MMODELYST
Evals/MMMU
EVL

MMMU

Multimodal understanding

Source ↗
benchmarkVision & MultimodalAccuracy across 30 college subjects (with images)11.5K items
What this measures

Multimodal college-level reasoning — questions that come with images, charts, diagrams, and chemical structures. Tests whether a model can reason over pictures, not just text.

Sample
Shown a circuit diagram: 'What is the voltage across resistor R2?' — requires reading the image and doing the physics.
Leaderboard · 8 models
Accuracy across 30 college subjects (with images) — higher is better
1Gemini 2.5 Pro
81.7
2GPT-4.1
74.8
3Claude 3.7 Sonnet
71.8
4Gemini 2.0 Flash
71.7
5GPT-4o
69.1
6Claude 3.5 Sonnet
68.3
7Gemini 1.5 Pro
62.2
8Claude 3 Opus
59.4
Benchmark health
active

Discriminating but climbing — the frontier is moving up this benchmark.

69.9
top-10 mean
81.7
best
8
models
What improves this score

Datasets & environments shown to raise it.

No data yet.
Research using this benchmark
all →
MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGICMMMU: A Chinese Massive Multi-discipline Multimodal Understanding BenchmarkAdversarial Orthogonal Disentanglement for LVLM Hallucination MitigationHEED: Density-Weighted Residual Alignment for Hybrid Vision-Language Model DistillationAre VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination

Matched by name in title/abstract.