MMLU-ProX: A Multilingual Benchmark for Advanced Large Language Model Evaluation

^*The University of Tokyo, Japan, ^†Duke-NUS Medical School, Singapore, ^‡Waseda University, Japan,
^§Northwestern University, United States, ^¶Carnegie Mellon University, United States,
^#Yale University, United States, ^♯University College Dublin, Ireland, ^♮Nanyang Technological University, Singapore,
^♭Smartor Inc, Japan, ^⊕University of California, Berkeley, United States, ^⊗University of New South Wales, Australia,
^ΔSingapore Management University, Singapore, ^ΛNew York University, United States,
^ΩPolytechnique Montreal, Canada, ^ΨUniversity of Geneva, Switzerland, ^ΓUniversity of Alberta, Canada

MMLU-ProX is a multilingual benchmark that builds upon MMLU-Pro, extending to 29 typologically diverse languages, designed to evaluate large language models' reasoning capabilities across linguistic and cultural boundaries.

MMLU-ProX addresses critical limitations in existing multilingual benchmarks by:

• Extending coverage to 29 typologically diverse languages
• Building upon the challenging, reasoning-focused design of MMLU-Pro
• Employing a rigorous semi-automatic translation process with expert validation
• Ensuring conceptual accuracy, terminological consistency, and cultural relevance

• [2025/08] 🎉 MMLU-ProX was accepted by EMNLP 2025 (Main)! We are working on more languages for the second version. Stay tuned!
• [2025/05] MMLU-ProX now contains 29 languages, all available on Huggingface! We provide both a lite version and a full version.
• [2025/03] MMLU-ProX's evaluation is now available on lm-evaluation-harness!
• [2025/03] MMLU-ProX is now available on Hugging Face!
• [2025/03] We are still expanding this dataset to more languages! Stay tuned!

To reproduce the results posted in our paper, we support vLLM evaluation by lm-evaluation-harness (Here) by the following command:

model_id=<your-target-model>
tensor_parallel_size=<number-of-gpu-you-want-to-use>
lang=<your-target-language>

python -m lm_eval \
  --model vllm \
  --model_args pretrained=${model_id},tensor_parallel_size=${tensor_parallel_size},dtype=auto,gpu_memory_utilization=0.9 \
  --batch_size auto \
  --tasks mmlu_prox_${lang}

Please refer to lm-evaluation-harness for more details about how to setup.

Note: Please install vllm=0.7.3 to reproduce our results other than Llama3.1-405B which is evaluated by vllm=0.6.6.

@article{xuan2025mmluprox,
  title={Mmlu-prox: A multilingual benchmark for advanced large language model evaluation},
  author={Xuan, Weihao and Yang, Rui and Qi, Heli and Zeng, Qingcheng and Xiao, Yunze and Feng, Aosong and Liu, Dairui and Xing, Yun and Wang, Junjue and Gao, Fan and others},
  journal={arXiv preprint arXiv:2503.10497},
  year={2025}
}

For questions or feedback about MMLU-ProX, please open an issue.