{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:TYUR5VO5Q3IF7UIQY4GEYZ7MXB","short_pith_number":"pith:TYUR5VO5","schema_version":"1.0","canonical_sha256":"9e291ed5dd86d05fd110c70c4c67ecb86bd935a84972881d4df420273b7c7f94","source":{"kind":"arxiv","id":"2404.09656","version":4},"attestation_state":"computed","paper":{"title":"Learn Your Reference Model for Real Good Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Alexey Gorbatovski, Alexey Malakhov, Boris Shaposhnikov, Daniil Gavrilov, Ian Maksimov, Nikita Balagansky, Nikita Surnachev, Yaroslav Aksenov","submitted_at":"2024-04-15T10:44:31Z","abstract_excerpt":"Despite the fact that offline methods for Large Language Models (LLMs) alignment do not require a direct reward model, they remain susceptible to overoptimization. This issue arises when the trained model deviates excessively from the reference policy, leading to a decrease in sample quality. We propose a new paradigm of offline alignment methods, called Trust Region (including variants TR-DPO, TR-IPO, TR-KTO), which dynamically updates the reference policy throughout the training process. Our results show that TR alignment methods effectively mitigate overoptimization, enabling models to main"},"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":"2404.09656","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-15T10:44:31Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"ba976991f7c1b30ff4ac3769e8a1030d9bd328012af0e987728d61a27aedb338","abstract_canon_sha256":"d66615526f330a9264996bedb3247ad7eba842a57e6e33363600f93a0e3c8d0d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:19:23.404993Z","signature_b64":"2YLoY5Bhm1dH/R8Zfo/2So41+Uxc8LwmTwsDHaynWzQoaZEW+GDo5TukBa7081w4X+BH0bGoTiWaCSlxSHLnDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"9e291ed5dd86d05fd110c70c4c67ecb86bd935a84972881d4df420273b7c7f94","last_reissued_at":"2026-07-05T10:19:23.404437Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:19:23.404437Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Learn Your Reference Model for Real Good Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Alexey Gorbatovski, Alexey Malakhov, Boris Shaposhnikov, Daniil Gavrilov, Ian Maksimov, Nikita Balagansky, Nikita Surnachev, Yaroslav Aksenov","submitted_at":"2024-04-15T10:44:31Z","abstract_excerpt":"Despite the fact that offline methods for Large Language Models (LLMs) alignment do not require a direct reward model, they remain susceptible to overoptimization. This issue arises when the trained model deviates excessively from the reference policy, leading to a decrease in sample quality. We propose a new paradigm of offline alignment methods, called Trust Region (including variants TR-DPO, TR-IPO, TR-KTO), which dynamically updates the reference policy throughout the training process. Our results show that TR alignment methods effectively mitigate overoptimization, enabling models to main"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2404.09656","kind":"arxiv","version":4},"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/2404.09656/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":"2404.09656","created_at":"2026-07-05T10:19:23.404500+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.09656v4","created_at":"2026-07-05T10:19:23.404500+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.09656","created_at":"2026-07-05T10:19:23.404500+00:00"},{"alias_kind":"pith_short_12","alias_value":"TYUR5VO5Q3IF","created_at":"2026-07-05T10:19:23.404500+00:00"},{"alias_kind":"pith_short_16","alias_value":"TYUR5VO5Q3IF7UIQ","created_at":"2026-07-05T10:19:23.404500+00:00"},{"alias_kind":"pith_short_8","alias_value":"TYUR5VO5","created_at":"2026-07-05T10:19:23.404500+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.23398","citing_title":"TPMM-DPO: Trajectory-aware Preference-guided Model Merging for Iterative Direct Preference Optimization","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2502.01237","citing_title":"The Differences Between Direct Alignment Algorithms are a Blur","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2602.06239","citing_title":"Provably avoiding over-optimization in Direct Preference Optimization without knowing the data distribution","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2411.10442","citing_title":"Enhancing the Reasoning Ability of Multimodal Large Language Models via Mixed Preference Optimization","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2602.06239","citing_title":"Provably avoiding over-optimization in Direct Preference Optimization without knowing the data distribution","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24804","citing_title":"Intrinsic Mutual Information as a Modulator for Preference Optimization","ref_index":1,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/TYUR5VO5Q3IF7UIQY4GEYZ7MXB","json":"https://pith.science/pith/TYUR5VO5Q3IF7UIQY4GEYZ7MXB.json","graph_json":"https://pith.science/api/pith-number/TYUR5VO5Q3IF7UIQY4GEYZ7MXB/graph.json","events_json":"https://pith.science/api/pith-number/TYUR5VO5Q3IF7UIQY4GEYZ7MXB/events.json","paper":"https://pith.science/paper/TYUR5VO5"},"agent_actions":{"view_html":"https://pith.science/pith/TYUR5VO5Q3IF7UIQY4GEYZ7MXB","download_json":"https://pith.science/pith/TYUR5VO5Q3IF7UIQY4GEYZ7MXB.json","view_paper":"https://pith.science/paper/TYUR5VO5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.09656&json=true","fetch_graph":"https://pith.science/api/pith-number/TYUR5VO5Q3IF7UIQY4GEYZ7MXB/graph.json","fetch_events":"https://pith.science/api/pith-number/TYUR5VO5Q3IF7UIQY4GEYZ7MXB/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/TYUR5VO5Q3IF7UIQY4GEYZ7MXB/action/timestamp_anchor","attest_storage":"https://pith.science/pith/TYUR5VO5Q3IF7UIQY4GEYZ7MXB/action/storage_attestation","attest_author":"https://pith.science/pith/TYUR5VO5Q3IF7UIQY4GEYZ7MXB/action/author_attestation","sign_citation":"https://pith.science/pith/TYUR5VO5Q3IF7UIQY4GEYZ7MXB/action/citation_signature","submit_replication":"https://pith.science/pith/TYUR5VO5Q3IF7UIQY4GEYZ7MXB/action/replication_record"}},"created_at":"2026-07-05T10:19:23.404500+00:00","updated_at":"2026-07-05T10:19:23.404500+00:00"}