MedREK: Retrieval-Based Editing for Medical LLMs with Key-Aware Prompts.

Source code for "MedREK: Retrieval-Based Editing for Medical LLMs with Key-Aware Prompts".

📰 News

• [2025/11/07] 🎉 MedREK has been accepted by ResponsibleFM @ NeurIPS 2025!
• [2025/10/16] 🚀 Our paper is available on arXiv!
• [2025/10/15] 🔥 We release the MedVersa dataset and code for MekREK!

👀 Introduction

• We present the first batch-editing benchmark in medical scenarios with broader coverage of medical subjects.

• We propose MedREK, a retrieval-based editing framework that integrates a shared query-key module for precise matching with an attention-based prompt encoder for informative guidance.

• Experimental results on various medical benchmarks demonstrate that our MedREK achieves superior performance across different core metrics and provides the first validated solution for batch-editing in medical LLMs.

🔧 Setup

conda create -n medrek python=3.10 -y
conda activate medrek
git clone https://github.com/mylittleriver/MedREK
pip install --upgrade medrek 
pip install -r requirements.txt

First enter the medrek directory: cd medrek

In utils/global_attrs.py, change ROOT_PATH and MODEL_PATH to your paths.

📊 MedVersa

Please check the /medversa floder to see the train\valid\test split of our enhanced Medical LLM Model Editing Benchmark.

⚙️ Train MedREK

Please run:

python train_medrek.py -mn 'meditron-7b' -dn 'medcf'  

Checkpoints will be saved in train_records/recipe/meditron-7b/train_name/checkpoints/.

You can view training information in train_records/recipe/meditron-7b/train_name/logs/ through Tensorboard.

We also provide the implementation of RECIPE for medical LLM model editing. Please check the RECIPE floder.

💫 Evaluate MedREK

Please run:

python test_medrek.py -en 'medrek' -mn 'meditron-7b' -et 'batch' -dvc 'cuda:0' \
    -ckpt 'train_records/recipe/meditron-7b/train_name/checkpoints/a_checkpoint' \
    -dn 'medcf' -edn 100 \

You can check results in eval_results/medrek.

🙏 Acknowledgement

This repo is built upon the following projects:

• RECIPE
• MedLaSA

We thank the authors for their codes.

📝 Citation

Please cite our work if you use our code or discuss our findings in your own research:

@article{xia2025medrek,
  title={MedREK: Retrieval-Based Editing for Medical LLMs with Key-Aware Prompts},
  author={Xia, Shujun and Lin, Haokun and Wu, Yichen and Zhou, Yinan and Li, Zixuan and Wan, Zhongwei and Xing, Xingrun and Zheng, Yefeng and Li, Xiang and Shan, Caifeng and others},
  journal={arXiv preprint arXiv:2510.13500},
  year={2025}
}