{"paper":{"title":"Beyond Hallucinations: Enhancing LVLMs through Hallucination-Aware Direct Preference Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"HA-DPO trains multimodal models to prefer accurate image descriptions over hallucinatory ones by optimizing on paired responses.","cross_cats":["cs.CL"],"primary_cat":"cs.CV","authors_text":"Bin Wang, Conghui He, Jiaqi Wang, Linke Ouyang, Xiaoyi Dong, Zhiyuan Zhao","submitted_at":"2023-11-28T14:54:37Z","abstract_excerpt":"Multimodal large language models have made significant advancements in recent years, yet they still suffer from a common issue known as the \"hallucination problem\", in which the models generate textual descriptions that inaccurately depict or entirely fabricate content from associated images. This paper introduces a novel solution, Hallucination-Aware Direct Preference Optimization (HA-DPO), which reframes the hallucination problem as a preference selection task. The model is trained to favor the non-hallucinating response when presented with two responses of the same image (one accurate and o"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"When applied to three mainstream multimodal models, HA-DPO significantly reduced hallucination issues and amplified the models' generalization capabilities. Notably, the MiniGPT-4 model, when enhanced with HA-DPO, demonstrated a substantial improvement: POPE accuracy rose from 51.13% to 86.13% (an absolute improvement of 35%), and the MME score surged from 932.00 to 1326.46 (a relative improvement of 42.32%).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The constructed positive and negative sample pairs are high-quality, style-consistent, and free of new biases that could undermine preference learning or generalization beyond the tested benchmarks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"HA-DPO reframes hallucination reduction in LVLMs as direct preference optimization over style-consistent positive and negative response pairs, yielding large gains such as 35-point POPE accuracy jumps on MiniGPT-4.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"HA-DPO trains multimodal models to prefer accurate image descriptions over hallucinatory ones by optimizing on paired responses.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6bcaca4640cf7cd7cb48662118152e2fd33ec01a53e36f970b9ae61704a55d89"},"source":{"id":"2311.16839","kind":"arxiv","version":2},"verdict":{"id":"0a33879e-fd3f-4015-b277-eaf4122f1c90","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T10:48:14.256416Z","strongest_claim":"When applied to three mainstream multimodal models, HA-DPO significantly reduced hallucination issues and amplified the models' generalization capabilities. Notably, the MiniGPT-4 model, when enhanced with HA-DPO, demonstrated a substantial improvement: POPE accuracy rose from 51.13% to 86.13% (an absolute improvement of 35%), and the MME score surged from 932.00 to 1326.46 (a relative improvement of 42.32%).","one_line_summary":"HA-DPO reframes hallucination reduction in LVLMs as direct preference optimization over style-consistent positive and negative response pairs, yielding large gains such as 35-point POPE accuracy jumps on MiniGPT-4.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The constructed positive and negative sample pairs are high-quality, style-consistent, and free of new biases that could undermine preference learning or generalization beyond the tested benchmarks.","pith_extraction_headline":"HA-DPO trains multimodal models to prefer accurate image descriptions over hallucinatory ones by optimizing on paired responses."},"references":{"count":79,"sample":[{"doi":"","year":2022,"title":"Brown, Jack Clark, Sam McCandlish, Chris Olah, Benjamin Mann, and Jared Kaplan","work_id":"42c49e87-0d71-4159-82b0-335c1cdf982a","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2014,"title":"The exploration-exploitation dilemma: a multidisci- plinary framework","work_id":"1b85bd8c-2331-4169-adae-4e756e673fca","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, Sand- hini Agarwal, Ariel Herbert-V oss, ","work_id":"36ef20ad-b8cb-4b11-9b0f-2e2d164b0174","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Christiano, Jan Leike, Tom B","work_id":"a2d819f8-2c2f-42a1-bd50-b96b93d59729","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instructblip: Towards general- purpose vision-language models wit","work_id":"ca0fe730-e9d7-436a-a761-ceba8bc2e1d3","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":79,"snapshot_sha256":"28bd0671d7c0b4edd136e93e111c48f18e82204dcaadae758ba18f98c8437de2","internal_anchors":1},"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"}