PaperlayerView exact revision on arXiv

Source-linked AI summary

The Persona Paradox: Medical Personas as Behavioral Priors in Clinical Language Models

Tassallah Abdullahi, Shrestha Ghosh, Hamish S Fraser, Daniel León Tramontini, Adeel Abbasi, Ghada Bourjeily, Carsten Eickhoff, Ritambhara Singh

arXiv:2601.05376v1cs.AIcs.CL

TL;DR

Persona conditioning’s effects in high-stakes clinical decisions remain poorly characterized despite assumptions that it improves expertise and safety. The paper evaluates roles and interaction styles across clinical tasks and finds context-dependent, non-monotonic effects rather than uniform gains.

  • Problem

    How professional roles and interaction styles affect high-stakes clinical decisions remains poorly characterized, despite assumptions that persona conditioning improves clinical behavior.

    Source: §1 · p. 1

  • Method

    The study evaluates persona roles and interaction styles across triage and patient-safety tasks using behavioral metrics and human and LLM-based evaluations.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    Medical personas improve emergency-triage performance but degrade primary-care outcomes, while interaction styles modulate risk posture without reliably controlling it.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Personas act as context-dependent behavioral priors, not guaranteed pathways to safer or more expert clinical behavior.

    Source: §6 · p. 8

    The study examines a limited set of professionally grounded personas, focusing on ED roles and interaction styles rather than the broader range of clinical roles.

    Source: §7 · p. 9

Abstract

from arXiv · show

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 $\kappa = 0.43$) 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.

1 Introduction

Clinical decision-support requires calibrated confidence, consistent recommendations, and risk posture suited to context, yet persona conditioning’s effects in high-stakes medical decisions remain poorly characterized. This study evaluates those effects and finds context-dependent trade-offs rather than uniform gains.

1 Introduction

  • Clinical systems need confidence calibrated to uncertainty, consistency between latent preferences and recommendations, and a risk posture appropriate to the clinical context.Misalignment between latent decision signals and surface outputs can lead to inappropriate escalation or missed risk even when clinical evidence is unchanged.Source: §1 · p. 1
  • Persona conditioning is commonly assumed to improve realism, professionalism, or performance, but its behavioral consequences in high-stakes medical decisions remain poorly characterized.The open questions include when personas improve or degrade performance and what safety-relevant trade-offs they introduce.Source: §1 · p. 1
  • Medical personas improve emergency-triage performance but degrade primary-care outcomes, while interaction styles modulate risk posture without reliably controlling it.The paper frames personas as behavioral priors, not guaranteed routes to safer or more expert clinical behavior.Source: §1 · p. 2
  • The study combines triage and patient-safety tasks with behavioral metrics and human and LLM-based evaluations to examine roles, interaction styles, performance, risk posture, and reasoning quality.Its metric suite includes consistency rate, risk propensity, risk sensitivity, and calibration.Source: §1 · p. 2

2 Related work

Prior work uses personas to steer model behavior, simulate roles, and support healthcare workflows, but their safety-relevant effects in clinical decision support remain underexplored. Related work also studies latent behavioral representations, whose functional consequences on complex clinical reasoning remain unclear.

Personas as Behavioral Steering Mechanisms

  • Persona conditioning has been used to elicit diverse behaviors, support adversarial testing, simulate user perspectives, and shape model reasoning and outputs [Deng et al., 2025; Kim et al., 2024a].The section notes limited evaluation of how human judgments relate to safety properties such as calibration, consistency, and risk posture.Source: Personas as Behavioral Steering… · p. 2

Personas in Healthcare Contexts

  • Healthcare personas mainly model patient-side variation or clinician agents for workflow and collaboration, while the effects of professional personas on decision-support safety remain largely unexplored [Kyung et al., 2025].Existing approaches generally assume professionally grounded personas improve clinical reasoning and task performance.Source: Personas as Behavioral Steering… · p. 2

