MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark
Open accessAbstract
In the age of large-scale language models, benchmarks like the Massive Multitask Language Understanding (MMLU) have been pivotal in pushing the boundaries of what AI can achieve in language comprehension and reasoning across diverse domains. However, as models continue to improve, their performance on these benchmarks has begun to plateau, making it increasingly difficult to discern differences in model capabilities. This paper introduces MMLU-Pro, an enhanced dataset designed to extend the mostly knowledge-driven MMLU benchmark by integrating more challenging, reasoning-focused questions and expanding the choice set from four to ten options. Additionally, MMLU-Pro eliminates the trivial and noisy questions in MMLU. Our experimental results show that MMLU-Pro not only raises the challenge, causing a significant drop in accuracy by 16% to 33% compared to MMLU but also demonstrates greater stability under varying prompts. 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. Additionally, we found that models utilizing Chain of Thought (CoT) reasoning achieved better performance on MMLU-Pro compared to direct answering, which is in stark contrast to the findings on the original MMLU, indicating that MMLU-Pro includes more complex reasoning questions. Our assessments confirm that MMLU-Pro is a more discriminative benchmark to better track progress in the field.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
When a benchmark's accuracy saturates, the prevailing response is to retire it and replace it with a more challenging version.
“When a benchmark's accuracy saturates, it is often retired and replaced with a more challenging version.”
From Life After Benchmark Saturation: A Case Study of CORE-Bench (2026) · cites this paper as [53]
Interventions to improve evaluation have been well explored, such as constructing higher quality benchmarks by identifying errors, expanding test sets, selecting high quality instances, or generating new synthetic benchmarks.
“Interventions to improve evaluation have been well explored, such as constructing higher quality benchmarks by identifying errors [62, 21], expanding test sets [64], selecting high quality instances from benchmarks [45], or generating entirely new synthetic benchmarks from a model [32].”
From Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation (2025) · cites this paper as [64]
Cited by
Sorted by citations
DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning
Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., et al. (2025) · Nature · doi:10.1038/s41586-025-09422-z · Cited by 1,701 · see citing works
Self-citationLarge language models for automated scholarly paper review: A survey
Zhuang, Z., Chen, J., Xu, H., Jiang, Y., & Lin, J. (2025) · Information Fusion · doi:10.1016/j.inffus.2025.103332 · Cited by 34 · see citing works
A benchmark of expert-level academic questions to assess AI capabilities
Center for AI Safety, Phan, L., Gatti, A., Li, N., Khoja, A., Kim, R., Ren, R., Hausenloy, J., Zhang, O., Mazeika, M., et al. (2026) · Nature · doi:10.1038/s41586-025-09962-4 · Cited by 23 · see citing works
Self-citationVoting or Consensus? Decision-Making in Multi-Agent Debate
Kaesberg, L. B., Becker, J., Wahle, J. P., Ruas, T., & Gipp, B. (2025) · Findings of the Association for Computational Linguistics: ACL 2025 · doi:10.18653/v1/2025.findings-acl.606 · Cited by 14 · see citing works
Absolute Zero: Reinforced Self-play Reasoning with Zero Data
Zhao, A., Wu, Y., Yue, Y., Wu, T., Xu, Q., Yue, Y., Lin, M., Wang, S., Wu, Q., Zheng, Z., et al. (2025) · arXiv · Cited by 13 · see citing works
Liu, R., Geng, J., Wu, A. J., Sucholutsky, I., Lombrozo, T., & Griffiths, T. L. (2024) · arXiv · doi:10.48550/arxiv.2410.21333 · Cited by 6 · see citing works
When AI Co-Scientists Fail: SPOT-a Benchmark for Automated Verification of Scientific Research
Son, G., Hong, J., Fan, H., Nam, H., Ko, H., Lim, S., Song, J., Choi, J., Paulo, G., Yu, Y., et al. (2025) · arXiv · doi:10.48550/arxiv.2505.11855 · Cited by 5 · see citing works
