HuggingFace Daily Papers（社区热门论文）· HuggingFace Daily Papers（社区热门论文）·· 2026-05-17精选AI 评分73
从可运行到可交付：基于多智能体测试驱动的开发范式用于从需求生成全栈Web应用
AI 导读
针对编码智能体生成的Web应用超70%不满足需求的问题，本文提出TDDev框架。该框架通过三阶段实现自动化闭环：先将需求转化为结构化测试，再通过浏览器模拟交互验证应用，最后将故障转化为修复报告。首次针对Web应用生成的TDD实证研究发现，引入TDD基础设施可提升质量34-48个百分点。关键结论是最佳协议需与模型生成风格匹配，不匹配将完全抵消TDD优势并最多增加25倍Token消耗。用户研究证实，该框架使人工干预降为零，开发转向自主反馈优化。
推荐理由
把TDD塞进多智能体代码生成，直接把Web应用的正确率从不到30%拉到70%以上，更重要的是他们发现给不同模型配错了开发协议反而会雪崩，做Agent工程的必读。
正文
Abstract:Coding agents can generate runnable web applications, but their outputs frequently fail to satisfy functional requirements. Although test-driven development (TDD) offers a principled repair loop, applying it to web applications requires deriving executable acceptance tests, validating behavior through dynamic browser interactions, and translating observed failures into actionable feedback. We present TDDev, an experimental instrument that automates these tasks and enables closed-loop TDD with minimal human mediation. Using TDDev, we conduct the first controlled study of TDD for full-stack web application generation across 20 diverse web applications, multiple backbone models, two coding agents, and three TDD implementations. With sufficiently capable backbones, TDD improves accuracy by 15.5--23.7 percentage points and remains effective across both minimal and full-featured coding agents; replacing unreliable feedback with a stronger tester restores positive gains for a lower-capability backbone. Incremental TDD provides bounded control, Whole-Project TDD enables coordinated repair and achieves the highest cost efficiency in most configurations, and Agentic TDD performs best with highly capable models. Based on these findings, we provide a practical decision tree that selects an implementation according to model capability, feedback reliability, and deployment objectives. By automating acceptance testing and failure-guided repair, TDDev also reduces the human testing and intervention required during development.
| Comments: | Accepted by the 41st IEEE/ACM International Conference on Automated Software Engineering (ASE'26) |
| Subjects: | Software Engineering (cs.SE) |
| Cite as: | arXiv:2605.17242 [cs.SE] |
| (or arXiv:2605.17242v2 [cs.SE] for this version) | |
| https://doi.org/10.48550/arXiv.2605.17242 arXiv-issued DOI via DataCite |
|
| Related DOI: | https://doi.org/10.1145/3832783.3834388
DOI(s) linking to related resources |
Submission history
From: Yuxuan Wan [view email]
[v1]
Sun, 17 May 2026 03:48:41 UTC (840 KB)
[v2]
Wed, 16 Sep 2026 03:08:26 UTC (8,891 KB)
来源：HuggingFace Daily Papers（社区热门论文） · arxiv.org