MMMU
BenchmarkFreeExpert-level multimodal understanding across 30 subjects.
Capabilities6 decomposed
expert-level multimodal reasoning evaluation across 30 college subjects
Medium confidenceEvaluates AI models' ability to understand and reason over college-level academic content spanning 6 core disciplines (Art & Design, Business, Science, Health & Medicine, Humanities & Social Science, Tech & Engineering) using 11,500 multiple-choice questions that integrate visual perception (charts, diagrams, chemical structures, music sheets, maps, tables) with domain-specific knowledge and deliberate reasoning. Questions are manually curated from textbooks, lecture materials, and online academic sources by college students across multiple disciplines, requiring integration of visual and textual information to select correct answers from multiple choices.
MMMU combines breadth (30 college subjects across 6 disciplines, 183 subfields) with visual heterogeneity (30+ image types) and expert-level difficulty (college exam questions) in a single 11.5K-question benchmark. Unlike MMVP or other multimodal benchmarks that focus on general visual understanding, MMMU specifically targets domain knowledge integration with visual reasoning, requiring models to understand specialized visual representations (chemical structures, music notation, technical diagrams) alongside subject-specific knowledge. The manual curation by college students across disciplines ensures academic authenticity rather than synthetic or simplified visual-text pairs.
MMMU provides significantly broader subject coverage (30 subjects vs. 5-10 in competitors like MMVP or LLaVA-Bench) and more challenging expert-level questions (college exams vs. general visual QA), making it the most comprehensive multimodal reasoning benchmark for academic domains, though it lacks real-world validation and contamination mitigation that some competitors provide.
leaderboard-based model ranking with evalai submission infrastructure
Medium confidenceProvides an official leaderboard accessible at https://mmmu-benchmark.github.io that ranks AI models by accuracy on the held-out test set. Models are submitted via an EvalAI evaluation server (available since 2023-12-04) which automatically scores submissions against the test set, or alternatively evaluated locally using released test set answers (available since 2026-02-12). The leaderboard tracks performance across all 11,500 questions and enables comparison against baseline models (GPT-4V at 56% accuracy) and human expert performance (added 2024-01-31).
MMMU provides dual evaluation pathways: cloud-based EvalAI submission (enabling real-time leaderboard updates and public ranking) and local evaluation with released test answers (enabling offline analysis and reproducibility). This hybrid approach balances transparency (local evaluation prevents evaluation server lock-in) with competitive incentives (public leaderboard encourages participation). The EvalAI infrastructure automates scoring at scale, eliminating manual evaluation bottlenecks that plague other academic benchmarks.
MMMU's dual evaluation pathway (EvalAI + local) provides more flexibility than single-server benchmarks like GLUE or SuperGLUE, while the public leaderboard with human baseline enables competitive benchmarking that pure research datasets lack.
subject-specific and discipline-level performance decomposition
Medium confidenceEnables evaluation of model performance broken down across 30 college subjects organized into 6 core disciplines (Art & Design, Business, Science, Health & Medicine, Humanities & Social Science, Tech & Engineering) and 183 subfields. While the leaderboard provides aggregate accuracy, the benchmark structure allows researchers to analyze which subjects or disciplines their models struggle with, identifying domain-specific knowledge gaps. The 11,500 questions are distributed across these subjects, enabling fine-grained capability assessment beyond overall accuracy.
MMMU's 30-subject structure enables fine-grained domain analysis that most benchmarks lack. Unlike MMVP or LLaVA-Bench which provide only aggregate metrics, MMMU's explicit subject organization allows researchers to identify whether model weaknesses are general (low accuracy across all subjects) or domain-specific (e.g., poor chemistry knowledge but strong history understanding). The 6-discipline grouping provides intermediate-level analysis between aggregate and subject-level granularity.
MMMU's 30-subject decomposition provides 3-5x more granular domain analysis than competitors like MMVP (which lacks subject labels) or LLaVA-Bench (which uses only 5-10 categories), enabling precise diagnosis of domain-specific knowledge gaps.
heterogeneous visual modality evaluation across 30 image types
Medium confidenceEvaluates model performance on 30 distinct visual modality types including charts, diagrams, chemical structures, music sheets, maps, tables, photographs, and other specialized academic visualizations. The 11,500 questions are distributed across these 30 image types, enabling analysis of which visual representations models struggle with. This heterogeneity tests whether models have robust visual understanding across diverse modalities rather than overfitting to common image types (e.g., natural photographs).
MMMU explicitly categorizes 30 distinct visual modality types and distributes questions across them, enabling systematic evaluation of visual understanding robustness. Unlike benchmarks that assume all images are natural photographs or simple diagrams, MMMU includes specialized academic visualizations (chemical structures, music notation, circuit diagrams, anatomical illustrations) that require domain-specific visual parsing. This forces models to develop robust visual understanding beyond natural image recognition.
MMMU's 30-modality structure provides 5-10x more visual diversity than competitors like MMVP or LLaVA-Bench which focus primarily on natural images and simple diagrams, making it the most comprehensive test of visual understanding robustness across academic domains.
mmmu-pro robust variant for contamination mitigation
Medium confidenceA revised version of MMMU released 2024-09-05 designed to address robustness concerns in the original benchmark. While specific improvements are not documented in provided material, the existence of MMMU-Pro suggests the original benchmark had contamination, evaluation stability, or data quality issues that motivated a 'robust version.' This variant enables researchers to evaluate models on a potentially cleaner or more challenging version of the benchmark.
