the target LLM to parameterize the reward model, and 2) tuning the reward model so that it has a 1:1 linear relationship with the true reward. However, we identify a significant issue: the DPO loss might have multiple minimizers, of which only one satisfies the required linearity condition. The problem arises from a well-known issue of the underlying Bradley-Te"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2412.13862","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/2412.13862/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":"2412.13862","created_at":"2026-07-05T09:51:09.915592+00:00"},{"alias_kind":"arxiv_version","alias_value":"2412.13862v1","created_at":"2026-07-05T09:51:09.915592+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2412.13862","created_at":"2026-07-05T09:51:09.915592+00:00"},{"alias_kind":"pith_short_12","alias_value":"6CUDVGOGXH6V","created_at":"2026-07-05T09:51:09.915592+00:00"},{"alias_kind":"pith_short_16","alias_value":"6CUDVGOGXH6V6JM7","created_at":"2026-07-05T09:51:09.915592+00:00"},{"alias_kind":"pith_short_8","alias_value":"6CUDVGOG","created_at":"2026-07-05T09:51:09.915592+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2604.04410","citing_title":"Relative Density Ratio Optimization for Stable and Statistically Consistent Model Alignment","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6CUDVGOGXH6V6JM7WGJJUMKLT4","json":"https://pith.science/pith/6CUDVGOGXH6V6JM7WGJJUMKLT4.json","graph_json":"https://pith.science/api/pith-number/6CUDVGOGXH6V6JM7WGJJUMKLT4/graph.json","events_json":"https://pith.science/api/pith-number/6CUDVGOGXH6V6JM7WGJJUMKLT4/events.json","paper":"https://pith.science/paper/6CUDVGOG"},"agent_actions":{"view_html":"https://pith.science/pith/6CUDVGOGXH6V6JM7WGJJUMKLT4","download_json":"https://pith.science/pith/6CUDVGOGXH6V6JM7WGJJUMKLT4.json","view_paper":"https://pith.science/paper/6CUDVGOG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2412.13862&json=true","fetch_graph":"https://pith.science/api/pith-number/6CUDVGOGXH6V6JM7WGJJUMKLT4/graph.json","fetch_events":"https://pith.science/api/pith-number/6CUDVGOGXH6V6JM7WGJJUMKLT4/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6CUDVGOGXH6V6JM7WGJJUMKLT4/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6CUDVGOGXH6V6JM7WGJJUMKLT4/action/storage_attestation","attest_author":"https://pith.science/pith/6CUDVGOGXH6V6JM7WGJJUMKLT4/action/author_attestation","sign_citation":"https://pith.science/pith/6CUDVGOGXH6V6JM7WGJJUMKLT4/action/citation_signature","submit_replication":"https://pith.science/pith/6CUDVGOGXH6V6JM7WGJJUMKLT4/action/replication_record"}},"created_at":"2026-07-05T09:51:09.915592+00:00","updated_at":"2026-07-05T09:51:09.915592+00:00"}