GUI-G1

The model checkpoint and training datasets are under review due to company policies and will be released soon. Thank you for your patience.

News

• [2025/9/16] Our model GUI-G1-3B-v1 in our paper is released.
• [2025/5/22] Our code is released.
• [2025/5/22] Our paper is released.

GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents

This repository is based on VLM-R1, with several improvements and adaptations for our use case, especially on Template, Reward Functions, and GRPO Objective.

In this work, we build upon the original VLM-R1 frameworks. We introduce GUI-G1, a VLM fine-tuned for GUI Grounding.

---

🔧 Major Modifications of GUI-G1

• Introduced a Fast Thinking Template that requires no model reasoning, accelerating training and inference
• Utilized diverse reward functions (Hit, IoU, Box) to prevent reward hacking and achieve multi-objective optimization
• Removed length correction from the GRPO objective and added a difficulty coefficient to enhance model robustness

---

Setup

conda create -n myproject python=3.10
conda activate myproject
bash setup.sh

---

Follow the steps below to prepare data and train the model:

1. [Data preparation instructions customized for your setup]
2. [Reference to your configuration files or modified scripts]
3. Use the following command to launch training:

bash src/open-r1-multimodal/run_scripts/run.sh

---

Results

Model	ScreenSpot	ScreenSPot-Pro
InfiGUI-R1-3B	87.5	35.7
GUI-G1-3B	90.3	37.1

---

Acknowledgements

This repository builds upon the great work from:

• VLM-R1

We thank the authors for their open-source contributions.

---

Citation

If you find our code or work useful for your research, please cite our work:

@article{zhou2025guig1,
  title        = {GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents},
  author       = {Zhou, Yuqi and Dai, Sunhao and Wang, Shuai and Zhou, Kaiwen and Jia, Qinglin and Xu, Jun},
  journal      = {arXiv preprint arXiv:2505.15810},
  year         = {2025}
}