{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:VDQNNFDM7HH726DYZD45TQ7B3Z","short_pith_number":"pith:VDQNNFDM","schema_version":"1.0","canonical_sha256":"a8e0d6946cf9cffd7878c8f9d9c3e1de4df7c4e2f6b4ea3184ebe7b3b35d03bb","source":{"kind":"arxiv","id":"2504.02193","version":3},"attestation_state":"computed","paper":{"title":"More is Less: The Pitfalls of Multi-Model Synthetic Preference Data in DPO Safety Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Ananth Grama, Bolian Li, David Cho, Junyuan Hong, Runjin Chen, Ruqi Zhang, Tianlong Chen, Yifan Wang, Yihe Deng, Zhangyang Wang","submitted_at":"2025-04-03T00:36:40Z","abstract_excerpt":"Aligning large language models (LLMs) with human values is an increasingly critical step in post-training. Direct Preference Optimization (DPO) has emerged as a simple, yet effective alternative to reinforcement learning from human feedback (RLHF). Synthetic preference data with its low cost and high quality enable effective alignment through single- or multi-model generated preference data. Our study reveals a striking, safety-specific phenomenon associated with DPO alignment: Although multi-model generated data enhances performance on general tasks (ARC, Hellaswag, MMLU, TruthfulQA, Winogran"},"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":"2504.02193","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-04-03T00:36:40Z","cross_cats_sorted":[],"title_canon_sha256":"dab600cf8e433591072ae74b0fcde03ad5f2adac61a8b676e2a25209f8f062ff","abstract_canon_sha256":"6f53b93b42543c5b5f424b9ee7081304ff6ffeb7f322a7a3b01ee0e6a11abe94"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:44:03.780845Z","signature_b64":"s4OEdm9INGNNf7A6zsPE+pOcDpCjPAa0N2wdm28pDdIb6VHUrovrJlAUOOrYv+sn3SaBKrE6YB4fVK138EeAAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a8e0d6946cf9cffd7878c8f9d9c3e1de4df7c4e2f6b4ea3184ebe7b3b35d03bb","last_reissued_at":"2026-07-05T11:44:03.780297Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:44:03.780297Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"More is Less: The Pitfalls of Multi-Model Synthetic Preference Data in DPO Safety Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Ananth Grama, Bolian Li, David Cho, Junyuan Hong, Runjin Chen, Ruqi Zhang, Tianlong Chen, Yifan Wang, Yihe Deng, Zhangyang Wang","submitted_at":"2025-04-03T00:36:40Z","abstract_excerpt":"Aligning large language models (LLMs) with human values is an increasingly critical step in post-training. Direct Preference Optimization (DPO) has emerged as a simple, yet effective alternative to reinforcement learning from human feedback (RLHF). Synthetic preference data with its low cost and high quality enable effective alignment through single- or multi-model generated preference data. Our study reveals a striking, safety-specific phenomenon associated with DPO alignment: Although multi-model generated data enhances performance on general tasks (ARC, Hellaswag, MMLU, TruthfulQA, Winogran"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.02193","kind":"arxiv","version":3},"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/2504.02193/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":"2504.02193","created_at":"2026-07-05T11:44:03.780371+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.02193v3","created_at":"2026-07-05T11:44:03.780371+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.02193","created_at":"2026-07-05T11:44:03.780371+00:00"},{"alias_kind":"pith_short_12","alias_value":"VDQNNFDM7HH7","created_at":"2026-07-05T11:44:03.780371+00:00"},{"alias_kind":"pith_short_16","alias_value":"VDQNNFDM7HH726DY","created_at":"2026-07-05T11:44:03.780371+00:00"},{"alias_kind":"pith_short_8","alias_value":"VDQNNFDM","created_at":"2026-07-05T11:44:03.780371+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2502.05159","citing_title":"A Lightweight Method to Disrupt Memorized Sequences in LLM","ref_index":63,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VDQNNFDM7HH726DYZD45TQ7B3Z","json":"https://pith.science/pith/VDQNNFDM7HH726DYZD45TQ7B3Z.json","graph_json":"https://pith.science/api/pith-number/VDQNNFDM7HH726DYZD45TQ7B3Z/graph.json","events_json":"https://pith.science/api/pith-number/VDQNNFDM7HH726DYZD45TQ7B3Z/events.json","paper":"https://pith.science/paper/VDQNNFDM"},"agent_actions":{"view_html":"https://pith.science/pith/VDQNNFDM7HH726DYZD45TQ7B3Z","download_json":"https://pith.science/pith/VDQNNFDM7HH726DYZD45TQ7B3Z.json","view_paper":"https://pith.science/paper/VDQNNFDM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.02193&json=true","fetch_graph":"https://pith.science/api/pith-number/VDQNNFDM7HH726DYZD45TQ7B3Z/graph.json","fetch_events":"https://pith.science/api/pith-number/VDQNNFDM7HH726DYZD45TQ7B3Z/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VDQNNFDM7HH726DYZD45TQ7B3Z/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VDQNNFDM7HH726DYZD45TQ7B3Z/action/storage_attestation","attest_author":"https://pith.science/pith/VDQNNFDM7HH726DYZD45TQ7B3Z/action/author_attestation","sign_citation":"https://pith.science/pith/VDQNNFDM7HH726DYZD45TQ7B3Z/action/citation_signature","submit_replication":"https://pith.science/pith/VDQNNFDM7HH726DYZD45TQ7B3Z/action/replication_record"}},"created_at":"2026-07-05T11:44:03.780371+00:00","updated_at":"2026-07-05T11:44:03.780371+00:00"}