Lost in the middle
Also known as: middle context degradation, long context recall problem, context position bias
Research showed that even when models have long context windows, their ability to actually use information isn't uniform across that window. Information placed at the very start or end of the context is recalled reliably. Information placed in the middle of a long document is often effectively invisible: models perform significantly worse on questions whose answers are buried there.
This has practical implications for RAG system design and for how you structure prompts that include long documents. If you need the model to attend carefully to a specific piece of information, putting it at the beginning or end of the context is safer than burying it in the middle. Chunking and reranking strategies in RAG systems partly exist to avoid this problem by surfacing only the most relevant passages, rather than feeding entire documents.
Newer models have improved substantially on this problem. Gemini 2.5 Flash, for example, shows much more uniform recall across its context window. But the issue isn't fully solved across all models, and it's worth knowing about when you're debugging a system where the model 'should know' something you gave it but keeps getting it wrong.