llm-lost-in-middle-shuffle-prior-removal
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 C, adding the instruction 'The search results are ordered randomly' and shuffling the k−1 distractor documents removes the decreasing-relevance positional prior but the U-shaped curve remains.
Summary
The "lost in the middle" attention gap isn't caused by documents being arranged in relevance order; even when the instruction explicitly says results are random and the distractor documents are shuffled, the model still attends well to the beginning and end of its context window but poorly to the middle. This means the positional blind spot is a structural feature of the model's attention, not an artifact of how inputs are sequenced, so simply reordering or randomizing context won't fix it.