llm-lost-in-middle-ablation-random-distractors-persist-u-shape
IN premise — summaries/2026/08/24/liu-2023-lost-in-middle-sR-references.md
Created 2026-08-25T02:58:09+00:00
In Liu et al. 2023 Appendix B, replacing hard-negative Wikipedia distractors with randomly sampled Wikipedia documents yields higher absolute accuracy for all models but the U-shaped positional curve persists.
Summary
The lost-in-the-middle effect is not just a byproduct of using tricky or confusing irrelevant passages. Even when the surrounding noise is made easy and random, models still perform worst on information buried in the middle of their context window, which points to a structural weakness in how they attend to position rather than a problem with specific document difficulty.