gpt35-30doc-mid-context-penalty

IN premise — summaries/2026/08/24/liu-2023-lost-in-middle-s20-t-otal-retrieved-documents.md

Created 2026-08-25T02:58:09+00:00

In the 30-document setting, GPT-3.5-Turbo (16K) drops from 73.4% accuracy at index 0 to 50.5% at index 9, then recovers to 63.7% at index 29—a >20-point mid-context penalty.

Summary

GPT-3.5-Turbo is significantly worse at finding the right document when it sits in the middle of a 30-document list compared to when it appears at the start or end, losing more than 20 percentage points of accuracy in the mid-range. This means any system that relies on this model to retrieve or reference information from a long context must account for position bias, either by reordering documents, chunking contexts shorter, or verifying that mid-position items aren't being silently missed.