Qodo出了个免费学院，专门教你怎么评估AI代码审查工具，不用注册就能看。里面讲清楚了基准测试怎么才算靠谱，还给了实测数据，挺实在的。
Qodo发布了AI Code Review Academy，这是一套免费的代码审查学习章节，无需注册即可访问。其中基准测试与工具对比章节建议，评估审查工具时应使用真实生产环境的PR数据，并让所有工具在默认设置下运行。该章节引用2025年一项研究，同一模型在孤立基准测试中得分84-89%，但在真实代码库中仅得25-34%。Qodo建议用自己代码库的10-20个PR来测试候选工具。
Qodo just launched the AI Code Review Academy. It’s a free set of chapters from @QodoAI on code rev...
Qodo just launched the AI Code Review Academy. It’s a free set of chapters from @QodoAI on code review in the era of AI-generated code. No sign-up required. I went through the benchmarks and tool comparison chapter. I would recommend it to anyone evaluating review tools right now. It gives you a way to judge whether a benchmark means anything before you act on it. - The dataset has to be real production PRs. - Every tool has to run at default settings; otherwise you are measuring tuning effort. - The methodology has to be open enough to rerun yourself. - And precision and recall have to be reported together, so you know how much a tool catches and how much of what it flags is real. The number that stuck with me came from a 2025 study cited in the chapter. Same model, two evaluations. It scored 84-89% on an isolated benchmark and 25-34% inside a real codebase with dependencies and conventions. The chapter's best advice is to run every tool you are considering against the same 10-20 pull requests from your own codebase. Academy: qodo.ai/academy/?utm_s… Thanks, Qodo, for partnering on this post. 💬 4 🔄 2 ❤️ 12 👀 1954 📊 6 ⚡