liu-2023-gpt35-turbo-position-swing
IN premise — summaries/2026/08/24/liu-2023-lost-in-middle-s0-abstract.md
Created 2026-08-25T02:58:08+00:00
GPT-3.5-Turbo accuracy swings from approximately 70% to approximately 55% depending on whether the answer-containing document is at the edges or middle of 20 retrieved documents (~4K tokens).
Summary
When the relevant document lands in the middle of a 20-document context, GPT-3.5-Turbo misses the answer about 45% of the time, versus roughly 30% when it sits near the start or end. This means document ordering isn't just a cosmetic detail; how a retrieval pipeline ranks and arranges results directly swings accuracy by 15 points, so context layout has to be treated as a first-order design decision, not an afterthought.