gpt4-20doc-accuracy-range
IN premise — summaries/2026/08/24/liu-2023-lost-in-middle-s20-t-otal-retrieved-documents.md
Created 2026-08-25T02:58:08+00:00
In the 20-document QA setting, GPT-4 (8K) achieves near 90% accuracy at positions 1 and 20 but drops to approximately 70% at middle positions.
Summary
GPT-4 shows a clear "lost in the middle" pattern when answering questions across 20 documents: it handles the first and last documents well but gets noticeably worse on anything in the middle. This means that in a retrieval or multi-document pipeline, simply adding more context does not guarantee uniform accuracy, because where information sits in the sequence actually affects whether the model can find and use it.