arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2605.23932cs.AIcs.CLcs.CYcs.LG
收藏

当正确信念崩溃:LLMs在临床压力下的认知韧性

When Correct Beliefs Collapse: Epistemic Resilience of LLMs under Clinical Pressure

Research Center for Social Computing and Interactive Robotics, Harbin Institute of Technology, China(社会计算与交互机器人研究院,哈尔滨工业大学,中国)

Boyu Xiao, Xiuqi Tian, Xuwen Song, Haochun Wang, Guanchun Song, Sendong Zhao, Bing Qin

AI总结:

研究LLMs在临床对话中面对逐步升级压力时信念稳定性问题,提出Med-Stress压力测试框架,发现知识-韧性差距,并设计RBED和R-FT方法提升鲁棒性。

摘要:

尽管在医学基准测试中准确率很高,但LLMs在临床对话中可能表现出严重的多轮谄媚行为,在逐步升级的压力下放弃最初正确的诊断。我们提出了\textbf{\textsc{Med-Stress}},一个针对性的压力测试框架,用于评估在逐步升级压力下的信念稳定性。在九个前沿大型语言模型(LLMs)中,我们发现医学知识与鲁棒性之间存在明显的分离:高初始诊断能力并不意味着高信念稳定性,导致多个LLMs存在较大的知识-鲁棒性差距。为了缓解这种失败模式,我们提出了一种轻量级的推理时防御方法\textbf{\texttt{RBED}}(\textbf{R}ole-\textbf{B}ased \textbf{E}pistemic \textbf{D}efense),以及一种训练时方法\textbf{\texttt{R-FT}}(\textbf{R}esilience-oriented \textbf{F}ine-\textbf{T}uning),该方法内化了基于证据的抗压能力。实验表明,\textbf{\texttt{R-FT}}几乎消除了信念变化,并显著提高了鲁棒性。

英文摘要:

Despite strong medical benchmark accuracy, LLMs can exhibit severe multi-turn sycophancy in clinical dialogue, abandoning initial correct diagnosis under escalating pressure. We propose \textbf{\textsc{Med-Stress}}, a targeted stress test framework that evaluates belief stability under escalating pressure. Across nine frontier large language models (LLMs), we find a clear dissociation between medical knowledge and robustness: high initial diagnostic capability does not imply high belief stability, yielding large knowledge-robustness gaps for several LLMs. To mitigate this failure mode, we propose a lightweight inference-time defense, \textbf{\texttt{RBED}} (\textbf{R}ole-\textbf{B}ased \textbf{E}pistemic \textbf{D}efense), and \textbf{\texttt{R-FT}} (\textbf{R}esilience-oriented \textbf{F}ine-\textbf{T}uning), a training-time approach that internalizes evidence-based resistance to pressure. Experiments show that \textbf{\texttt{R-FT}} nearly eliminates belief change and substantially improves robustness.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