MMMU-Pro represents an iterative improvement on the original benchmark, suggesting the benchmark maintainers identified and addressed specific issues (likely contamination or evaluation stability). This demonstrates commitment to benchmark quality and provides researchers with a more reliable evaluation target than the original version.
MMMU-Pro's existence indicates the benchmark maintainers actively address quality issues, unlike static benchmarks that accumulate contamination over time; however, the lack of documentation on specific improvements limits its utility.
college-level academic question curation from diverse sources
Medium confidenceProvides 11,500 manually curated multiple-choice questions sourced from college textbooks, lecture materials, and online academic sources. Questions are collected by college students across multiple disciplines and cover 30 college subjects spanning 6 core disciplines and 183 subfields. This manual curation approach ensures questions reflect authentic academic difficulty and content rather than synthetic or simplified question generation, though it introduces potential quality variance and lacks documented inter-annotator agreement.
MMMU's manual curation by college students across disciplines ensures questions reflect authentic academic content and difficulty rather than synthetic generation. The sourcing from textbooks, lectures, and online materials grounds questions in real educational contexts. However, this approach trades scalability and quality control for authenticity — unlike synthetic benchmarks that can guarantee consistency, MMMU's manual curation introduces potential quality variance and contamination risks.
MMMU's authentic college-level questions provide more realistic evaluation than synthetic benchmarks like MMVP or LLaVA-Bench, but lack the quality control and decontamination procedures that some competitors implement.
Capabilities are decomposed by AI analysis. Each maps to specific user intents and improves with match feedback.
Related Artifactssharing capabilities
Artifacts that share capabilities with MMMU, ranked by overlap. Discovered automatically through the match graph.
MathVista
Visual mathematical reasoning benchmark.
MMLU (Massive Multitask Language Understanding)
57-subject benchmark, the standard metric for comparing LLMs.
MMLU
57-subject knowledge benchmark — 15K+ questions across STEM, humanities, professional domains.
chinese-llm-benchmark
ReLE评测：中文AI大模型能力评测（持续更新）：目前已囊括359个大模型，覆盖chatgpt、gpt-5.2、o4-mini、谷歌gemini-3-pro、Claude-4.6、文心ERNIE-X1.1、ERNIE-5.0、qwen3-max、qwen3.5-plus、百川、讯飞星火、商汤senseChat等商用模型， 以及step3.5-flash、kimi-k2.5、ernie4.5、MiniMax-M2.5、deepseek-v3.2、Qwen3.5、llama4、智谱GLM-5、GLM-4.7、LongCat、gemma3、mistral等开源大模型。不仅提供排行榜，也提供规模超20
LMSYS Chatbot Arena
Crowdsourced LLM evaluation — side-by-side blind voting, Elo ratings, most trusted LLM benchmark.
MATH Benchmark
12.5K competition math problems — AMC/AIME/Olympiad level, 7 subjects, standard math benchmark.
Best For
- ✓AI researchers developing multimodal large language models (LMMs)
- ✓Teams evaluating proprietary or open-source vision-language models for academic applications
- ✓Organizations building domain-specific AI assistants for education or professional training
- ✓Benchmark maintainers seeking comprehensive evaluation of multimodal reasoning capabilities
- ✓AI researchers publishing papers on multimodal models who need standardized benchmark results
- ✓Model developers seeking public validation and competitive positioning on a recognized benchmark
- ✓Teams evaluating multiple model variants and needing automated scoring infrastructure
- ✓Organizations tracking progress on multimodal reasoning over time
Known Limitations
- ⚠Only measures accuracy on multiple-choice questions — does not evaluate explanation quality, reasoning transparency, or multi-turn dialogue capabilities
- ⚠No real-world validation study correlating MMMU performance to actual expert performance or downstream task success
- ⚠Data contamination risk unaddressed — questions sourced from publicly available textbooks and online materials likely overlap with training data of web-scale models
- ⚠Exact train/dev/test split proportions not documented, making reproducibility and contamination assessment difficult
- ⚠No inter-annotator agreement metrics or quality control procedures documented for manual question curation
- ⚠Benchmark does not measure few-shot learning, adversarial robustness, or generalization to novel domains outside the 30 fixed subjects
Requirements
Input / Output
UnfragileRank
UnfragileRank is computed from adoption signals, documentation quality, ecosystem connectivity, match graph feedback, and freshness. No artifact can pay for a higher rank.
About
Massive Multi-discipline Multimodal Understanding benchmark with 11,500 expert-level questions across 30 subjects requiring college-level domain knowledge and deliberate reasoning over images, diagrams, and charts.
Categories
Alternatives to MMMU
Build high-quality LLM apps - from prototyping, testing to production deployment and monitoring.
Compare →Amplication brings order to the chaos of large-scale software development by creating Golden Paths for developers - streamlined workflows that drive consistency, enable high-quality code practices, simplify onboarding, and accelerate standardized delivery across teams.
Compare →Are you the builder of MMMU?
Claim this artifact to get a verified badge, access match analytics, see which intents users search for, and manage your listing.
Get the weekly brief
New tools, rising stars, and what's actually worth your time. No spam.
Data Sources
Looking for something else?
Search →