📄 中文摘要
在生产环境中使用大型语言模型（LLMs）时，系统性失败往往比随机错误更为常见。这些失败源于模型的架构和训练过程。研究者识别了11种行为失效模式，并引用了60多篇学术文献。其中，'幻觉'或'虚构'是一个主要问题，模型可能会自信地引用不存在的库，并在被问及原因时编造出看似合理的解释。研究者Farquhar等人提出了'语义熵'的概念来检测这种现象，通过聚类语义上等价的答案并计算熵值来判断。高熵值通常意味着信息的虚构。为应对这一问题，建议使用检索增强生成（RAG）和验证链等方法。
大型语言模型在生产中的11种失败方式（附学术来源）
出处: 11 Ways LLMs Fail in Production (With Academic Sources)
发布: 2026年3月19日
📄 English Summary
11 Ways LLMs Fail in Production (With Academic Sources)
In production environments, systematic failures of large language models (LLMs) are often more prevalent than random errors, stemming from the models' architecture and training processes. Researchers have identified 11 behavioral failure modes, supported by over 60 academic sources. A significant issue is 'hallucination' or 'confabulation,' where the model confidently references non-existent libraries and fabricates plausible justifications when questioned. Farquhar et al. introduced the concept of 'semantic entropy' to detect this phenomenon by clustering semantically equivalent answers and calculating entropy values. High entropy typically indicates probable fabrication. To mitigate this issue, strategies such as Retrieval-Augmented Generation (RAG) and Chain-of-Verification are recommended.