🔗 Full paper: https://arxiv.org/abs/2505.20286

conda env create -f environment.yaml
conda activate eigen1
pip install -e .

# install tool dependencies
cd mcp_sandbox/
pip install -r requirements.txt

1. Configure the tools following the instructions in mcp_sandbox/README.md.

2. Configure configs/common_config.py file, set the following parameters:

• DEEPSEEK_CONFIG: Deepseek v3.1 model, which is used as the main agent
• OPENAI_CONFIG: for RAG
• O3-MINI_CONFIG: for evaluation
• SANDBOX: for MCP toolbox URL

To run the agent on the HLE Bio/Chem dataset(data/hle-bio.json):

1. Run the Agent

   python -m functions.eigen1_hle

2. Calculate Score Once the agent has finished, run the scoring script:

   python utils/hle_score.py

This repository benefit from X-Master.

@article{tang2025eigen,
  title={Eigen-1: Adaptive Multi-Agent Refinement with Monitor-Based RAG for Scientific Reasoning},
  author={Tang, Xiangru and Xu, Wanghan and Wang, Yujie and Guo, Zijie and Shao, Daniel and Chen, Jiapeng and Zhang, Cixuan and Wang, Ziyi and Zhang, Lixin and Wan, Guancheng and others},
  journal={arXiv preprint arXiv:2509.21193},
  year={2025}
}