RAG formatting can distort model attention even when content is unchanged

A new arXiv paper argues that RAG systems can be affected not just by what they retrieve, but by how the retrieved context is formatted.

The authors call the effect a structural attention tax: knowledge-graph triples can attract much more model attention per token than semantically equivalent natural-language text. That can crowd out demonstrations or other useful context.

The finding is a reminder that context engineering is not only retrieval quality. Layout, delimiters, and representation choices can change model behavior even when the underlying facts are the same.