一篇新发表在arXiv上的研究调查了大型语言模型（LLM）生成的自然语言注释如何影响代码生成性能。研究人员发现，源自成功代码解决方案的注释能显著提高接收模型生成正确代码的能力，平均将pass@1率提高了17.2%。相反，描述失败解决方案或不相关问题的注释则没有益处，甚至会降低性能。研究结果表明，注释的效用在于其传达正确解决方案内容的能力，而不仅仅是其存在或描述意图。 AI
影响 强调了高质量、面向解决方案的注释对于提高LLM代码生成能力的重要性。
排序理由 学术论文，详细介绍了关于LLM代码生成的最新研究成果。
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
完整方法见我们的编辑标准。
arXiv:2609.09242v1 Announce Type: cross Abstract: Large Language Models (LLMs) often generate natural-language comments while writing code, and these comments become part of the context used to generate the code that follows. However, it remains unclear which properties of commen…
Large Language Models (LLMs) often generate natural-language comments while writing code, and these comments become part of the context used to generate the code that follows. However, it remains unclear which properties of comments affect code-generation performance. We study th…