VeriOS

Research code for the paper "VeriOS: Query-Driven Proactive Human-Agent-GUI Interaction for Trustworthy OS Agents".

Paper link: https://arxiv.org/abs/2509.07553

🚀 Quick Start

1. Environment Setup

1. Clone the repository:

   git clone https://github.com/Wuzheng02/VeriOS

2. Navigate into the project directory:

   cd VeriOS

3. Download the VeriOS-Bench dataset:

   https://huggingface.co/datasets/wuuuuuz/VeriOS-Bench

4. Download the pre-trained models:

   VeriOS-Agent-7B: https://huggingface.co/wuuuuuz/VeriOS-Agent-7B

   VeriOS-Agent-32B: https://huggingface.co/wuuuuuz/VeriOS-Agent-32B

2. Evaluation

1. Evaluate VeriOS-Agent performance:

   python test_interaction_loop.py --model_path /path/to/VeriOS-Agent --json_path /path/to/test.json

2. Evaluate dual-agent system performance:

   python dual_agent.py --model_path1 /path/to/scenarioagent --model_path2 /path/to/actionagent --json_path /path/to/test.json

3. Evaluate other baselines:

   python test_loop_{name}.py --model_path /path/to/agent --json_path /path/to/test.json

3. Training

This work is based on full fine-tuning of LLMs using LLaMA-Factory. We gratefully acknowledge the support from the LLaMA-Factory project.

To reproduce the training process of VeriOS-Agent from scratch:

1. Replace the .yaml files in the LLaMA-Factory repository with those provided in this repository.
2. Follow the official training tutorials provided in the LLaMA-Factory repository.

📋 Citation

@article{wu2025verios,
  title={VeriOS: Query-Driven Proactive Human-Agent-GUI Interaction for Trustworthy OS Agents},
  author={Zheng Wu and Heyuan Huang and Xingyu Lou and Xiangmou Qu and Pengzhou Cheng and Zongru Wu and Weiwen Liu and Weinan Zhang and Jun Wang and Zhaoxiang Wang and Zhuosheng Zhang},
  journal={arXiv preprint arXiv:2509.07553},
  year={2025}
}