一句话：用RAG增强LLM，将自然语言测试用例规格自动生成可执行的汽车ECU集成测试脚本，测试工时减少43%。
论文针对汽车行业中电子控制单元（ECU）架构日益复杂、集成测试效率低下的问题，聚焦于测试工程师将自然语言测试用例规格翻译为可在含硬件调试接口的集成测试环境中执行脚本这一环节。为弥补LLM与领域专用测试工具API之间的知识鸿沟，作者构建了面向任务的向量库，除API手册外还纳入以工作流为中心的补充信息，并与包含代码前缀的提示相结合，使LLM能够通过检索增强生成（RAG）产出健壮且正确的集成测试脚本。评估在多核ECU的典型测试场景上进行，这些用例虽非直接取自特定开发项目，但贴近众多汽车ECU开发中的常规实践。结果显示该方法对所有用例均成功生成可执行脚本，相比人工执行的现实基线总测试执行工时降低43%，表明上下文增强的LLM可有效加速汽车领域的专业软件工程任务，同时也指出了在测试覆盖率、可维护性与流程无缝集成等方面尚存挑战。
Improving the efficiency of software integration testing is a critical challenge in the automotive industry, particularly as Electronic Control Unit (ECU) architectures become increasingly complex. This paper addresses the automation of integration test script generation by leveraging Large Language Models (LLMs) with Retrieval Augmented Generation (RAG). Specifically, we target the phase in which test engineers translate natural language test case specifications into executable scripts for integration test environments containing hardware debug interfaces. To bridge the knowledge gap between LLMs and domain-specific test tool APIs, we construct a task-oriented vector store that incorporates both API manuals and supplemental, workflow-centric information. By combining these with prompts containing code prefixes, our method enables LLMs to generate robust and correct integration test scripts. We evaluated our approach on typical test scenarios reflecting industry practices for multi-core ECUs. While the test cases used were not directly taken from a specific development project, they closely mirror those routinely employed across numerous automotive ECU development initiatives. The proposed method successfully generated executable scripts for all cases and reduced total test execution man-hours by 43% compared to a realistic baseline of manual execution. These results demonstrate the practical benefit of context-enriched LLMs in accelerating specialized software engineering tasks within the automotive domain, and it also identifies remaining challenges in extending automation to broader aspects such as test coverage, maintainability, and seamless process integration.
DOI 原文 ·
@inproceedings{paperbot9309,
title = {Acceleration of Automotive Software Development by Retrieval Augmented Integration Test Script Generation},
author = {Masashi Mizoguchi and Kentaro Yoshimura and Keita Nakazawa and Yasuomi Sato and Takahiro Iida and Fumio Narisawa},
booktitle = {ASE},
year = {2025},
doi = {10.1109/ase63991.2025.00314}
}