Benchmark · multimodal

MMMU-Pro

11 results 8 models

MMMU-Pro is a robustness-hardened version of the MMMU benchmark that measures college-level multimodal (image + text) understanding and reasoning across six disciplines; the metric is multiple-choice accuracy.

Read more
Example
A discipline question pairs an image — a diagram, chart, chemical structure, or medical scan — with a college-level question and up to 10 answer options; in the vision-only setting the whole question and its options are embedded inside a screenshot or photo, so the model must read the text out of the image itself.
Scoring
Accuracy: the model's chosen option is exact-matched against the gold key and the score is the percentage of items answered correctly. Candidate options are augmented from 4 to 10 to blunt random guessing, and an overall MMMU-Pro score averages the Standard (10-option) and Vision-input settings (Chain-of-Thought prompting is typically used).
Verification
Every item is single-answer multiple choice, so grading is automatic: the final selected option is parsed from the model's (often chain-of-thought) output and compared for exact match with the reference answer. Robustness is baked into the data — questions that text-only models can answer are filtered out and the option count is raised to 10 — so a high score reflects genuine multimodal reasoning rather than text shortcuts or lucky guesses.
Why it matters
The original MMMU inflated scores because models could exploit text-only shortcuts or guess among four options; MMMU-Pro removes both, so measured accuracy drops sharply and better separates models that truly integrate vision and text, making it a more honest gauge of expert-level multimodal ability.
Worked example
Task
Representative Science (Physics) item in MMMU-Pro format. [Image: a block on a frictionless incline of angle θ = 30°.] Question: What is the magnitude of the block's acceleration along the incline? Options (10): (A) 0 m/s² (B) 1.6 m/s² (C) 2.5 m/s² (D) 3.3 m/s² (E) 4.9 m/s² (F) 5.7 m/s² (G) 6.9 m/s² (H) 7.4 m/s² (I) 8.5 m/s² (J) 9.8 m/s². In the vision-only setting this entire prompt appears embedded inside a screenshot.
Solution
a = g·sin θ = 9.8 × sin 30° = 9.8 × 0.5 = 4.9 m/s² → (E)
Walkthrough
On a frictionless incline the only force acting along the surface is the gravity component mg·sin θ, so the acceleration is g·sin θ and is independent of the block's mass; g·sin 30° = 4.9 m/s², which matches option (E). Grading compares the parsed option letter to the gold key by exact match.
0 22.5 45 67.5 90 2023-12-06 2025-04-04 2026-08-03 Gemini Ultra · 59.4 · 2023-12-06 Claude 3.5 Sonnet · 68.3 · 2024-06-20 o1 · 78.2 · 2024-09-12 Gemini 2.5 Pro Deep Think · 84 · 2025-05-20 Gemini 3 Pro · 81 · 2025-11-18 Gemini 3 Pro · 81 · 2025-11-18 Claude Opus 4.6 · 73.9 · 2026-02-05 Claude Opus 4.6 · 70.6 · 2026-02-05 Claude Opus 4.6 · 70.6 · 2026-03-06 Kimi K3 · 81.6 · 2026-07-17 Inkling-Small · 74 · 2026-08-03
Gemini Ultra Claude 3.5 Sonnet o1 Gemini 2.5 Pro Deep Think Gemini 3 Pro Claude Opus 4.6 Kimi K3 Inkling-Small
Timeline
Date Model Score Source
2026-08-03 Inkling-Small 74.0% Thinking Machines Lab releases Inkling-Small, a 276B open weights multimodal MoE model
2026-07-17 Kimi K3 81.6% Moonshot AI releases open Kimi K3, a 2.8T-parameter MoE model with 1M context
2026-03-06 Claude Opus 4.6 70.6% Anthropic releases Claude Opus 4.6 system card detailing capabilities and safety assessments
2026-02-05 Claude Opus 4.6 73.9% Anthropic releases Claude Opus 4.6 with 1M context window and agentic improvements
2026-02-05 Claude Opus 4.6 70.6% Anthropic publishes Claude Opus 4.6 system card detailing capabilities and safety evaluations
2025-11-18 Gemini 3 Pro 81.0% Google releases Gemini 3 Pro with state-of-the-art reasoning and multimodal capabilities
2025-11-18 Gemini 3 Pro 81.0% Google releases Gemini 3 Pro with state-of-the-art reasoning and multimodal capabilities
2025-05-20 Gemini 2.5 Pro Deep Think 84.0% Google DeepMind updates Gemini 2.5 with Deep Think, audio output, and computer use
2024-09-12 o1 78.2% OpenAI releases o1-preview, a reasoning model that surpasses human experts on math and science benchmarks
2024-06-20 Claude 3.5 Sonnet 68.3% Anthropic releases Claude 3.5 Sonnet addendum with improved coding and vision benchmarks
2023-12-06 Gemini Ultra 59.4% Google introduces Gemini 1.0, its largest multimodal AI model