AI生成软件的环境可复现性测量

AI 摘要

该论文提出一个三层框架（声明依赖、运行时安装依赖、必要且充分依赖）来评估AI编码代理生成软件环境规范的能力。研究在3个生产级编码代理、4种语言生态和50个编程任务上测试，发现代理在依赖规范上存在系统性泛化失败，同一任务下不同代理的依赖集一致性低至7%，且新代理没有明显改善。最大分歧出现在声明层与运行时层之间，表明模型训练分布中的环境先验是主要驱动因素。

正文节选

Code That Works, Environments That Don’t: Measuring Environment Reproducibility in AI-Generated Software Abstract Code generation has emerged as a central capability of large language models, with coding agents now able to produce functionally correct software projects from natural language prompts. However, functional correctness alone does not capture a critical dimension of generation quality: environment specification, which is defined as the accurate identification of the dependencies requi

发布时间：2026-10-02 12:00

抓取时间：2026-10-02 13:54

来源机构：arXiv