Bud Ecosystem mirror of weidawang/CMPhysBench — a verbatim copy for offline, reproducible model evaluation. License unchanged (Apache-2.0); all rights remain with the original authors.

CMPhysBench: A Benchmark for Evaluating Large Language Models in Condensed Matter Physics

🎉🎉🎉 This paper is accpeted by ICLR 2026.

         

We introduce CMPhysBench, designed to assess the proficiency of Large Language Models (LLMs) in Condensed Matter Physics, as a novel Benchmark. CMPhysBench is composed of more than 520 graduate-level meticulously curated questions covering both representative subfields and foundational theoretical frameworks of condensed matter physics, such as magnetism, superconductivity, strongly correlated systems, etc. To ensure a deep understanding of the problem-solving process,we focus exclusively on calculation problems, requiring LLMs to independently generate comprehensive solutions. Meanwhile, leveraging tree-based representations of expressions, we introduce the Scalable Expression Edit Distance (SEED) score, which provides fine-grained (non-binary) partial credit and yields a more accurate assessment of similarity between prediction and ground-truth. Our results show that even the best models, Grok-4, reach only 36 average SEED score and 28% accuracy on CMPhysBench, underscoring a significant capability gap, especially for this practical and frontier domain relative to traditional physics.

Citations

@article{wang2025cmphysbench,
  title={CMPhysBench: A Benchmark for Evaluating Large Language Models in Condensed Matter Physics},
  author={Wang, Weida and Huang, Dongchen and Li, Jiatong and Yang, Tengchao and Zheng, Ziyang and Zhang, Di and Han, Dong and Chen, Benteng and Luo, Binzhao and Liu, Zhiyu and others},
  journal={arXiv preprint arXiv:2508.18124},
  year={2025}
}