Guardian: A Runtime Framework for LLM-based UI Exploration

Dezhi Ran, Hao Wang, Zihe Song, Mengzhou Wu, Yuan Cao, Ying Zhang, Wei Yang, Tao Xie

Summary

Guardian improves LLM-based feature-driven mobile UI testing through a runtime framework. It constrains the available UI actions to enforce instructions and checks whether new information invalidates earlier plans. When necessary, it restores the application state and prompts replanning. Evaluation on the FestiVal benchmark of 58 tasks across 23 apps demonstrates improved task success and completion over prior approaches.