arXiv cs.CL· Krithik Vishwanath, Brandon Ye, Anton Alyakin, John E. Markert, Aaron Hsieh, Micha{\l} Ma\'nkowski, Eric K. Oermann·· 10 天前AI 评分42
大语言模型易受临床文档与推理中附带信息干扰
Large language models are vulnerable to incidental information in clinical documentation and reasoning
AI 导读
研究评估了 LLM 对患者就诊中附带信息的敏感性：在 576 段医患对话中，前沿模型将闲聊内容写入 35% 的笔记，3.7% 的笔记出现信息误归因或临床误用；在 57 段模拟录音中，-10 dB 的另一患者背景语音泄漏进 48.2% 的转录文本，四个开放权重模型生成的 5.3% 下游笔记被检出污染。研究提出临床推理与分心的双编码假说，建议临床使用前评估模型对附带信息的抵抗能力。
正文 · 原文
Abstract:Large language models (LLMs) are increasingly relied upon to support ambient documentation and clinical reasoning. Here we examine the impact of a failure mode shared between these two applications by assessing their sensitivity to information incidental to the patient encounter. In 576 patient-clinician dialogues, we found that frontier models inserted small-talk exchanges into 35% of notes, while mean quality scores changed by at most 0.20 points on five-point scales. In 3.7% of frontier notes, models misattributed the asides or used them clinically. In 57 mock recorded consultations, background speech from a separate patient encounter at -10 dB leaked into 48.2% of transcripts, with contamination detected in 5.3% of downstream notes generated by four open-weight models. We propose a dual encoding hypothesis of clinical reasoning and distraction in LLMs, with preliminary evidence that LLM components associated with disruption by incidental information also support clinical reasoning. These findings support evaluating resistance to incidental information before clinical use, with safeguards that prevent contamination while preserving clinical reasoning.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.08585 [cs.CL] |
| (or arXiv:2610.08585v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.08585 arXiv-issued DOI via DataCite |
Submission history
From: Krithik Vishwanath Mr. [view email]
[v1]
Tue, 6 Oct 2026 15:55:29 UTC (4,129 KB)
[v2]
Wed, 7 Oct 2026 13:16:20 UTC (4,129 KB)
来源：arXiv cs.CL · arxiv.org