The Middle Penalty: Why LLMs Struggle to Find Information Buried in Long Contexts

This finding exposes a critical flaw in how modern LLMs process information: accuracy depends not just on context length, but on where information sits in the sequence. For RAG systems and long-context applications, this means simple document concatenation fails—strategic re-ranking and information placement become as critical as retrieval quality itself.