MiMo: Unlocking the Reasoning Potential of Language Model - From Pretraining to Posttraining
Xiaomi, L.-C., :, Xia, B., Shen, B., Cici, Zhu, D., Zhang, D., Wang, G., Zhang, H., Liu, H., et al. (2025) · arXiv · doi:10.48550/arxiv.2505.07608 · Cited by 1 · see citing works
Self-citationTowards a Science of Scaling Agent Systems
Kim, Y., Gu, K., Park, C., Park, C., Schmidgall, S., Heydari, A. A., Yan, Y., Zhang, Z., Zhuang, Y., Liu, Y., et al. (2025) · arXiv · doi:10.48550/arxiv.2512.08296 · Cited by 1 · see citing works
A-MemGuard: A Proactive Defense Framework for LLM-Based Agent Memory
Wei, Q., Yang, T., Wang, Y., Li, X., Li, L., Yin, Z., Zhan, Y., Holz, T., Lin, Z., & Wang, X. (2025) · arXiv · doi:10.48550/arxiv.2510.02373 · Cited by 1 · see citing works
Self-citationLife After Benchmark Saturation: A Case Study of CORE-Bench
Nadgir, N., Kapoor, S., Liu, K., Kirgis, P., Orona, M., Rabanser, S., Bayer, T., Shetty, A., Ling, Y., Chan-Sew, D., et al. (2026) · arXiv · doi:10.48550/arxiv.2606.26158
AlphaResearch: Accelerating New Algorithm Discovery with Language Models
Yu, Z., Feng, K., Zhao, Y., He, S., Zhang, X.-P., & Cohan, A. (2025) · arXiv · doi:10.48550/arxiv.2511.08522
From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs
Wang, H., Liang, W., Fu, Z., Zheng, N., Zhang, Y., Tong, Y., Zhu, T., Jiang, H., Li, C., Wu, J., et al. (2025) · arXiv · doi:10.48550/arxiv.2509.23196
Protecting Language Models Against Unauthorized Distillation through Trace Rewriting
Ma, X., Yeoh, W., Zhang, N., & Vorobeychik, Y. (2026) · arXiv · doi:10.48550/arxiv.2602.15143
Self-citationDynamic Cheatsheet: Test-Time Learning with Adaptive Memory
Suzgun, M., Yuksekgonul, M., Bianchi, F., Jurafsky, D., & Zou, J. (2025) · arXiv · doi:10.48550/arxiv.2504.07952
Nanbeige4.2-3B: Unlocking Agentic Capabilities in a Compact Model
Lab, N., :, Yang, C., Huang, C., Lan, F., Chen, H., Zhou, H., Song, H., Cao, J., Zhu, J., et al. (2026) · arXiv · doi:10.48550/arxiv.2607.22083
Self-citationMeasuring Collapse and Correction in Homogeneous-Panel LLM Debate
Li, X., Liu, M., & Yuen, C. (2026) · arXiv · doi:10.48550/arxiv.2609.35279
Stop Overvaluing Multi-Agent Debate -- We Must Rethink Evaluation and Embrace Model Heterogeneity
Zhang, H., Cui, Z., Chen, J., Wang, X., Zhang, Q., Wang, Z., Wu, D., & Hu, S. (2025) · arXiv · doi:10.48550/arxiv.2502.08788
A Survey of Agent Memory in the Second Half: Towards Self-Evolving and Long-Horizon Agents
Huang, W.-C., Zhang, W., Liang, Y., Bei, Y., Chen, Y., Feng, T., Pan, X., Tan, Z., Wang, Y., Wei, T., et al. (2026) · arXiv · doi:10.48550/arxiv.2602.06052
Self-citationSignal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation
Heineman, D., Hofmann, V., Magnusson, I., Gu, Y., Smith, N. A., Hajishirzi, H., Lo, K., & Dodge, J. (2025) · arXiv · doi:10.48550/arxiv.2508.13144
Critique-GRPO: Advancing LLM Reasoning with Natural Language and Numerical Feedback
Zhang, X., Zhang, Y., Sun, H., Feng, K., Lu, C., Yang, C., & Meng, H. (2025) · arXiv · doi:10.48550/arxiv.2506.03106
LLMs Cannot Reliably Judge (Yet?): A Comprehensive Assessment on the Robustness of LLM-as-a-Judge
Li, S., Xu, C., Wang, J., Gong, X., Chen, C., Zhang, J., Wang, J., Lam, K.-Y., & Ji, S. (2025) · arXiv · doi:10.48550/arxiv.2506.09443
- Authors
- Wang, Y., Ma, X., Zhang, G., Ni, Y., Chandra, A., Guo, S., Ren, W., Arulraj, A., He, X., Jiang, Z., Li, T., Ku, M., Wang, K., Zhuang, A., Fan, R., Yue, X., & Chen, W.
- Venue
- Advances in Neural Information Processing Systems 37
- Published
- 2024
- Pages
- 95266-95290
- ISSN
- 1049-5258 (ISSN-L)
- OpenAlex
- W4415796274
- Cited by
- 131
- References
- 0
- Citations by year
- 2026: 87 · 2025: 39 · 2024: 5
- FWCI
- 37.82
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- cc-by
- Locations
- neural information processing systems (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Funders
- DeepMind
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
- Keywords
- benchmark, MMLU-Pro, language model evaluation, robustness, challenging tasks
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
