<h1 align="center">
<img src="./assets/sparkui_logo.png" width="30" style="vertical-align: middle; margin-right: 3px;" />
M<sup>2</sup>-Miner: Multi-Agent Enhanced Monte Carlo Tree Search for Mobile GUI Agent Data Mining
</h1>

M2-Miner: Multi-Agent Enhanced Monte Carlo Tree Search for Mobile GUI Agent Data Mining

Overview

News
Motivation
Highlights
Experimental Results
Citation

🎉 News

2026-04-18: We released the inference code.
2026-02-05: We released our paper, M<sup>2</sup>-Miner: Multi-Agent Enhanced Monte Carlo Tree Search for Mobile GUI Agent Data Mining.

🚀 Motivation

Comparison of the different GUI data structure between prior methods.

Compared to existing methods, our algorithm produces a more diverse set of intents from the same initial intent.

✨ Highlights

🤖 Innovative MCTS-based GUI Agent Mining: We introduce M<sup>2</sup>-Miner, the first automated MCTS-driven framework for GUI agent data mining, featuring a collaborative multi-agent system—InferAgent, OrchestraAgent, and JudgeAgent—that jointly boost mining efficiency and data quality.
🔄 Intent Recycling Strategy: We propose a novel intent recycling mechanism that extracts a wealth of valuable interaction trajectories, significantly enriching intent diversity and further enhancing mining efficiency.
🧠 Progressive Model-in-the-Loop Training: Our progressive model-in-the-loop training strategy greatly improves the mining success rate and enables strong generalization to unseen environments, supporting robust startup even in novel scenarios.
👑 Superior Mining Quality and SOTA Agent Performance: Extensive experiments confirm that our method excels in mining high-quality, diverse intent trajectories, and GUI agents trained with our data consistently achieve state-of-the-art results on standard benchmarks.

📊 Experimental Results

Table 1. Performance comparison on GUI agent benchmarks. Bold and underline indicate the best and second-best results.

🤝 Acknowledgements

We thank the teams behind Qwen2.5-VL and AndroidControl for their foundational work, and ms-swift for the efficient training and inference framework. We also thank AgentCPM-GUI for providing evaluation code.

📄 License

Code in this repository is released under the Apache-2.0 license.

📄 Citation

If you use works related to M<sup>2</sup>-Miner, please cite our work:

@misc{lv2026m2miner,
      title={M2-Miner: Multi-Agent Enhanced MCTS for Mobile GUI Agent Data Mining}, 
      author={Rui Lv, Juncheng Mo, Tianyi Chu, Chen Rao, Hongyi Jing, Jiajie Teng, Jiafu Chen, Shiqi Zhang, Liangzi Ding, Shuo Fang, Huaizhong Lin, Ziqiang Dang, Chenguang Ma, Lei Zhao},
      journal={arXiv preprint arXiv:2602.05429},
      year={2026},
      doi={10.48550/arXiv.2602.05429},
      url={https://doi.org/10.48550/arXiv.2602.05429},
      note={Accepted by ICLR 2026}
}

If you are interested in our method or it helps your research, please give us a star 🌟 on GitHub.

README