GEM: Gaussian Embedding Modeling for Out-of-Distribution Detection in GUI Agents

Research code for the paper "GEM: Gaussian Embedding Modeling for Out-of-Distribution Detection in GUI Agents".

Paper link: https://arxiv.org/abs/2505.12842

🚀 Quick Start

1. Clone the Repository

git clone https://github.com/Wuzheng02/GEM-OODforGUIagents
cd GEM-OODforGUIagents

2. Run Evaluation

Example: AITZ (ID) vs. OmniAct-Desktop (OOD)

To evaluate GEM on the AITZ train set (ID) and test using AITZ test (ID) and OmniAct-Desktop test (OOD):

1. Extract input scores (for both ID and OOD datasets):

   python run.py

2. Fit GMM and perform OOD detection:

   python GEM.py

🔍 Note: Baseline methods (e.g., MSP, Energy, Mahalanobis) are also available in run.py (see commented sections).

📋 Citation

@article{wu2025gem,
  title={GEM: Gaussian Embedding Modeling for Out-of-Distribution Detection in GUI Agents},
  author={Wu, Zheng and Cheng, Pengzhou and Wu, Zongru and Dong, Lingzhong and Zhang, Zhuosheng},
  journal={arXiv preprint arXiv:2505.12842},
  year={2025}
}