{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:W7WKZ35QGWEDNZUBOSVHJEC2QK","short_pith_number":"pith:W7WKZ35Q","schema_version":"1.0","canonical_sha256":"b7ecacefb0358836e68174aa74905a82a708ff25b2a27ae3d0344f0516db7417","source":{"kind":"arxiv","id":"2503.02846","version":1},"attestation_state":"computed","paper":{"title":"Mask-DPO: Generalizable Fine-grained Factuality Alignment of LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Chengqi Lyu, Dahua Lin, Kai Chen, Wenwei Zhang, Yuzhe Gu","submitted_at":"2025-03-04T18:20:24Z","abstract_excerpt":"Large language models (LLMs) exhibit hallucinations (i.e., unfaithful or nonsensical information) when serving as AI assistants in various domains. Since hallucinations always come with truthful content in the LLM responses, previous factuality alignment methods that conduct response-level preference learning inevitably introduced noises during training. Therefore, this paper proposes a fine-grained factuality alignment method based on Direct Preference Optimization (DPO), called Mask-DPO. Incorporating sentence-level factuality as mask signals, Mask-DPO only learns from factually correct sent"},"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":"2503.02846","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-03-04T18:20:24Z","cross_cats_sorted":[],"title_canon_sha256":"d1ca2fdae7bc634b9fb9741579e1c1d297ceb86533a0fd1075564ba6978f6025","abstract_canon_sha256":"a284e7d2cdf3b3742404a2368e7c685a0c8ff21142948c162f90bc4ecbdec56e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:24:15.185020Z","signature_b64":"Zc0tauqsyBN3AplQP0kmvFjfQngs85vhI/jW4O4pnUjYsOonNAW32uuPQIwk+BtWCncH+V0s7KohEumUFDgsCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b7ecacefb0358836e68174aa74905a82a708ff25b2a27ae3d0344f0516db7417","last_reissued_at":"2026-07-05T10:24:15.184470Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:24:15.184470Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mask-DPO: Generalizable Fine-grained Factuality Alignment of LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Chengqi Lyu, Dahua Lin, Kai Chen, Wenwei Zhang, Yuzhe Gu","submitted_at":"2025-03-04T18:20:24Z","abstract_excerpt":"Large language models (LLMs) exhibit hallucinations (i.e., unfaithful or nonsensical information) when serving as AI assistants in various domains. Since hallucinations always come with truthful content in the LLM responses, previous factuality alignment methods that conduct response-level preference learning inevitably introduced noises during training. Therefore, this paper proposes a fine-grained factuality alignment method based on Direct Preference Optimization (DPO), called Mask-DPO. Incorporating sentence-level factuality as mask signals, Mask-DPO only learns from factually correct sent"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.02846","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/2503.02846/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":"2503.02846","created_at":"2026-07-05T10:24:15.184530+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.02846v1","created_at":"2026-07-05T10:24:15.184530+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.02846","created_at":"2026-07-05T10:24:15.184530+00:00"},{"alias_kind":"pith_short_12","alias_value":"W7WKZ35QGWED","created_at":"2026-07-05T10:24:15.184530+00:00"},{"alias_kind":"pith_short_16","alias_value":"W7WKZ35QGWEDNZUB","created_at":"2026-07-05T10:24:15.184530+00:00"},{"alias_kind":"pith_short_8","alias_value":"W7WKZ35Q","created_at":"2026-07-05T10:24:15.184530+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.12590","citing_title":"Analyzing and Improving Fine-grained Preference Optimization in Medical LVLMs","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03503","citing_title":"ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2507.12455","citing_title":"Mitigating Object Hallucinations via Sentence-Level Early Intervention","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2511.02626","citing_title":"Understanding New-Knowledge-Induced Factual Hallucinations in LLMs: Analysis and Interpretation","ref_index":10,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/W7WKZ35QGWEDNZUBOSVHJEC2QK","json":"https://pith.science/pith/W7WKZ35QGWEDNZUBOSVHJEC2QK.json","graph_json":"https://pith.science/api/pith-number/W7WKZ35QGWEDNZUBOSVHJEC2QK/graph.json","events_json":"https://pith.science/api/pith-number/W7WKZ35QGWEDNZUBOSVHJEC2QK/events.json","paper":"https://pith.science/paper/W7WKZ35Q"},"agent_actions":{"view_html":"https://pith.science/pith/W7WKZ35QGWEDNZUBOSVHJEC2QK","download_json":"https://pith.science/pith/W7WKZ35QGWEDNZUBOSVHJEC2QK.json","view_paper":"https://pith.science/paper/W7WKZ35Q","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.02846&json=true","fetch_graph":"https://pith.science/api/pith-number/W7WKZ35QGWEDNZUBOSVHJEC2QK/graph.json","fetch_events":"https://pith.science/api/pith-number/W7WKZ35QGWEDNZUBOSVHJEC2QK/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/W7WKZ35QGWEDNZUBOSVHJEC2QK/action/timestamp_anchor","attest_storage":"https://pith.science/pith/W7WKZ35QGWEDNZUBOSVHJEC2QK/action/storage_attestation","attest_author":"https://pith.science/pith/W7WKZ35QGWEDNZUBOSVHJEC2QK/action/author_attestation","sign_citation":"https://pith.science/pith/W7WKZ35QGWEDNZUBOSVHJEC2QK/action/citation_signature","submit_replication":"https://pith.science/pith/W7WKZ35QGWEDNZUBOSVHJEC2QK/action/replication_record"}},"created_at":"2026-07-05T10:24:15.184530+00:00","updated_at":"2026-07-05T10:24:15.184530+00:00"}