ModelRefs / MMLU-Pro Leaderboard — AI Model Scores
MMLU-Pro Leaderboard — AI Model Scores
Harder MMLU successor with 10-option questions and reduced contamination. Current leaders, methodology, and citation sources for MMLU-Pro.
Overview
Harder MMLU successor with 10-option questions and reduced contamination.
How it is measured: 5-shot CoT across 14 disciplines; 12k questions, 10 answer choices.
What this benchmark measures
- multi-discipline reasoning
- robust multiple-choice selection
Relevant to:
- hard reasoning evaluation portfolios
- broad discipline screening
Failure modes it exercises:
- reasoning errors under larger distractor sets
Method and its limits
Five-shot chain-of-thought evaluation across 14 disciplines with ten answer choices.
- Discipline-level performance can be obscured by aggregate accuracy.
- Chain-of-thought and harness settings can affect comparability.
Dataset
- Dataset
- MMLU-Pro
- Type
- academic multiple-choice questions with ten options
- Freshness
- aging
The registered methodology states over 12,000 questions across 14 disciplines and ten answer choices. The 2024-11-06 arXiv revision anchors the reviewed methodology, but it is not a current-world knowledge cutoff.
How to read this score
Accuracy supports matched-harness comparison, but should be decomposed by discipline and paired with real-task evaluation.
Similarity to real tasks: Low — Ten-option questions increase difficulty, but remain controlled academic multiple-choice tasks.
Data contamination risk: Medium — MMLU-Pro was designed to reduce trivial/noisy MMLU artifacts and is newer than MMLU, but the data and evaluation code are public; model-specific training overlap still requires run-level evidence.
Benchmark gaming risk: Medium — The paper reports lower prompt sensitivity than MMLU, but chain-of-thought use, prompt style, dataset revision, and benchmark-specific tuning still affect comparability.
What you still need to test yourself
- Test open-ended, current, domain-specific, and production-format tasks.
- Pin the dataset revision and evaluate prompt sensitivity, calibration, safety, cost, and latency separately.
This benchmark supports decisions about:
- Compare broad multi-discipline reasoning under a matched MMLU-Pro harness.
- Use discipline-level results to identify where deeper evaluation is needed.
Limitations
- Does not test deployment constraints, safety, privacy, tool use, or open-ended production performance.
- A higher aggregate score does not establish universal model superiority.
- Every score needs a canonical run record with exact model and dataset versions, prompt/harness settings, source scope, and limitations; no model-specific contamination conclusion is inferred.
Sources re-reviewed 2026-07-11. MMLU-Pro is treated as an aging 2024 benchmark with a reviewed v6 paper revision; each score must still identify the dataset and harness revision used.
Sources
- MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark MMLU-Pro authors / NeurIPS 2024 · accessed 2026-07-11
- MMLU-Pro code and data TIGER-AI-Lab · accessed 2026-07-11
- MMLU-Pro dataset card and maintenance history TIGER-Lab · accessed 2026-07-11
How this benchmark is scored
| Category | reasoning |
|---|---|
| Maximum score | 100 % accuracy |
| Direction | Higher is better |
Primary source: https://arxiv.org/abs/2406.01574
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to MMLU-Pro Leaderboard — AI Model Scores.