抑制性注意力缓解临床长上下文丢失中间效应
原标题：Inhibitory Attention for Clinical Long-Context Reasoning: Characterizing and Mitigating Lost-in-the-Middle Effects in EHR Processing
AI 摘要
该研究首次系统刻画了电子健康记录（EHR）处理中的临床长上下文丢失中间效应（CLitM），发现大语言模型对位于长上下文中间位置的临床信息检索准确率显著下降，峰值与谷值准确率差距达21.9个百分点。作者提出了一种轻量级的查询条件临床抑制（QCCS）上下文选择门控方法，并在多臂实验中证明其端到端指令遵循性能优于BM25、密集检索和交叉编码器等基线。研究表明，查询对齐的上下文选择比金句检索召回率更能预测EHR指令遵循准确率。
正文节选
Inhibitory Attention for Clinical Long-Context Reasoning: Characterizing and Mitigating Lost-in-the-Middle Effects in EHR Processing Abstract Electronic health records now routinely exceed 100,000 tokens per patient. Yet large language models exhibit a systematic positional retrieval bias, the lost-in-the-middle (LitM) effect, in which information located near the center of a long input context is retrieved far less reliably than information near the edges. For clinical applications, this bias i