[论文解读] The Persona Paradox: Medical Personas as Behavioral Priors in Clinical Language Models

Tassallah Abdullahi, Shrestha Ghosh+6 more|arXiv (Cornell University)|Jan 8, 2026

Persona Design and Applications被引用 0

一句话总结

本研究系统评估医疗人设（专业角色与互动风格）如何在分诊与患者安全任务中影响临床大语言模型，揭示受情境影响的非单调效应，而治理级提示并不能保证安全性或专业性。

ABSTRACT

Persona conditioning can be viewed as a behavioral prior for large language models (LLMs) and is often assumed to confer expertise and improve safety in a monotonic manner. However, its effects on high-stakes clinical decision-making remain poorly characterized. We systematically evaluate persona-based control in clinical LLMs, examining how professional roles (e.g., Emergency Department physician, nurse) and interaction styles (bold vs.\ cautious) influence behavior across models and medical tasks. We assess performance on clinical triage and patient-safety tasks using multidimensional evaluations that capture task accuracy, calibration, and safety-relevant risk behavior. We find systematic, context-dependent, and non-monotonic effects: Medical personas improve performance in critical care tasks, yielding gains of up to $\sim+20\%$ in accuracy and calibration, but degrade performance in primary-care settings by comparable margins. Interaction style modulates risk propensity and sensitivity, but it's highly model-dependent. While aggregated LLM-judge rankings favor medical over non-medical personas in safety-critical cases, we found that human clinicians show moderate agreement on safety compliance (average Cohen's $κ= 0.43$) but indicate a low confidence in 95.9\% of their responses on reasoning quality. Our work shows that personas function as behavioral priors that introduce context-dependent trade-offs rather than guarantees of safety or expertise. The code is available at https://github.com/rsinghlab/Persona\_Paradox.

研究动机与目标

• Investigate how professional roles (ED physician/nurse) and interaction styles (bold vs. cautious) as personas influence clinical decision-making in LLMs.
• Characterize how persona conditioning affects task performance, calibration, consistency, and risk behavior in high-stakes clinical tasks.
• Develop a multidimensional evaluation framework to assess safety-relevant trade-offs introduced by persona priors in clinical LLMs.

提出的方法

• Instantiate personas via system prompts defining professional role and interaction style.
• Evaluate on two clinical tasks: emergency triage (high-acuity) and primary-care triage (lower-acuity) plus patient-safety compliance tasks.
• Use automated metrics (accuracy, risk propensity, risk sensitivity, consistency rate, calibration) to quantify behavioral shifts.
• Supplement with qualitative assessments from three judge LLMs and blinded clinician evaluation to gauge safety and reasoning quality.
• Compare medical vs non-medical persona baselines and analyze model- and task-dependent effects.

Figure 1: Experimental framework for analyzing personas as behavioral priors in clinical LLMs. Personas are injected via system prompts (A). Models are evaluated on two clinical tasks, yielding decision labels, free-text justifications, and latent logit scores (B). Behavioral effects are quantified

实验结果

研究问题

• RQ1 RQ1: How does persona conditioning affect clinical performance and safety across triage and patient-safety tasks?
• RQ2 RQ2: How do interaction styles modulate model risk posture and associated trade-offs?
• RQ3 RQ3: How do LLM judges and clinicians perceive persona-induced differences in safety and reasoning?

主要发现

• Medical personas improve emergency triage performance by up to ~20 percentage points in accuracy and calibration compared with baselines.
• In primary-care triage, the same medical personas often degrade performance by a similar magnitude, showing context-dependent effects.
• Interaction styles (bold vs. cautious) produce non-monotonic, model-dependent shifts in risk propensity and risk sensitivity, not a reliable control mechanism for safety posture.
• LLM-based judges generally prefer medical personas in safety-critical cases, but human clinicians show only moderate agreement on safety compliance and low confidence on reasoning quality for many cases.
• Across models, physicians’ personas yield stronger perceived safety, with clinicians demonstrating some confidence in these judgments; however, justification quality remains ambiguous and inconsistent across tasks.
• Overall, personas act as context-dependent behavioral priors that trade safety/expertise for context, rather than universally improving clinical decision-making.

Figure 2: Persona effects on Clinical Triage. Bars show $\Delta$ relative to no‑persona baseline. On average, medical Personas improve emergency performance but degrade primary care performance, with model‑dependent effects on consistency. Arrows represent the directionality of the metric. ’*’ repre