{"paper":{"title":"MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Multimodal models show 17 to 27 percent lower accuracy on MMMU-Pro than on MMMU.","cross_cats":["cs.CV"],"primary_cat":"cs.CL","authors_text":"Botao Yu, Ge Zhang, Graham Neubig, Huan Sun, Kai Zhang, Shengbang Tong, Tianyu Zheng, Wenhu Chen, Xiang Yue, Yuansheng Ni, Yubo Wang, Yu Su, Yuxuan Sun","submitted_at":"2024-09-04T15:31:26Z","abstract_excerpt":"This paper introduces MMMU-Pro, a robust version of the Massive Multi-discipline Multimodal Understanding and Reasoning (MMMU) benchmark. MMMU-Pro rigorously assesses multimodal models' true understanding and reasoning capabilities through a three-step process based on MMMU: (1) filtering out questions answerable by text-only models, (2) augmenting candidate options, and (3) introducing a vision-only input setting where questions are embedded within images. This setting challenges AI to truly \"see\" and \"read\" simultaneously, testing a fundamental human cognitive skill of seamlessly integrating"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Results show that model performance is substantially lower on MMMU-Pro than on MMMU, ranging from 16.8% to 26.9% across models.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that questions solvable by text-only models require no visual understanding and that embedding questions in images cleanly tests seamless visual-textual integration without introducing new confounds or biases.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MMMU-Pro is a stricter multimodal benchmark that removes text-only solvable questions, augments options, and requires reading text from images, yielding substantially lower model scores of 16.8-26.9%.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Multimodal models show 17 to 27 percent lower accuracy on MMMU-Pro than on MMMU.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a3977abe5b8b889b88d2a55b779339683307583cf28d62d0b016d5a6cb4fe83f"},"source":{"id":"2409.02813","kind":"arxiv","version":3},"verdict":{"id":"95cb3904-d6dd-435d-a5a5-e384df4cfa8d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T00:45:48.228603Z","strongest_claim":"Results show that model performance is substantially lower on MMMU-Pro than on MMMU, ranging from 16.8% to 26.9% across models.","one_line_summary":"MMMU-Pro is a stricter multimodal benchmark that removes text-only solvable questions, augments options, and requires reading text from images, yielding substantially lower model scores of 16.8-26.9%.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that questions solvable by text-only models require no visual understanding and that embedding questions in images cleanly tests seamless visual-textual integration without introducing new confounds or biases.","pith_extraction_headline":"Multimodal models show 17 to 27 percent lower accuracy on MMMU-Pro than on MMMU."},"references":{"count":73,"sample":[{"doi":"","year":2024,"title":"Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone","work_id":"feef9556-a016-493c-abd2-0c97a23a7ebf","ref_index":1,"cited_arxiv_id":"2404.14219","is_internal_anchor":true},{"doi":"","year":2022,"title":"Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. 2022. Flamingo: a visual language model","work_id":"6b329e9b-8d3c-474e-ac3e-7f0faf71a56e","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Anthropic. 2024. https://www.anthropic.com/news/claude-3-5-sonnet Claude 3.5 sonnet. https://www.anthropic.com/news/claude-3-5-sonnet","work_id":"96b9c6f9-1b0f-4ba0-8672-291abf39729a","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1109/iccv.2015.279","year":2015,"title":"Lawrence Zitnick, and Devi Parikh","work_id":"acf6495e-ae4c-4644-a52d-ff5e1c2ca351","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"OpenFlamingo: An Open-Source Framework for Training Large Autoregressive Vision-Language Models","work_id":"87bfa84a-e663-4165-806f-93ef439d88d0","ref_index":5,"cited_arxiv_id":"2308.01390","is_internal_anchor":true}],"resolved_work":73,"snapshot_sha256":"58820cbc906ff527e64ce8277deef03a21627afc9ebe631942072f2079f71631","internal_anchors":24},"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"}