📄 中文摘要
在使用大型语言模型（LLM）时，许多用户倾向于将所有相关信息放入一个长提示中，以期获得最佳效果。然而，这种方法并不总是有效，尤其是在模型可能忽视明确的约束或在提示中出现自相矛盾的情况下。研究表明，提供过多的上下文反而可能导致模型生成的答案质量下降。对于复杂的代码库，简单的指令如“分析我们的整个代码库并遵循我们的编码模式”可能显得过于乐观，反映出在处理大量信息时，模型可能会陷入“中间迷失”的问题。
迷失在中间：为什么更大的上下文窗口并不总能提升大型语言模型的性能
📄 English Summary
Lost in the Middle: Why Bigger Context Windows Don’t Always Improve LLM Performance
When utilizing large language models (LLMs), many users tend to input all relevant information into a single long prompt, hoping for optimal results. However, this approach does not always yield effective outcomes, particularly when the model may overlook explicit constraints or produce contradictory responses. Research indicates that providing excessive context can lead to a decline in the quality of the generated answers. For complex codebases, simplistic instructions like 'Analyze our entire codebase and follow our coding patterns' can be overly optimistic, highlighting the 'lost in the middle' problem that arises when handling large amounts of information.