The Persona Paradox: Medical Personas as Behavioral Priors in Clinical Language Models
Abstract
Persona conditioning in clinical language models produces context-dependent effects on performance and safety that vary systematically with professional role and interaction style, challenging assumptions of monotonic improvement in expert behavior.
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.
Community
This paper investigates how "persona conditioning" (e.g., instructing an LLM to act as a specific medical professional) impacts clinical decision-making. The authors challenge the assumption that assigning a medical persona consistently improves accuracy or safety, labeling this inconsistency the "Persona Paradox."
Key Insights:
Non-Monotonic Effects: Assigning a medical persona (like an Emergency Department physician) does not always improve performance. It acts as a behavioral prior that can help in some contexts but hurt in others.
The Context Gap: Medical personas improved accuracy and calibration by up to 20% in critical-care tasks (triage) but degraded performance by a similar margin in primary-care settings.
Interaction Styles: Adding styles such as "bold" or "cautious" changes the model’s risk propensity, but these effects vary widely across base models.
The Alignment Gap: While "LLM judges" preferred medical personas for safety-critical cases, human clinicians were much more skeptical. Human experts showed low confidence in the AI's reasoning quality in 95.9% of cases, despite moderate agreement on safety compliance.
Conclusion
The study concludes that personas are not "expertise switches" but rather priors that introduce context-dependent trade-offs. Relying on personas for clinical safety is risky because they do not provide a universal guarantee of better judgment. Personas should be used with caution in high-stakes medicine, as they can inadvertently trigger biases or performance drops depending on the specific clinical task.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Beyond the ACE Score: Replicable Combinations of Adverse Childhood Experiences That Worsen Depression Risk (2025)
- WET -- Weighted Ensemble Transformer for Identifying Psychiatric Stressors Related to Suicide X (formerly Twitter) (2025)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper