MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark
Summary
This paper introduces MMLU-Pro, an enhanced dataset designed to extend the mostly knowledge-driven MMLU benchmark. It adds more reasoning-focused questions, expands the choice set from four to ten options, and drops trivial or noisy MMLU items. Experimental results show a significant drop in accuracy by 16% to 33% compared to MMLU. With 24 different prompt styles tested, the sensitivity of model scores to prompt variations decreased from 4-5% in MMLU to just 2% in MMLU-Pro. Models utilizing Chain of Thought (CoT) reasoning achieved better performance on MMLU-Pro compared to direct answering, in contrast to findings on the original MMLU, indicating more complex reasoning questions. Please visit our leaderboard at https://huggingface.co/spaces/TIGER-Lab/MMLU-Pro. The v6 paper is the NeurIPS 2024 Datasets and Benchmarks version. [S1]
Claims
- id: C1 claim: "MMLU-Pro expands each question from four answer choices to ten and adds harder, reasoning-focused items." source: S1 evidence: "expanding the choice set from four to ten options"
- id: C2 claim: "Accuracy drops by 16% to 33% compared with MMLU, and prompt sensitivity falls to about 2%." source: S1 evidence: "significant drop in accuracy by 16% to 33% compared to MMLU"
- id: C3 claim: "The authors point to a public Hugging Face leaderboard for the benchmark." source: S1 evidence: "https://huggingface.co/spaces/TIGER-Lab/MMLU-Pro"
Verification
- Code available: Paper: https://huggingface.co/spaces/TIGER-Lab/MMLU-Pro, which returned HTTP 200 on 2026-11-12. Not run.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: The original benchmark is covered in 2009-03300-mmlu. Chain-of-thought prompting is in 2201-11903-chain-of-thought-prompting.
- Limits: Results depend on the model list and prompt styles in the v6 paper.
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 3 Jun 2024 (v1); last revised v6 (6 Nov 2024). Sources: [S1] https://arxiv.org/abs/2406.01574, v6 (6 Nov 2024). checked_on: 2026-11-12.
When to apply
Use when comparing models on a harder, less prompt-fragile successor to MMLU, or when asking why CoT helps more here than on the original set.
Facts
- id: F1 claim: "MMLU-Pro extends MMLU with harder reasoning questions and ten-option multiple choice." confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F2 claim: "The authors report a 16% to 33% accuracy drop versus MMLU and lower sensitivity to prompt wording." confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F3 claim: "A leaderboard is published at https://huggingface.co/spaces/TIGER-Lab/MMLU-Pro." confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
Sources
- [S1] arXiv 2406.01574, https://arxiv.org/abs/2406.01574. License: CC BY 4.0. v6 (6 Nov 2024). Accessed 2026-11-12.
License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.
License: CC-BY-4.0 · Summary by iinventedlight (AI-assisted). Source: Yubo Wang, Xueguang Ma, Ge Zhang, and 14 coauthors, MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark (CC BY 4.0).
Cite this
'iinventedlight (AI-assisted)'. 2026. "MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark." iinventedlight Library, iil:note:tzxdexvz, version of 2026-11-12. https://iinventedlight.com/library/notes/2406-01574-mmlu-pro/. License: CC-BY-4.0.
@misc{iil_note_tzxdexvz,
author = {{'iinventedlight (AI-assisted)'}},
title = {{MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark}},
howpublished = {iinventedlight Library},
year = {2026},
date = {2026-11-12},
url = {https://iinventedlight.com/library/notes/2406-01574-mmlu-pro/},
note = {Permanent ID iil:note:tzxdexvz. License: CC-BY-4.0.},
}
Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.
For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2406-01574-mmlu-pro.md.