# Autonomous Large Language Model Agents Enabling Intent-Driven Mobile GUI Testing

**Authors:** Juyeon Yoon, Robert Feldt, Shin Yoo  
**arXiv:** [2311.08649](https://arxiv.org/abs/2311.08649)  
**Date:** 2023-11-15  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2311.08649v1/

## Key points

- The paper introduces DroidAgent, an autonomous GUI testing agent for Android
- DroidAgent aims to revolutionize GUI testing through semantic, intent-driven approaches
- Traditional GUI testing requires manual selection based on lower-level metrics like code coverage or activity coverage
- DroidAgent leverages Large Language Models and support mechanisms for setting task goals and interacting with Android apps autonomously
- Empirical evaluation using 15 apps from the Themis benchmark showed DroidAgent achieved higher autonomy in performing realistic tasks compared to current techniques
- DroidAgent autonomously created tasks like creating a second account and adding a friend in a messaging app without human intervention
- On average, DroidAgent achieved 61% activity coverage, surpassing existing methods' 51%
- Manual analysis revealed that out of 374 autonomously created tasks, 317 were considered realistic and relevant to app functionalities
- DroidAgent deeply interacted with apps and effectively covered more features
- The study highlights the potential of autonomous agents like DroidAgent in improving efficiency and effectiveness through intent-driven automation powered by Large Language Models

## Summary

The paper "Autonomous Large Language Model Agents Enabling Intent-Driven Mobile GUI Testing" by Juyeon Yoon, Robert Feldt, and Shin Yoo introduces DroidAgent, an autonomous GUI testing agent for Android that aims to revolutionize GUI testing through semantic, intent-driven approaches. Traditional GUI testing requires manual selection of what to test based on lower-level metrics like code coverage or activity coverage. In contrast, DroidAgent leverages Large Language Models and support mechanisms such as long- and short-term memory to autonomously set relevant task goals and interact with Android apps to achieve them. The authors conducted an empirical evaluation of DroidAgent using 15 apps from the Themis benchmark. The results showed that DroidAgent was able to autonomously perform realistic tasks with a higher level of autonomy compared to current state-of-the-art GUI testing techniques. For instance, when testing a messaging app, DroidAgent autonomously created a second account and added a first account as a friend, demonstrating its ability to test realistic use cases without human intervention. On average, DroidAgent achieved 61% activity coverage, surpassing the 51% achieved by existing GUI testing methods. Manual analysis revealed that out of the 374 autonomously created tasks, 317 were deemed realistic and relevant to app functionalities. Additionally, DroidAgent was found to deeply interact with the apps and effectively cover more features. Overall, this study showcases the potential of autonomous agents like DroidAgent in improving the efficiency and effectiveness of  through intent-driven automation powered by Large Language Models.