liu-2023-20to50-diminishing-returns

IN premise — summaries/2026/08/24/liu-2023-lost-in-middle-s1-introduction.md

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

In open-domain QA, increasing retrieved documents from 20 to 50 yields only approximately 1-1.5% absolute accuracy improvement, indicating model reader saturation before retriever recall saturates.

Summary

Once the model has roughly 20 relevant documents, it can't effectively make use of much more, so pulling in 50 instead of 20 barely improves accuracy. This means the real bottleneck is the model's reading comprehension rather than its ability to find documents, so spending extra effort retrieving more context past that point is wasted compute.