https://github.com/sjtu-sai-agents/ML-Master

The official implementation of "ML-Master: Towards AI-for-AI via Integration of Exploration and Reasoning"
https://github.com/sjtu-sai-agents/ML-Master
Last synced: 4 days ago
JSON representation
The official implementation of "ML-Master: Towards AI-for-AI via Integration of Exploration and Reasoning"
Host: GitHub
URL: https://github.com/sjtu-sai-agents/ML-Master
Owner: sjtu-sai-agents
Created: 2025-06-16T08:01:16.000Z (8 months ago)
Default Branch: main
Last Pushed: 2026-01-16T11:14:47.000Z (24 days ago)
Last Synced: 2026-01-17T01:01:26.402Z (23 days ago)
Language: Python
Homepage:
Size: 47 MB
Stars: 333
Watchers: 12
Forks: 41
Open Issues: 3
Metadata Files:
- Readme: README.md
Awesome Lists containing this project

awesome-agentic-machine-learning - ML-Master - for-AI agent integrating exploration and reasoning with adaptive memory. By SJTU SAI. | ![GitHub stars](https://img.shields.io/github/stars/sjtu-sai-agents/ML-Master?style=flat-square) | (AutoML Agents)
README

          


![ML-Master Logo](./assets/logo.gif)



## 📰 What's New

- [2026/01/16] Release the preprint version of ML-Master 2.0! See the [ArXiv](https://arxiv.org/abs/2601.10402).

- [2025/12/16] 🎉 **ML-Master 2.0 reaches new heights!**  Achieving #1 on [MLE-Bench](https://github.com/openai/mle-bench) Leaderboard with 56.44% overall performance (92.7% improvement over 1.0). Thanks to [EigenAI](https://www.eigenai.com/) for their high-performance AI infrastructure support.

- [2025/10/30] We upload a new branch `feature-dev` with improved readability and maintainability. If you need to continue developing on ML-Master or apply ML-Master to downstream tasks, please switch the branch to `feature-dev`. 

- [2025/10/29] We now provide a Docker image for environment setup! Check it out [here](https://hub.docker.com/r/sjtuagents/ml-master).

- [2025/10/27] Add support for gpt-5.

- [2025/08/08] Initial code release is now available on GitHub!

- [2025/06/19] Release the preprint version! See the [ArXiv](https://arxiv.org/abs/2506.16499).

- [2025/06/17] Release the initial version! See the initial manuscript [here](./assets/ML-Master_github.pdf).

# ML-Master 2.0: Cognitive Accumulation for Ultra-Long-Horizon Agentic Science in Machine Learning Engineering

[![project](https://img.shields.io/badge/project-Page-blue)](https://sjtu-sai-agents.github.io/ML-Master)

[![arXiv](https://img.shields.io/badge/arXiv-2601.10402-b31b1b.svg)](https://arxiv.org/abs/2601.10402)

[![WeChat](https://img.shields.io/badge/WeChat-新智元-lightgreen)](https://mp.weixin.qq.com/s/dv1MD5S2vr3MB-skV4Thrw)

## 🚀 Overview

**ML-Master 2.0** is a pioneering agentic science framework that tackles the challenge of ultra-long-horizon autonomy through cognitive accumulation, facilitated by a Hierarchical Cognitive Caching (HCC) architecture that dynamically distills transient execution traces into stable long-term knowledge, ensuring that tactical execution and strategic planning remain decoupled yet co-evolve throughout complex, long-horizon scientific explorations. 

![ML-Master 2.0](./assets/ML-Master2.0-figure.png)

## 📊 Performance Highlights

![ML-Master 2.0 Score](./assets/ML-Master2.0_score.png)

**ML-Master 2.0** achieves **#1 on [MLE-Bench](https://github.com/openai/mle-bench) Leaderboard** with massive performance gains:

| Metric (%)                  | ML-Master 1.0 | ML-Master 2.0 | Relative Improvement |

|----------------------------|---------------|---------------|---------------------|

| 🥇 Overall (All)           | 29.33         | **56.44**     | **+92.7% ↑**       |

| 🟢 Low Complexity          | 48.48         | **75.76**     | **+56.2% ↑**       |

| 🟡 Medium Complexity       | 20.18         | **50.88**     | **+152.2% ↑**      |

| 🔴 High Complexity         | 24.44         | **42.22**     | **+72.8% ↑**       |

