ChestX-Reasoner: Advancing Radiology Foundation Models with Reasoning through Step-by-Step Verification.

| 📑 Paper | 🐱 Github Repo | 🐱 ChestX-Reasoner-7B | 🐱 RadRBench |

Ziqing Fan^1,2 , Cheng Liang^1,2 , Chaoyi Wu^1,2 , Ya Zhang^1,2, Yanfeng Wang^1,2, Weidi Xie^1,2

¹ Shanghai Jiao Tong University, ² Shanghai AI Laboratory.

The official codes for "ChestX-Reasoner: Advancing Radiology Foundation Models with Reasoning through Step-by-Step Verification".

Training

In the following, we provide an overview and detailed guidance on the code used to train our ChestX-Reasoner and its variants.

• Note that SFT step requires at least 4 A100 80GB GPUs and training for about 2 days.
• Note that RL step requires at least 8 A100 80GB GPUs and training for about 3 days.
• VLLM for inference and Verl engine are essential to save training times.

Environment

You can install the code environment used for training our model. Our code is established based on VERL(https://github.com/volcengine/verl) engine. You may see for more detailed instructions. Besides, we provide a copy of our env list in ./env.txt.

conda create -n env_name python==3.10
conda activate env_name
pip3 install torch torchvision
pip3 install flash-attn --no-build-isolation
git clone https://github.com/volcengine/verl.git
cd verl
pip3 install -e .[vllm]

• Python: Version >= 3.9
• CUDA: Version >= 12.1
• VLLM: Version >= 0.7

Supervised Fine-Tuning

cd ChestXReasoner
bash run_SFT.sh

Notably, before run the bash file, there are configs and data paths should set in your devices. Please see details in ./ChestXReasoner/readme.md

Reinforcement Learning

To be continue

Reinforcement Learning with Process Reward

Evaluation

Benchmark Data

In eval/data, we present our benchmark construction code and our data.

Evaluation

We provide:

1. The evaluation code on both reasoning and accuracy in eval/
2. The baseline inference code in eval/inference
3. The evaluation results on both reasoning and accuracy of all baselines in eval/res

Citation

If you find this work is relevant with your research or applications, please feel free to cite our work!

@article{fan2025chestx,
  title={ChestX-Reasoner: Advancing Radiology Foundation Models with Reasoning through Step-by-Step Verification},
  author={Fan, Ziqing and Liang, Cheng and Wu, Chaoyi and Zhang, Ya and Wang, Yanfeng and Xie, Weidi},
  journal={arXiv preprint arXiv:2504.20930},
  year={2025}
}