{"paper":{"title":"Aligning Modalities in Vision Large Language Models via Preference Fine-tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Preference fine-tuning on automatically generated hallucinated responses aligns vision and language modalities in large models while cutting hallucinations.","cross_cats":["cs.CL","cs.CV"],"primary_cat":"cs.LG","authors_text":"Chelsea Finn, Chenhang Cui, Huaxiu Yao, Rafael Rafailov, Yiyang Zhou","submitted_at":"2024-02-18T00:56:16Z","abstract_excerpt":"Instruction-following Vision Large Language Models (VLLMs) have achieved significant progress recently on a variety of tasks. These approaches merge strong pre-trained vision models and large language models (LLMs). Since these components are trained separately, the learned representations need to be aligned with joint training on additional image-language pairs. This procedure is not perfect and can cause the model to hallucinate - provide answers that do not accurately reflect the image, even when the core LLM is highly factual and the vision backbone has sufficiently complete representation"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"In experiments across broad benchmarks, we show that we can not only reduce hallucinations, but improve model performance across standard benchmarks, outperforming prior approaches.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The two-stage automated generation of dispreferred responses (GPT-4V hallucination injection and image distortion) produces high-quality preference pairs that accurately reflect and correct the model's hallucination behavior when used in DPO.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"POVID generates AI-created preference data to fine-tune vision-language models with DPO, reducing hallucinations and improving benchmark scores.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Preference fine-tuning on automatically generated hallucinated responses aligns vision and language modalities in large models while cutting hallucinations.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"fc41d789a43373108e2b1e47bfe0168f29991977cfe7ec321abbbf203a40db8a"},"source":{"id":"2402.11411","kind":"arxiv","version":1},"verdict":{"id":"c94e6e5a-3650-40e1-8bb0-efed02110008","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T10:54:31.659670Z","strongest_claim":"In experiments across broad benchmarks, we show that we can not only reduce hallucinations, but improve model performance across standard benchmarks, outperforming prior approaches.","one_line_summary":"POVID generates AI-created preference data to fine-tune vision-language models with DPO, reducing hallucinations and improving benchmark scores.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The two-stage automated generation of dispreferred responses (GPT-4V hallucination injection and image distortion) produces high-quality preference pairs that accurately reflect and correct the model's hallucination behavior when used in DPO.","pith_extraction_headline":"Preference fine-tuning on automatically generated hallucinated responses aligns vision and language modalities in large models while cutting hallucinations."},"references":{"count":155,"sample":[{"doi":"","year":2023,"title":"InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning , author=. 2023 , eprint=","work_id":"2a233147-17ae-4a84-a51a-aeaa6303e8e1","ref_index":18,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2000,"title":"Langley , title =","work_id":"6cd283dc-0548-45e9-af07-6bc1005593ad","ref_index":20,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1980,"title":"T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980","work_id":"6b09bca6-ef5d-4a83-8c8e-219f23cbd761","ref_index":21,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"M. J. Kearns , title =","work_id":"8efd8073-6f5d-45c5-94d5-62d366b52518","ref_index":22,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1983,"title":"Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983","work_id":"51835800-f16e-4534-8339-d3ea09147556","ref_index":23,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":155,"snapshot_sha256":"5b1f65218dc218aa0a2bbc2f22c81feb7663e2379061bb550038898361d7e30c","internal_anchors":22},"formal_canon":{"evidence_count":2,"snapshot_sha256":"41ded6d331100e664479d2b235b3d8c2d8b5b1aa5b27fbfa16f9871c6754a422"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}