What does research say about: Lost-in-the-middle effect degrades long-context accuracy?

Information placed in the middle of a 10K-token context is recalled 20% less accurately than information at the start or end of the same context. (Source: Liu et al., 'Lost in the Middle: How Language Models Use Long Contexts', Stanford NLP, 2023). Without positional awareness, critical instructions buried in mid-context are ignored by the model's attention mechanism.

Lost-in-the-middle effect degrades long-context accuracy.

Context & Methodology

Without positional awareness, critical instructions buried in mid-context are ignored by the model's attention mechanism.

Applicable Use Cases

workflow

Applies To

openaianthropicgoogle

Primary Impact

quality

Confidence Level

High

Platform Status

Built

Implementation Effort

low

Recommendation

Execution Priority

Put This Evidence to Work

Use the STCO framework to implement findings like this in structured, testable prompts.

Start Building Free Browse All 141 Citations

ROI Calculator Token Calculator Prompt Templates