Term
Lost in the middle
A failure mode of long-context LLMs where retrieval accuracy is high for content at the start and end of the prompt but drops in the middle. Documented in a 2023 paper and still observable in 2026 frontier models at extreme context lengths.
Background
The U-shaped attention pattern means that important facts buried in the middle of a 500k-token prompt are systematically under-attended. Mitigations: re-rank retrieval so high-relevance chunks land near the end, summarise long histories before the next turn, and avoid concatenating dozens of full documents without compression.