ML-Master: Towards AI-for-AI via Intergration of Exploration and Reasoning

Status: ⌛ Initial code release is now available!

🚀 Overview

ML-Master is a novel AI4AI (AI-for-AI) agent that integrates exploration and reasoning into a coherent iterative methodology, facilitated by an adaptive memory mechanism that selectively captures and summarizes relevant insights and outcomes, ensuring each component mutually reinforces the other without compromising either.

📰 What's New

• [2025/08/08] Initial code release is now available on GitHub!
• [2025/06/19] Release the preprint version! See the ArXiv.
• [2025/06/17] Release the initial version! See the initial manuscript here.

📊 Performance Highlights

ML-Master outperforms prior baselines on the MLE-Bench:

Metric	Result
🥇 Average Medal Rate	29.3%
🧠 Medium Task Medal Rate	20.2%, more than doubling the previous SOTA
🕒 Runtime Efficiency	12 hours, 50% budget

📆 Coming Soon

• Grading report release
• Paper release of ML-Master
• Initial code release of ML-Master (expected early August)
• Code refactoring for improved readability and maintainability

🚀 Quick Start

🛠️ Environment Setup

To get started, make sure to first install the environment of MLE-Bench. After that, install additional packages based on requirements.txt.

git clone https://github.com/sjtu-sai-agents/ML-Master.git
cd ML-Master
conda create -n ml-master python=3.12
conda activate ml-master

# 🔧 Install MLE-Bench environment here
# (Follow the instructions in its README)

pip install -r requirements.txt

---

📦 Download MLE-Bench Data

The full MLE-Bench dataset is over 2TB. We recommend downloading and preparing the dataset using the scripts and instructions provided by MLE-Bench.

Once prepared, the expected dataset structure looks like this:

/path/to/mle-bench/plant-pathology-2020-fgvc7/
└── prepared
    ├── private
    │   └── test.csv
    └── public
        ├── description.md
        ├── images/
        ├── sample_submission.csv
        ├── test.csv
        └── train.csv

🪄 ML-Master uses symbolic links to access the dataset. You can download the data to your preferred location and ML-Master will link it accordingly.

---

🧠 Configure DeepSeek and GPT

ML-Master requires LLMs to return custom <think></think> tags in the response. Ensure your DeepSeek API supports this and follows the OpenAI client interface below:

self.client = OpenAI(
    api_key=self.api_key,
    base_url=self.base_url
)
response = self.client.completions.create(**params)

Set your base_url and api_key in the run.sh script. GPT-4o is used only for evaluation and feedback, consistent with MLE-Bench.

# Basic configuration
AGENT_DIR=./
EXP_ID=plant-pathology-2020-fgvc7   # Competition name
dataset_dir=/path/to/mle-bench      # Path to prepared dataset
MEMORY_INDEX=0                      # GPU device ID

# DeepSeek config
code_model=deepseek-r1
code_temp=0.5
code_base_url="your_base_url"
code_api_key="your_api_key"

# GPT config (used for feedback & metrics)
feedback_model=gpt-4o-2024-08-06
feedback_temp=0.5
feedback_base_url="your_base_url"
feedback_api_key="your_api_key"

# CPU allocation
start_cpu=0
CPUS_PER_TASK=36
end_cpu=$((start_cpu + CPUS_PER_TASK - 1))

# Time limit (in seconds)
TIME_LIMIT_SECS=43200

---

▶️ Start Running

Before running ML-Master, you need to launch a server which tells agent whether the submission is valid or not, allowed and used by MLE-Bench.

bash launch_server.sh

After that, simply run the following command:

bash run.sh

📝 Logs and solutions will be saved in:

• ./logs (for logs)
• ./workspaces (for generated solutions)

---

📊 Evaluation

For evaluation details, please refer to the official MLE-Bench evaluation guide.

🙏 Acknowledgements

We would like to express our sincere thanks to the following open-source projects that made this work possible:

• 💡 MLE-Bench — for providing a comprehensive and professional AutoML benchmarking platform.
• 🌲 AIDE — for offering a powerful tree-search-based AutoML code framework that inspired parts of our implementation.

✍️ Citation

If you find our work helpful, please use the following citations.

@misc{liu2025mlmasteraiforaiintegrationexploration,
      title={ML-Master: Towards AI-for-AI via Integration of Exploration and Reasoning}, 
      author={Zexi Liu and Yuzhu Cai and Xinyu Zhu and Yujie Zheng and Runkun Chen and Ying Wen and Yanfeng Wang and Weinan E and Siheng Chen},
      year={2025},
      eprint={2506.16499},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2506.16499}, 
}