一句话：将 EvoSuite 加固与按簇 LLM 重构相结合，在提升覆盖率的同时大幅降低 token 消耗并改善可读性。
在工业环境中采用自动 Java 单元测试生成，需要同时满足高覆盖率、可读性、token 效率以及对现代 Java 的可部署性，而已有技术无法四者兼顾：基于搜索的生成器在可读性和新版本 Java 支持上表现不佳，以 LLM 为中心的生成器需消耗数千万 token，先前的 LLM 测试重构方法在受控实验中会使基线覆盖率下降 8–21 个百分点。为此，作者提出并部署了混合式生成器 EvoLLM：一方面使用 EvoSuiteNG，即通过失败分析驱动的 EvoSuite 加固版本，解决可部署性问题并强化基线覆盖率；另一方面进行按簇的 LLM 重构，将测试按覆盖率目标分组，并在单次 LLM 调用中基于紧凑的待测程序上下文重写整个簇。在 6 个开源项目和 3 个工业项目上，EvoLLM 优于 6 个最先进基线，行覆盖率提升 8.2–20.7 个百分点，分支覆盖率提升 6.2–19.4 个百分点，变异得分提升 6.8–24.5 个百分点，同时 LLM token 用量减少 47.3%–93.0%。其测试生成阶段在 JDK 8、11 和 17 构建的项目上均展现出稳健的可部署性，并在需要 JDK 17 的 Spring-core 版本上效果显著；14 名开发者参与的调研进一步表明，EvoLLM 生成的测试在可读性上优于先前工具与开发者手写的测试。
Adopting automated Java unit-test generation in industrial settings requires high coverage, readability, token efficiency, and modern-Java deployability. No existing technique meets all four: search-based test generators struggle with readability and modern-Java deployability, LLM-centric generators consume tens of millions of tokens, and prior LLM-based test refactoring methods drop baseline coverage by 8–21 percentage points (pp) in our controlled setups. We report on EvoLLM, a hybrid Java unit-test generator we built and deployed to address all four requirements. EvoLLM combines EvoSuiteng—a failure-analysis-driven hardening of EvoSuite that resolves deployability issues and strengthens baseline coverage—with cluster-wise LLM-based refactoring that groups tests by coverage goals and rewrites each cluster in a single LLM call grounded in compact program-under-test context. Across six open-source and three industrial projects, EvoLLM outperforms six state-of-the-art baselines, improving line coverage by 8.2–20.7 pp, branch coverage by 6.2–19.4 pp, and mutation score by 6.8–24.5 pp, while reducing LLM token usage by 47.3%–93.0%. In addition, the test generation stage of EvoLLM demonstrates robust deployability across target projects built with JDKs 8, 11, and 17, while achieving strong effectiveness on a Spring-core release requiring JDK 17. Finally, a 14-participant developer survey further suggests that EvoLLM-generated tests ranked higher in readability than those from prior tools and developer-written tests.
@inproceedings{paperbot12345,
title = {EvoLLM: High-Coverage, Readable, and Token-Efficient Unit Test Generation for Industrial Java Codebases},
author = {Seokhyeon Moon and Myungho Lee and Jihun Kim and Yoon-Chan Jhi and Jihoon Cho and Hakjoo Oh},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834525}
}