X-Master: Can We Lead on Humanity’s Last Exam?

📑 ArXiv Paper

🔥 News

• [2025/08/22] Initial codes of X-Master is now available on GitHub!
• [2025/07/26] Play with our SciMaster, a general-purpose scientific AI agent product!

This is the official implementation of X-Master, a general-purpose tool-augmented reasoning agent.

Key Features

• 🧠 Interact with Environments during Reasoning: X-Master emulates human researchers by fluidly pivoting between internal reasoning and external tool use.

• 💻 Code as Interaction Language: X-Master communicates its intentions and interacts with environments—including Python libraries, custom tools, and even self-generated code—by formulating precise Python code snippets.

• 🔬 Scattered-and-Stacked Workflow: X-Masters enhances problem-solving performance by strategically increasing both the breadth of exploration and the depth of reasoning.

Examples

• Some response examples for each HLE category are in logs/example.jsonl.

🚀 QuickStart

Environment Setup

First install requirements using the following command.

conda create -n xmaster python=3.10
conda activate xmaster
pip install -r requirements.txt
cd src
pip install -e.

---

MCP Tools

The source code is available at code_server. You can clone the repository and deploy the code execution server using MCP Tools.

---

X-Master Configuration

1. Set DeepSeek-R1-0528 model url and ToolBox url in configs/common_config.py.

   Note that we use locally deployed DeepSeek-R1-0528 model, instead of api.

2. For Humanity's Last Exam (HLE) evaluation, set o3-mini api in configs/common_config.py.

---

Run X-Masters

Before running X-Masters, ensure that the environment, toolbox, and configuration are properly set up. We provide the text-only subset of HLE in data/hle.json.

• For single query inference with X-Masters, run

  python -m agents.XMaster.xmaster_agent --query "YOUR_QUERY"
  

• For X-Masters on HLE benchmark,
  • Generate solutions using X-Masters workflow.

    python -m functions.xmaster_hle
    

  • Evaluate generated solutions using o3-mini.

    python utils/hle_score.py
    

Citation

@article{xmaster,
  title={SciMaster: Towards General-Purpose Scientific AI Agents, Part I. X-Master as Foundation: Can We Lead on Humanity's Last Exam?},
  author={Jingyi, Chai and Tang, Shuo and Ye, Rui and Du, Yuwen and Zhu, Xinyu and Zhou, Mengcheng and Yanfeng, Wang and E, Weinan and Chen, Siheng},
  journal={arXiv preprint arXiv:2507.05241},
  year={2025}
}