{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:XXUTTKOKLH75TEY6GZQ7K2XNEL","short_pith_number":"pith:XXUTTKOK","schema_version":"1.0","canonical_sha256":"bde939a9ca59ffd9931e3661f56aed22facb3951fc63452a5eb4a16e9159e252","source":{"kind":"arxiv","id":"2605.28440","version":1},"attestation_state":"computed","paper":{"title":"AdaDPO: Self-Adaptive Direct Preference Optimization with Balanced Gradient Updates","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Madalina Ciobanu, Qingqing Mao, Ritankar Das, Shaolong Chen","submitted_at":"2026-05-27T13:05:49Z","abstract_excerpt":"DPO has become a widely adopted alternative to RLHF for aligning LLMs with human preferences, eliminating the need for a separate reward model or RL loop. Recent theoretical analysis uncovers an asymmetric gradient behavior in DPO: the loss suppresses dispreferred responses substantially faster than it promotes preferred ones, causing the model to learn to avoid bad answers rather than to generate good ones. We propose AdaDPO, a Self-Adaptive variant of the DPO algorithm that introduces per-preference-pair, stop-gradient-based coefficients derived directly from the policy model's generation pr"},"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":"2605.28440","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-27T13:05:49Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"44c53748bb192232ee2ca7590086e392a5ee49a2c0aef57e5ef1f660cce87866","abstract_canon_sha256":"29c8b5ae52e430eeacdbb4efa7233d628d56e4a687a7aab80e2c8b9cf01a0fb6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-28T01:05:18.194447Z","signature_b64":"Cb/084EQEweWBQ/rOC30j6yyZPqHmLBQlJQ/a65qDRGtZBwmnYE/VnaXJj4CEZ/d9xE0SvcXmCcm/dHvdwJgAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"bde939a9ca59ffd9931e3661f56aed22facb3951fc63452a5eb4a16e9159e252","last_reissued_at":"2026-05-28T01:05:18.194027Z","signature_status":"signed_v1","first_computed_at":"2026-05-28T01:05:18.194027Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"AdaDPO: Self-Adaptive Direct Preference Optimization with Balanced Gradient Updates","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Madalina Ciobanu, Qingqing Mao, Ritankar Das, Shaolong Chen","submitted_at":"2026-05-27T13:05:49Z","abstract_excerpt":"DPO has become a widely adopted alternative to RLHF for aligning LLMs with human preferences, eliminating the need for a separate reward model or RL loop. Recent theoretical analysis uncovers an asymmetric gradient behavior in DPO: the loss suppresses dispreferred responses substantially faster than it promotes preferred ones, causing the model to learn to avoid bad answers rather than to generate good ones. We propose AdaDPO, a Self-Adaptive variant of the DPO algorithm that introduces per-preference-pair, stop-gradient-based coefficients derived directly from the policy model's generation pr"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.28440","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/2605.28440/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":"2605.28440","created_at":"2026-05-28T01:05:18.194092+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.28440v1","created_at":"2026-05-28T01:05:18.194092+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.28440","created_at":"2026-05-28T01:05:18.194092+00:00"},{"alias_kind":"pith_short_12","alias_value":"XXUTTKOKLH75","created_at":"2026-05-28T01:05:18.194092+00:00"},{"alias_kind":"pith_short_16","alias_value":"XXUTTKOKLH75TEY6","created_at":"2026-05-28T01:05:18.194092+00:00"},{"alias_kind":"pith_short_8","alias_value":"XXUTTKOK","created_at":"2026-05-28T01:05:18.194092+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/XXUTTKOKLH75TEY6GZQ7K2XNEL","json":"https://pith.science/pith/XXUTTKOKLH75TEY6GZQ7K2XNEL.json","graph_json":"https://pith.science/api/pith-number/XXUTTKOKLH75TEY6GZQ7K2XNEL/graph.json","events_json":"https://pith.science/api/pith-number/XXUTTKOKLH75TEY6GZQ7K2XNEL/events.json","paper":"https://pith.science/paper/XXUTTKOK"},"agent_actions":{"view_html":"https://pith.science/pith/XXUTTKOKLH75TEY6GZQ7K2XNEL","download_json":"https://pith.science/pith/XXUTTKOKLH75TEY6GZQ7K2XNEL.json","view_paper":"https://pith.science/paper/XXUTTKOK","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.28440&json=true","fetch_graph":"https://pith.science/api/pith-number/XXUTTKOKLH75TEY6GZQ7K2XNEL/graph.json","fetch_events":"https://pith.science/api/pith-number/XXUTTKOKLH75TEY6GZQ7K2XNEL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/XXUTTKOKLH75TEY6GZQ7K2XNEL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/XXUTTKOKLH75TEY6GZQ7K2XNEL/action/storage_attestation","attest_author":"https://pith.science/pith/XXUTTKOKLH75TEY6GZQ7K2XNEL/action/author_attestation","sign_citation":"https://pith.science/pith/XXUTTKOKLH75TEY6GZQ7K2XNEL/action/citation_signature","submit_replication":"https://pith.science/pith/XXUTTKOKLH75TEY6GZQ7K2XNEL/action/replication_record"}},"created_at":"2026-05-28T01:05:18.194092+00:00","updated_at":"2026-05-28T01:05:18.194092+00:00"}