{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:NSTWLEWH45NCICDHQOPENPERUE","short_pith_number":"pith:NSTWLEWH","schema_version":"1.0","canonical_sha256":"6ca76592c7e75a240867839e46bc91a13ec4c88da170a77705d227bb14c9e2cc","source":{"kind":"arxiv","id":"2402.05749","version":2},"attestation_state":"computed","paper":{"title":"Generalized Preference Optimization: A Unified Approach to Offline Alignment","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Bernardo \\'Avila Pires, Bilal Piot, Daniele Calandriello, Mark Rowland, Michal Valko, Pierre Harvey Richemond, R\\'emi Munos, Yunhao Tang, Zeyu Zheng, Zhaohan Daniel Guo","submitted_at":"2024-02-08T15:33:09Z","abstract_excerpt":"Offline preference optimization allows fine-tuning large models directly from offline data, and has proved effective in recent alignment practices. We propose generalized preference optimization (GPO), a family of offline losses parameterized by a general class of convex functions. GPO enables a unified view over preference optimization, encompassing existing algorithms such as DPO, IPO and SLiC as special cases, while naturally introducing new variants. The GPO framework also sheds light on how offline algorithms enforce regularization, through the design of the convex function that defines t"},"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":"2402.05749","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-02-08T15:33:09Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"399b4266573587af3da7357e7348292609fbe60be04f904ea8f3a817931401b3","abstract_canon_sha256":"642ab2edff7bb673cef0be4181e62bbc3871e29cf398102641ebe4aca03b861f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:24:24.038552Z","signature_b64":"xaNlmWrGwiY3Km4azUhXXasWcSbxhAfS4Rmxp6v0+ywKiP5UFphqzb1X2Y/Tt+2pPASr4U56LnQBk/wiV1nCDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6ca76592c7e75a240867839e46bc91a13ec4c88da170a77705d227bb14c9e2cc","last_reissued_at":"2026-07-05T08:24:24.038014Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:24:24.038014Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Generalized Preference Optimization: A Unified Approach to Offline Alignment","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Bernardo \\'Avila Pires, Bilal Piot, Daniele Calandriello, Mark Rowland, Michal Valko, Pierre Harvey Richemond, R\\'emi Munos, Yunhao Tang, Zeyu Zheng, Zhaohan Daniel Guo","submitted_at":"2024-02-08T15:33:09Z","abstract_excerpt":"Offline preference optimization allows fine-tuning large models directly from offline data, and has proved effective in recent alignment practices. We propose generalized preference optimization (GPO), a family of offline losses parameterized by a general class of convex functions. GPO enables a unified view over preference optimization, encompassing existing algorithms such as DPO, IPO and SLiC as special cases, while naturally introducing new variants. The GPO framework also sheds light on how offline algorithms enforce regularization, through the design of the convex function that defines t"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.05749","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/2402.05749/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":"2402.05749","created_at":"2026-07-05T08:24:24.038082+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.05749v2","created_at":"2026-07-05T08:24:24.038082+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.05749","created_at":"2026-07-05T08:24:24.038082+00:00"},{"alias_kind":"pith_short_12","alias_value":"NSTWLEWH45NC","created_at":"2026-07-05T08:24:24.038082+00:00"},{"alias_kind":"pith_short_16","alias_value":"NSTWLEWH45NCICDH","created_at":"2026-07-05T08:24:24.038082+00:00"},{"alias_kind":"pith_short_8","alias_value":"NSTWLEWH","created_at":"2026-07-05T08:24:24.038082+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2502.01237","citing_title":"The Differences Between Direct Alignment Algorithms are a Blur","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12288","citing_title":"TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching","ref_index":139,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12288","citing_title":"TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching","ref_index":139,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06987","citing_title":"Response Time Enhances Alignment with Heterogeneous Preferences","ref_index":163,"is_internal_anchor":false},{"citing_arxiv_id":"2604.05341","citing_title":"Curr-RLCER:Curriculum Reinforcement Learning For Coherence Explainable Recommendation","ref_index":21,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NSTWLEWH45NCICDHQOPENPERUE","json":"https://pith.science/pith/NSTWLEWH45NCICDHQOPENPERUE.json","graph_json":"https://pith.science/api/pith-number/NSTWLEWH45NCICDHQOPENPERUE/graph.json","events_json":"https://pith.science/api/pith-number/NSTWLEWH45NCICDHQOPENPERUE/events.json","paper":"https://pith.science/paper/NSTWLEWH"},"agent_actions":{"view_html":"https://pith.science/pith/NSTWLEWH45NCICDHQOPENPERUE","download_json":"https://pith.science/pith/NSTWLEWH45NCICDHQOPENPERUE.json","view_paper":"https://pith.science/paper/NSTWLEWH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.05749&json=true","fetch_graph":"https://pith.science/api/pith-number/NSTWLEWH45NCICDHQOPENPERUE/graph.json","fetch_events":"https://pith.science/api/pith-number/NSTWLEWH45NCICDHQOPENPERUE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NSTWLEWH45NCICDHQOPENPERUE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NSTWLEWH45NCICDHQOPENPERUE/action/storage_attestation","attest_author":"https://pith.science/pith/NSTWLEWH45NCICDHQOPENPERUE/action/author_attestation","sign_citation":"https://pith.science/pith/NSTWLEWH45NCICDHQOPENPERUE/action/citation_signature","submit_replication":"https://pith.science/pith/NSTWLEWH45NCICDHQOPENPERUE/action/replication_record"}},"created_at":"2026-07-05T08:24:24.038082+00:00","updated_at":"2026-07-05T08:24:24.038082+00:00"}