SeePhys Pro: Diagnosing Modality Transfer and Blind-Training Effects in Multimodal RLVR for Physics Reasoning

PhysRL accompanies our SeePhys Pro paper, which studies whether multimodal models truly use visual evidence when solving physics reasoning problems. This repository provides the paper's physics RL training corpora in one place: the full PhysRL-38K collection and the vision-necessary PhysRL-8K subset used as the main physics RL training set.

Dataset Configs

• PhysRL-8K: 8,876 vision-necessary physics VQA examples.
• PhysRL-38K: 38,167 examples in the full PhysRL physics VQA training collection.

Both configs use the same public fields: id, problem, images, and answer. The PhysRL-8K identifiers are inherited from matching examples in PhysRL-38K, so shared samples have the same public id.

Project Family

These datasets are part of the same SeePhys Pro project:

• Benchmark: SeePhysPro
• RL training set: PhysRL

Links

• Paper: SeePhys Pro: Diagnosing Modality Transfer and Blind-Training Effects in Multimodal RLVR for Physics Reasoning
• Project page: seephyspro.github.io
• GitHub: AI4Phys/SeePhy-Pro
• Workshop challenge: AI for Math Workshop at ICML 2026
• Challenge submission: Codabench Track 3

Loading

from datasets import load_dataset

physrl8k = load_dataset("Kun-Xiang/PhysRL", "PhysRL-8K", split="train")
physrl38k = load_dataset("Kun-Xiang/PhysRL", "PhysRL-38K", split="train")

Citation

@article{xiang2026seephyspro,
  title={SeePhys Pro: Diagnosing Modality Transfer and Blind-Training Effects in Multimodal RLVR for Physics Reasoning},
  author={Xiang, Kun and Zhang, Terry Jingchen and Liu, Zirong and Zhou, Bokai and Tang, Yueling and Yu, Junjie and Lu, Jiacong and Huang, Shangrui and Li, Heng and Zhang, Likui and Liu, Kunkun and Zhang, Changzheng and Fang, Yangle and Guo, Boqiang and Zhen, Hui-Ling and Tu, Dandan and Huang, Yinya and Liang, Xiaodan},
  journal={arXiv preprint arXiv:2605.09266},
  year={2026}
}