一句话：BuilDroid 用自校正 LLM 智能体自动修复 Android 构建错误，在 245 个开源项目上达到 90.2% 成功率。
Android 生态快速演化且高度碎片化，开源项目常因配置漂移、版本不兼容、依赖缺失和工程配置不一致而陷入构建失败，影响可复现性与维护。BuilDroid 是一个基于 LLM 的自动化构建智能体，在隔离且自包含的环境中运行迭代式自校正循环：捕获构建错误后，优先使用预定义启发式规则修复，必要时借助底层 LLM 的推理能力自主解决。作者在 245 个开源 Android 项目上评估，构建成功率达 90.2%，比现有方案高出 30.2 个百分点以上，降低了构建门槛并提升了软件可复现性，支持更广泛可靠的实证研究。
The continuous evolution of the Android ecosystem has led to a highly dynamic and fragmented development environment. This constant churn makes building Android projects, especially from open-source repositories, a notoriously difficult task. Developers and researchers encounter a daunting build barrier due to the rapid configuration drift, which results in a cascade of errors. These errors include version incompatibilities, missing dependencies, and inconsistent project configurations, hindering reproducibility and maintainability.To address these issues, we present BuilDroid, an LLM-based agent that automates the build process of Android projects. Operating within a self-contained, isolated environment, BuilDroid runs an iterative, self-correcting loop. Through this operation, BuilDroid captures errors and autonomously resolves them, either through predefined heuristics or by leveraging the reasoning capabilities of its underlying LLM.Across 245 open-source Android projects, BuilDroid effectively resolves complex and evolving build errors, achieving a build success rate of 90.2%, surpassing existing solutions by a margin of over 30.2 percentage points. Consequently, BuilDroid reduces the barrier for researchers and developers, fostering greater software reproducibility and enabling more extensive and reliable empirical research within this rapidly evolving ecosystem.Video demo: https://youtu.be/YAFLu7NSl5E
DOI 原文 ·
@inproceedings{paperbot9212,
title = {BuilDroid: A Self-Correcting LLM Agent for Automated Android Builds},
author = {Jaehyeon Kim and Rui Rua and Karim Ali},
booktitle = {ASE},
year = {2025},
doi = {10.1109/ase63991.2025.00370}
}