Skip to main content
Back to List
Natural Language Processing

Lost in the Middle

A long-context failure mode where mid-document information is underused compared with beginning or end segments

#Lost in the Middle#context loss#mid-context degradation#long context

What is Lost in the Middle?

Lost in the Middle is a long-context behavior where models rely less on information in the middle of a document than on the beginning or end.

How does it show up?

As context length grows, models may skip central evidence and cite less relevant sections.

  1. Missing key evidence from middle sections
  2. Citation bias toward opening or closing segments
  3. Weaker reasoning across multi-part documents

Why does it matter?

This directly affects reliability in RAG, long-form QA, and report generation.
Mitigations typically include chunk design, reranking, and explicit context refresh.

Related terms