文章背景与核心概要
在医疗健康等高风险领域中，大语言模型（LLMs）的可靠性正受到越来越多的审视。本文探讨了当面对临床信息不确定或缺失时，大语言模型表现出的行为特征。研究发现，尽管大模型在标准医学基准测试中表现优异，但在信息模糊或缺失的情况下，它们往往会表现出危险的“过度自信”倾向。
该研究通过在 MedMCQA 数据集中引入语言提示和故意移除正确答案的方法，揭示了模型置信度常与实际准确率严重脱节，从而导致“不安全的自信错误”。这些发现强调了在将大语言模型部署到现实世界的临床工作流之前，亟需建立更具不确定性感知能力的评估框架。
当置信度失效：大语言模型在不确定性与临床信息缺失情况下的过度自信 (When Confidence Fails: Overconfidence in LLMs under Uncertainty and Missing Clinical Information)
作者： Maryam Tahermazandarani, Adnan Mahmood, Fahmida Islam, Quan Z. Sheng
日期： 2026年8月10日
标识符： arXiv:2608.09080
学科领域： 计算与语言 (cs.CL)；人工智能 (cs.AI)；人机交互 (cs.HC)；机器学习 (cs.LG)
Authors: Maryam Tahermazandarani, Adnan Mahmood, Fahmida Islam, Quan Z. Sheng
Date: August 10, 2026
Identifier: arXiv:2608.09080
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC); Machine Learning (cs.LG)
摘要 (Summary)
本文研究了大语言模型（LLMs）在高风险临床环境中的可靠性，重点关注这些模型在面对不确定性和信息缺失时的行为表现。作者证明了虽然大语言模型在标准医学基准测试中表现良好，但在信息模糊或缺失时，它们会表现出危险的过度自信倾向。通过使用语言线索和故意移除答案的方式对 MedMCQA 数据集进行模型测试，该研究表明模型的置信度往往与准确性不匹配，从而导致“不安全的自信错误”。这些发现强调了在将大语言模型部署到现实世界的临床工作流之前，迫切需要更好的不确定性感知评估框架。
This paper investigates the reliability of Large Language Models (LLMs) in high-stakes clinical environments, specifically focusing on how these models behave when faced with uncertainty and missing information. The authors demonstrate that while LLMs perform well on standard medical benchmarks, they exhibit a dangerous tendency toward overconfidence when information is ambiguous or absent. By testing models against the MedMCQA dataset using linguistic cues and deliberate answer removal, the study reveals that model confidence is often misaligned with accuracy, leading to "unsafe confident errors." The findings underscore a critical need for better uncertainty-aware evaluation frameworks before deploying LLMs in real-world clinical workflows.
核心研究发现 (Key Research Findings)
1. 过度自信问题 (The Overconfidence Problem)
研究发现了一种持续存在的失效模式：随着临床不确定性的增加，模型的准确率会下降，但模型自我报告的置信度水平却并未按比例下降。这种错位营造了一种虚假的可靠性错觉，可能导致有害的临床决策。
The research identifies a consistent failure mode: as clinical uncertainty increases, model accuracy drops, but the models' self-reported confidence levels do not decrease proportionally. This misalignment creates a false sense of reliability that could lead to harmful clinical decision-making.
2. 评估框架 (Evaluation Framework)
作者引入了一种双管齐下的评估策略来测试模型的鲁棒性： * 语言不确定性（Linguistic Uncertainty）： 修改提示词，在临床场景中引入模糊性。 * 答案移除（Answer Removal）： 刻意排除正确答案，以测试模型是否能够识别自身知识的缺乏并选择弃权。
The authors introduced a two-pronged evaluation strategy to test model robustness: * Linguistic Uncertainty: Modifying prompts to introduce ambiguity into clinical scenarios. * Answer Removal: Deliberately excluding the correct answer to test whether the model can recognize its own lack of knowledge and choose to abstain.
3. 失效指标 (Metrics of Failure)
该研究利用了几个校准指标来量化模型性能与置信度之间的差距： * 校准差距 (Calibration Gap) * 期望校准误差 (Expected Calibration Error, ECE) * 不安全自信错误率 (Unsafe Confident Error Rate, UCER)
The study utilized several calibration metrics to quantify the gap between model performance and confidence: * Calibration Gap * Expected Calibration Error (ECE) * Unsafe Confident Error Rate (UCER)
4. 结论 (Conclusion)
研究结果突显了不同模型在处理缺失信息时的显著差异。一些模型表现出一种持久的倾向，即在无法获取正确信息时，仍然高度自信地“幻觉”出答案。作者得出结论认为，在缺乏严格且具备不确定性感知测试的情况下，当前的大语言模型尚不具备自主临床使用所需的认识论可靠性。
The results highlight significant variation in how different models handle missing information. Some models demonstrate a persistent tendency to "hallucinate" answers with high confidence even when the correct information is unavailable. The authors conclude that current LLMs lack the necessary epistemic reliability for autonomous clinical use without rigorous, uncertainty-aware testing.