一句话：提出 EVATest，用领域探索日志与事件奖励引导 LLM 检索增强生成 Android GUI 测试序列。
移动应用 GUI 测试成本高，通用大语言模型难以理解领域特定的界面设计与交互模式。论文提出 EVATest，其直觉是同一领域的应用往往共享相似的 GUI 设计和交互模式。方法上将同领域应用的随机探索日志及其中输入事件的奖励值作为检索增强上下文，引导 LLM 推理 GUI 测试的目标与机制，从而生成更有效的 GUI 输入事件序列。在快餐和旅行两个领域的 10 个流行应用上进行初步评估，与最先进的基于 LLM 的方法相比，交互到的和见到的可点击元素数量分别提升 41.9% 和 23.7%。
GUI Testing is an important phase in the quality assurance of mobile applications. When performed manually, it is a costly and tedious process requiring much effort of human testers to explore the software GUI and write scripts. Large Language Models (LLMs) provide a novel solution to largely reduce such effort, but generic LLMs often face challenges in understanding domain-specific app user interface and interaction patterns. In this paper, we present a novel LLM-based approach for mobile GUI testing that leverages a reward-guided retrieval-augmented generation technique and domain knowledge to address these challenges. Our intuition is that apps in the same domain tend to share similar GUI designs and interaction patterns, so we can leverage interaction history with similar GUIs to improve the coverage of the current testing. In particular, our approach uses random exploration logs of apps in the same domain and reward values of input events in those logged explorations as additional context information, which is then used to help LLM reason the goal and mechanisms of GUI testing, and thus generate more effective GUI input event sequences. Our initial evaluation on 10 popular apps from two domains (fast food and traveling) shows promising results, boosting the number of interacted and seen clickables by 41.9%, and 23.7%, respectively, compared with the state-of-the-art LLM-based approach.
@inproceedings{paperbot11957,
title = {EVATest: Domain-Oriented Android GUI Testing based on Reward-Guided Retrieval-Augmented Generation},
author = {Bhavana Kondeti and Guanqun Yang and Yui Takashima and Xueqing Liu and Xiaoyin Wang},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3786582.3786810}
}