一句话：提出编译器感知的迭代式测试适配技术，用大语言模型修复 fork 间因 API 漂移等导致的测试不可复用问题。
基于 fork 的软件生态允许各分支独立演化，但也导致质量保证工作碎片化：一个 fork 中开发的测试往往对其它 fork 有价值，却因 API 漂移、重构、依赖演化等原因难以直接复用，此前研究虽表明大量测试具备潜在可复用性，适配工作却基本依赖人工。本文探索以大语言模型辅助测试用例适配，作为在 fork 生态中自动传播测试的新方向。作者提出一种编译器感知的迭代式适配技术，将某 fork 的测试集成到另一 fork，并利用 LLM 根据构建失败所暴露的不兼容信息反复修复测试代码。在 Java fork 对上的实证结果显示，对于因测试代码本身需要修改而无法直接复用的情形，LLM 引导可在 80% 的案例中成功完成适配。文章最后展望了实现高效、可扩展测试传播的研究方向。
Fork-based software ecosystems enable independent evolution, but fragment the quality assurance efforts. Although tests developed in one fork are often relevant to others, direct reuse is rare and can be difficult, due to API drift, refactorings, dependency evolution, or other changes. Prior work has shown that many tests are potentially reusable across forks, yet adapting them remains a largely manual task. This paper explores LLM-assisted test case adaptation as a new direction for automating test propagation in fork ecosystems. We propose a compiler-aware, iterative adaptation technique that integrates tests from one fork into another and leverages LLMs to repair incompatibilities revealed by build failures. We report empirical results on Java fork pairs, showing that LLM guidance can successfully adapt tests in 80% of cases where direct reuse fails due to modifications required to the test code itself. We outline research directions toward effective and scalable test propagation.
@inproceedings{MukelabaiSNB26,
title = {Towards Automated Test Adaptation in Fork Ecosystems via Large Language Models},
author = {Mukelabai Mukelabai and Keanu-Wesley Schurkus and Yannic Noller and Thorsten Berger},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3805594}
}