This is the source code for paper "DxDirector: an agentic large language model driving the full-process clinical diagnosis"

pip install -r requirements.txt
Typical install time: 30 minutes

Instruction-tuning for full-process clinical diagnosis:

1. make training data

python Instruction-Tuning-data-construction/make_data_stage1.py
python Instruction-Tuning-data-construction/make_data_stage2.py
python Instruction-Tuning-data-construction/make_data_stage3_add-think.py

2. Instruction-tuning

The training needs 4 Nvidia A100 80G GPUS.

sh Instruction-Tuning/inst-ft_code_with_marks_think_at_step.sh

Step-level strategy optimization:

1. make training data

python Step-level-strategy-optimization/data_construction/rewards_for_multiple_strategy.py

2. Training

The training needs 4 Nvidia A100 80G GPUS.

sh Step-level-strategy-optimization/train/slso_train.sh

Inference

Need one Nvidia A100 80G GPU.

sh Inference/DxDirector.sh
Expected run time: 2s for a sample.

Citation

@article{xu2026dxdirector,
  title={DxDirector: an agentic large language model driving the full-process clinical diagnosis},
  author={Xu, Shicheng and Huang, Xin and Wei, Zihao and Pang, Liang and Shen, Huawei and Cheng, Xueqi},
  journal={Nature Communications},
  year={2026},
  publisher={Nature Publishing Group}
}