{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:BRHDIKEQFYVUEYNYHJOFDMYEPQ","short_pith_number":"pith:BRHDIKEQ","schema_version":"1.0","canonical_sha256":"0c4e3428902e2b4261b83a5c51b3047c0a94302027400a610ede05b88bd8065f","source":{"kind":"arxiv","id":"2404.11999","version":5},"attestation_state":"computed","paper":{"title":"Token-level Direct Preference Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Guoqing Liu, Haifeng Zhang, Jun Wang, Ning Yang, Weiyu Ma, Yongcheng Zeng","submitted_at":"2024-04-18T08:49:38Z","abstract_excerpt":"Fine-tuning pre-trained Large Language Models (LLMs) is essential to align them with human values and intentions. This process often utilizes methods like pairwise comparisons and KL divergence against a reference LLM, focusing on the evaluation of full answers generated by the models. However, the generation of these responses occurs in a token level, following a sequential, auto-regressive fashion. In this paper, we introduce Token-level Direct Preference Optimization (TDPO), a novel approach to align LLMs with human preferences by optimizing policy at the token level. Unlike previous method"},"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":"2404.11999","kind":"arxiv","version":5},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-04-18T08:49:38Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"c8c0399dac725367951e6eb9468b1dd4d912ba6f1d04e21f3e0447c3010abaf4","abstract_canon_sha256":"dd976dfb6b2c6d8b2bbfe68ea9055987125151c152a7642ca04efae43c0996e9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:00:59.505339Z","signature_b64":"RcpSo7bSt9r0QXCsT/tnaYBJ55hbnfB/wmhQ+YxgHv3LsWtyY2S0bZxTIrM2a/CLt0MtO395u51y7Xn7Q1QiCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0c4e3428902e2b4261b83a5c51b3047c0a94302027400a610ede05b88bd8065f","last_reissued_at":"2026-07-05T09:00:59.504850Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:00:59.504850Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Token-level Direct Preference Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Guoqing Liu, Haifeng Zhang, Jun Wang, Ning Yang, Weiyu Ma, Yongcheng Zeng","submitted_at":"2024-04-18T08:49:38Z","abstract_excerpt":"Fine-tuning pre-trained Large Language Models (LLMs) is essential to align them with human values and intentions. This process often utilizes methods like pairwise comparisons and KL divergence against a reference LLM, focusing on the evaluation of full answers generated by the models. However, the generation of these responses occurs in a token level, following a sequential, auto-regressive fashion. In this paper, we introduce Token-level Direct Preference Optimization (TDPO), a novel approach to align LLMs with human preferences by optimizing policy at the token level. Unlike previous method"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2404.11999","kind":"arxiv","version":5},"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/2404.11999/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":"2404.11999","created_at":"2026-07-05T09:00:59.504909+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.11999v5","created_at":"2026-07-05T09:00:59.504909+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.11999","created_at":"2026-07-05T09:00:59.504909+00:00"},{"alias_kind":"pith_short_12","alias_value":"BRHDIKEQFYVU","created_at":"2026-07-05T09:00:59.504909+00:00"},{"alias_kind":"pith_short_16","alias_value":"BRHDIKEQFYVUEYNY","created_at":"2026-07-05T09:00:59.504909+00:00"},{"alias_kind":"pith_short_8","alias_value":"BRHDIKEQ","created_at":"2026-07-05T09:00:59.504909+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":18,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.12590","citing_title":"Analyzing and Improving Fine-grained Preference Optimization in Medical LVLMs","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08035","citing_title":"DyCo-RL: Dynamic Cross-Modal Coordination for Visual Reasoning","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12288","citing_title":"TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2606.29150","citing_title":"Flow Reasoning Models: Scaling Reasoning Through Iterative Self-Refinement","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2408.15339","citing_title":"UNA: A Unified Supervised Framework for Efficient LLM Alignment Across Feedback Types","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2508.04149","citing_title":"Difficulty-Based Preference Data Selection by DPO Implicit Reward Gap","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2509.03526","citing_title":"Enhancing Speech Large Language Models through Reinforced Behavior Alignment","ref_index":58,"is_internal_anchor":false},{"citing_arxiv_id":"2602.07832","citing_title":"rePIRL: Learn PRM with Inverse RL for LLM Reasoning","ref_index":42,"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":37,"is_internal_anchor":false},{"citing_arxiv_id":"2507.05179","citing_title":"From Fragments to Facts: A Curriculum-Driven DPO Approach for Generating Hindi News Veracity Explanations","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2512.05929","citing_title":"LLM Harms: A Taxonomy and Discussion","ref_index":226,"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":37,"is_internal_anchor":false},{"citing_arxiv_id":"2602.23964","citing_title":"RAD-DPO: Robust Adaptive Denoising Direct Preference Optimization for Generative Retrieval in E-commerce","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2605.12288","citing_title":"TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching","ref_index":192,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00380","citing_title":"ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2604.11259","citing_title":"Mobile GUI Agent Privacy Personalization with Trajectory Induced Preference Optimization","ref_index":39,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00380","citing_title":"ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14379","citing_title":"Step-level Denoising-time Diffusion Alignment with Multiple Objectives","ref_index":40,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BRHDIKEQFYVUEYNYHJOFDMYEPQ","json":"https://pith.science/pith/BRHDIKEQFYVUEYNYHJOFDMYEPQ.json","graph_json":"https://pith.science/api/pith-number/BRHDIKEQFYVUEYNYHJOFDMYEPQ/graph.json","events_json":"https://pith.science/api/pith-number/BRHDIKEQFYVUEYNYHJOFDMYEPQ/events.json","paper":"https://pith.science/paper/BRHDIKEQ"},"agent_actions":{"view_html":"https://pith.science/pith/BRHDIKEQFYVUEYNYHJOFDMYEPQ","download_json":"https://pith.science/pith/BRHDIKEQFYVUEYNYHJOFDMYEPQ.json","view_paper":"https://pith.science/paper/BRHDIKEQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.11999&json=true","fetch_graph":"https://pith.science/api/pith-number/BRHDIKEQFYVUEYNYHJOFDMYEPQ/graph.json","fetch_events":"https://pith.science/api/pith-number/BRHDIKEQFYVUEYNYHJOFDMYEPQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BRHDIKEQFYVUEYNYHJOFDMYEPQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BRHDIKEQFYVUEYNYHJOFDMYEPQ/action/storage_attestation","attest_author":"https://pith.science/pith/BRHDIKEQFYVUEYNYHJOFDMYEPQ/action/author_attestation","sign_citation":"https://pith.science/pith/BRHDIKEQFYVUEYNYHJOFDMYEPQ/action/citation_signature","submit_replication":"https://pith.science/pith/BRHDIKEQFYVUEYNYHJOFDMYEPQ/action/replication_record"}},"created_at":"2026-07-05T09:00:59.504909+00:00","updated_at":"2026-07-05T09:00:59.504909+00:00"}