Benchmark · general

MMLU-Pro

11 results 10 models

MMLU-Pro is a harder, reasoning-heavy successor to MMLU that tests a model's knowledge and problem-solving across many academic disciplines with multiple-choice questions, each offering 10 answer options. Results are reported as percentage accuracy — the share of questions answered correctly.

Read more
Example
A typical item is a multiple-choice question from a field like mathematics, physics, law, or engineering that needs several reasoning steps, where the model must choose the single correct answer among 10 options (versus 4 in the original MMLU).
Scoring
The metric is accuracy: the percentage of questions for which the model picks the correct option, computed as correct answers divided by total questions.
Verification
Each response is graded automatically by exact-matching the model's chosen option against the known correct label, so no human judging is involved.
Why it matters
Because leading models had nearly saturated the original MMLU, MMLU-Pro's tougher questions and 10 options make it more discriminating and more sensitive to real reasoning, giving a cleaner measure of progress.
Worked example
Task
MMLU-Pro (physics, 10 options). A projectile is launched from ground level at 20 m/s, 30° above the horizontal (g = 10 m/s², no air resistance). What is its maximum height? A) 2.5 m B) 5.0 m C) 7.5 m D) 10.0 m E) 12.5 m F) 15.0 m G) 17.3 m H) 20.0 m I) 3.5 m J) 8.7 m
Solution
v_y = v·sin30° = 20 × 0.5 = 10 m/s; h_max = v_y² / (2g) = 10² / (2 × 10) = 5.0 m → B
Walkthrough
At the apex the vertical velocity is zero, so h_max = v_y²/(2g) with v_y = v·sin30°; only the vertical component sets the height. MMLU-Pro grades by exact-match of the chosen letter against the gold key (accuracy over its 10 options).
0 24 48 72 96 2024-06-03 2024-12-30 2025-07-28 GPT-4o · 72.6 · 2024-06-03 GPT-4o · 72.6 · 2024-06-03 GPT-4-Turbo · 63.7 · 2024-06-03 Claude 3.5 Sonnet · 90.4 · 2024-06-20 Falcon3-3B-Instruct · 29.7 · 2024-12-17 Falcon3-7B-Base · 39.2 · 2024-12-17 Falcon3-10B-Base · 42.5 · 2024-12-17 DeepSeek-V3-0324 · 81.2 · 2025-03-24 Gemini 2.0 Flash · 77.6 · 2025-04-15 Deepseek-R1-0528 · 85 · 2025-05-30 GLM-4.5 · 84.6 · 2025-07-28
GPT-4o GPT-4-Turbo Claude 3.5 Sonnet Falcon3-3B-Instruct Falcon3-7B-Base Falcon3-10B-Base DeepSeek-V3-0324 Gemini 2.0 Flash Deepseek-R1-0528 GLM-4.5
Timeline
Date Model Score Source
2025-07-28 GLM-4.5 84.6% Zhipu AI releases GLM-4.5 and GLM-4.5-Air foundation models for intelligent agents
2025-05-30 Deepseek-R1-0528 85.0% DeepSeek updates R1 model with improved reasoning and benchmark scores
2025-04-15 Gemini 2.0 Flash 77.6% Google releases Gemini 2.0 Flash with enhanced quality and twice the speed of Gemini 1.5 Pro
2025-03-24 DeepSeek-V3-0324 81.2% DeepSeek-V3-0324 improves reasoning, coding, and Chinese writing over DeepSeek-V3
2024-12-17 Falcon3-3B-Instruct 29.7% Falcon3 family releases five open models with improved science, math, and code capabilities
2024-12-17 Falcon3-7B-Base 39.2% Falcon3 family releases five open models with improved science, math, and code capabilities
2024-12-17 Falcon3-10B-Base 42.5% Falcon3 family releases five open models with improved science, math, and code capabilities
2024-06-20 Claude 3.5 Sonnet 90.4% Anthropic releases Claude 3.5 Sonnet addendum with improved coding and vision benchmarks
2024-06-03 GPT-4o 72.6% MMLU-Pro introduces a more robust benchmark with 10 options and reasoning-focused questions
2024-06-03 GPT-4-Turbo 63.7% MMLU-Pro introduces a more robust benchmark with 10 options and reasoning-focused questions
2024-06-03 GPT-4o 72.6% MMLU-Pro paper