## 📆 Coming Soon

- [x] Grading report release

- [x] Paper release of ML-Master 2.0

- [ ] Initial code release of ML-Master 2.0

## 🙏 Acknowledgements

  

    

      

        

      

      


      SJTU SAI

    

    

      

        

        

      

      


      EigenAI

    

  

## ✍️ Citation

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

```bibtex

@misc{zhu2026ultralonghorizonagenticsciencecognitive,

      title={Toward Ultra-Long-Horizon Agentic Science: Cognitive Accumulation for Machine Learning Engineering}, 

      author={Xinyu Zhu and Yuzhu Cai and Zexi Liu and Bingyang Zheng and Cheng Wang and Rui Ye and Jiaao Chen and Hanrui Wang and Wei-Chen Wang and Yuzhi Zhang and Linfeng Zhang and Weinan E and Di Jin and Siheng Chen},

      year={2026},

      eprint={2601.10402},

      archivePrefix={arXiv},

      primaryClass={cs.AI},

      url={https://arxiv.org/abs/2601.10402}, 

}

```

```bibtex

@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}, 

}

```

---

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

[![project](https://img.shields.io/badge/project-Page-blue)](https://sjtu-sai-agents.github.io/ML-Master/1.0.html)

[![arXiv](https://img.shields.io/badge/arXiv-2506.16499-b31b1b.svg)](https://arxiv.org/abs/2506.16499)

[![WeChat](https://img.shields.io/badge/WeChat-新智元-lightgreen)](https://mp.weixin.qq.com/s/8Dn7Hvpmp59-0xDD28nQkw)

[![DockerHub](https://img.shields.io/badge/DockerHub-repository-blue.svg)](https://hub.docker.com/r/sjtuagents/ml-master)

> **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. 

![ML-Master](./assets/ML-Master_figure.png)

## 📊 Performance Highlights

ML-Master outperforms prior baselines on the **[MLE-Bench](https://github.com/openai/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 |

![ML-Master](./assets/ML-Master_score.png)

## 📆 Coming Soon

- [x] Grading report release

- [x] Paper release of ML-Master

- [x] Initial code release of ML-Master (expected early August)

- [x] Code refactoring for improved readability and maintainability

## 🚀 Quick Start

### 🛠️ Environment Setup

#### Pull and Start Docker Container

Please execute the following commands to pull the latest image and start an interactive container:

```bash

# Pull the latest image

docker pull sjtuagents/ml-master:latest

# Start the container

docker run --rm --gpus all --ipc=host --shm-size=64g \

  --runtime=nvidia --ulimit memlock=-1 --ulimit stack=67108864 \

  -it sjtuagents/ml-master:latest /bin/bash

# Clone the repository

git clone https://github.com/sjtu-sai-agents/ML-Master.git

cd ML-Master

conda activate ml-master

```

#### Install ml-master

To get started, make sure to first install the environment of **[MLE-Bench](https://github.com/openai/mle-bench)**. After that, install additional packages based on `requirements.txt`.

```bash

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](https://github.com/openai/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 `` tags in the response. Ensure your **DeepSeek** API supports this and follows the `OpenAI` client interface below:

```python

self.client = OpenAI(

    api_key=self.api_key,

    base_url=self.base_url

)

response = self.client.completions.create(**params)

```

If your API does not support this interface or you are using a closed source model(e.g. gpt-5) as coding model, please add `agent.steerable_reasoning=false` to `run.sh`. This may result in some performance loss.

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](https://github.com/openai/mle-bench)**.

```bash

# 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

bash launch_server.sh

```

After that, simply run the following command:

```bash

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](https://github.com/openai/mle-bench)**.

## 🙏 Acknowledgements

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

* 💡 **[MLE-Bench](https://github.com/openai/mle-bench)** — for providing a comprehensive and professional AutoML benchmarking platform.

* 🌲 **[AIDE](https://github.com/WecoAI/aideml)** — for offering a powerful tree-search-based AutoML code framework that inspired parts of our implementation.

## 💬 Contact Us

We welcome discussions, questions, and feedback! Join our WeChat group:







## ✍️ Citation

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

```bibtex

@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}, 

}

```
ecosyste.ms

Data

Tools

Indexes

Applications

Experiments

Awesome

https://github.com/sjtu-sai-agents/ML-Master

Awesome Lists containing this project

README
