"Lost in the Middle: How Language Models Use Long Contexts" is a paper published in 2023 by Nelson F. Liu and colleagues at Stanford University. In experiments that gave models multiple documents and asked them to find an answer, accuracy was high when the relevant information sat at the very beginning or end of the input but dropped sharply when it was in the middle, producing a U-shaped performance curve. The same tendency appeared in models that support long context windows.
The phenomenon has come to be known as "lost in the middle" and is frequently cited in practical LLM work, including how documents are ordered in RAG (retrieval-augmented generation), how long prompts are structured, and designs that place instructions to agents at the start or end of the context.