{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:CF4Z7IRIXGLQH2PCSABMOOQ6WR","short_pith_number":"pith:CF4Z7IRI","schema_version":"1.0","canonical_sha256":"11799fa228b99703e9e29002c73a1eb447f600e514a96b15930fc7348d6726d4","source":{"kind":"arxiv","id":"2603.07211","version":2},"attestation_state":"computed","paper":{"title":"CompassDPO: Dynamics-Controlled Direct Preference Optimization for Robust Safety Alignment","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Haokai Ma, Hao Zhan, Jilong Liu, Pengyang Shao, Richang Hong, Wei Qin, Wenjian Tao, Yonghui Yang","submitted_at":"2026-03-07T13:30:53Z","abstract_excerpt":"Direct Preference Optimization (DPO) has become a standard framework for safety alignment, but its reliance on pairwise preference updates makes training sensitive to imperfect supervision. Existing robust DPO methods often address this sensitivity through global loss corrections or external data-level interventions, while largely overlooking how unreliable comparisons distort batch-level optimization dynamics. We propose CompassDPO, a reward-free DPO framework that stabilizes preference optimization through dynamics control. Using the implicit DPO reward margin as a training-time compass, Com"},"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":"2603.07211","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-07T13:30:53Z","cross_cats_sorted":[],"title_canon_sha256":"c6553342df53be9419633df1c0eff01fd8748acee72bcf5dbeef82eb4841ba3a","abstract_canon_sha256":"0627562e585067ae19b327d8299f37db25e627e3ce9e3384f3c6aa89f21ffa3d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-27T01:05:45.661591Z","signature_b64":"DF2Kg6cIvDalh2JeCDsyklZQAKdFzW0F4zhEhJ7CVlcQmjKDeqjhRTIAal3JZXzVvR+GFjDfdGOppyQESHWzCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"11799fa228b99703e9e29002c73a1eb447f600e514a96b15930fc7348d6726d4","last_reissued_at":"2026-05-27T01:05:45.660734Z","signature_status":"signed_v1","first_computed_at":"2026-05-27T01:05:45.660734Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"CompassDPO: Dynamics-Controlled Direct Preference Optimization for Robust Safety Alignment","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Haokai Ma, Hao Zhan, Jilong Liu, Pengyang Shao, Richang Hong, Wei Qin, Wenjian Tao, Yonghui Yang","submitted_at":"2026-03-07T13:30:53Z","abstract_excerpt":"Direct Preference Optimization (DPO) has become a standard framework for safety alignment, but its reliance on pairwise preference updates makes training sensitive to imperfect supervision. Existing robust DPO methods often address this sensitivity through global loss corrections or external data-level interventions, while largely overlooking how unreliable comparisons distort batch-level optimization dynamics. We propose CompassDPO, a reward-free DPO framework that stabilizes preference optimization through dynamics control. Using the implicit DPO reward margin as a training-time compass, Com"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.07211","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/2603.07211/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":"2603.07211","created_at":"2026-05-27T01:05:45.660851+00:00"},{"alias_kind":"arxiv_version","alias_value":"2603.07211v2","created_at":"2026-05-27T01:05:45.660851+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.07211","created_at":"2026-05-27T01:05:45.660851+00:00"},{"alias_kind":"pith_short_12","alias_value":"CF4Z7IRIXGLQ","created_at":"2026-05-27T01:05:45.660851+00:00"},{"alias_kind":"pith_short_16","alias_value":"CF4Z7IRIXGLQH2PC","created_at":"2026-05-27T01:05:45.660851+00:00"},{"alias_kind":"pith_short_8","alias_value":"CF4Z7IRI","created_at":"2026-05-27T01:05:45.660851+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/CF4Z7IRIXGLQH2PCSABMOOQ6WR","json":"https://pith.science/pith/CF4Z7IRIXGLQH2PCSABMOOQ6WR.json","graph_json":"https://pith.science/api/pith-number/CF4Z7IRIXGLQH2PCSABMOOQ6WR/graph.json","events_json":"https://pith.science/api/pith-number/CF4Z7IRIXGLQH2PCSABMOOQ6WR/events.json","paper":"https://pith.science/paper/CF4Z7IRI"},"agent_actions":{"view_html":"https://pith.science/pith/CF4Z7IRIXGLQH2PCSABMOOQ6WR","download_json":"https://pith.science/pith/CF4Z7IRIXGLQH2PCSABMOOQ6WR.json","view_paper":"https://pith.science/paper/CF4Z7IRI","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2603.07211&json=true","fetch_graph":"https://pith.science/api/pith-number/CF4Z7IRIXGLQH2PCSABMOOQ6WR/graph.json","fetch_events":"https://pith.science/api/pith-number/CF4Z7IRIXGLQH2PCSABMOOQ6WR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CF4Z7IRIXGLQH2PCSABMOOQ6WR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CF4Z7IRIXGLQH2PCSABMOOQ6WR/action/storage_attestation","attest_author":"https://pith.science/pith/CF4Z7IRIXGLQH2PCSABMOOQ6WR/action/author_attestation","sign_citation":"https://pith.science/pith/CF4Z7IRIXGLQH2PCSABMOOQ6WR/action/citation_signature","submit_replication":"https://pith.science/pith/CF4Z7IRIXGLQH2PCSABMOOQ6WR/action/replication_record"}},"created_at":"2026-05-27T01:05:45.660851+00:00","updated_at":"2026-05-27T01:05:45.660851+00:00"}