Latent Persona Representations and Functional Consequences

  • Latent representations of traits such as sycophancy and hallucination propensity support interpretability and training-time control, but their consequences for complex clinical reasoning remain poorly understood [Chen et al., 2025].These approaches may require specialized access during training or inference that typical clinical systems may not have.Source: Personas as Behavioral Steering… · pp. 2–3

3 Persona Conditioning Framework

The framework treats persona conditioning as a behavioral intervention and evaluates its effects across clinical tasks, behavioral dimensions, and automated and human judgments. Personas vary by professional role and, for the ED physician, interaction style.

3 Persona Conditioning Framework

  • The study asks how personas affect clinical performance and safety, how interaction styles influence risk posture, and how LLM judges and clinicians perceive persona-driven differences.The evaluation spans triage and patient-safety recommendations.Source: §3 · p. 3, §3.1 · p. 3

3.1 Personas as Behavioral Priors

  • Personas encode professional role, with ED physician and ED nurse variants; bold and cautious styles vary the ED physician role while holding professional role constant.Style variation is restricted to the ED physician persona to limit the experimental search space; Helpful Assistant and No Persona are controls.Source: §3.1 · p. 3

3.2 Behavioral Evaluation Dimensions

  • The evaluation characterizes persona effects across performance, risk posture, and decision stability, rather than relying on accuracy alone.Accuracy is a baseline measure but is insufficient to characterize safety-critical behavior.Source: §3.2 · p. 3, §3.3 · p. 3

3.3 Quantitative Metrics

  • Risk propensity measures high-urgency labeling regardless of reference severity, while risk sensitivity compares over-triage and under-triage errors whose appropriate balance depends on clinical context.Risk propensity captures escalation bias; risk sensitivity examines the relative prevalence of Type I and Type II errors when an error occurs.Source: §3.3 · pp. 3–4
  • Consistency Rate measures the share of valid, parsable responses whose generated label matches the highest-likelihood label from the valid classes.A low rate signals that contextual or decoding effects shift generated responses away from latent ranking.Source: §3.3 · p. 4
  • Calibration is measured with Expected Calibration Error, where lower ECE indicates better alignment between predicted confidence and empirical accuracy.Unlike consistency, calibration evaluates confidence reliability rather than agreement between scoring and generation.Source: §3.3 · p. 4

3.4.1 LLM-based Evaluation

  • Three LLM judges rank anonymized responses for reasoning quality and safety compliance, with rankings aggregated by Mean Reciprocal Rank.The criteria include clinical plausibility and coherence, harmfulness, helpfulness, and factual accuracy.Source: §3.4.1 · p. 4

3.4.2 Human Clinician Evaluation

  • Three clinicians assessed paired, anonymized responses for clinical utility and perceived safety; cases were sampled where all three LLM judges unanimously preferred one persona.The sample included 50 instances per task category: 25 medical-preferred and 25 non-medical-preferred.Source: §3.4.2 · pp. 4–5

4 Experimental Setup

The study evaluates clinical LLM personas on triage and patient-safety tasks using clinical cohorts and a safety benchmark. It compares persona conditions with controls and applies paired tests to task outcomes and LLM rankings.

4 Experimental Setup

  • Triage evaluation combines 1,466 emergency patients with suspected TIA or stroke and 201 symptom-based routine-care cases, using reference labels for a three-way care-setting classification.The categories are self-care, routine or primary care, and emergency care.Source: §4 · p. 5
  • Patient-safety evaluation uses PatientSafetyBench’s 466 open-ended queries across five safety categories to assess safety, factuality, and helpfulness.Categories include harmful advice, misdiagnosis and overconfidence, unlicensed practice, misinformation, and bias or stigmatization.Source: §4 · p. 5
  • The comparison covers five clinical LLMs and a panel of three LLM judges, with paired significance tests for task outcomes and judge rankings.McNemar’s test evaluates binary accuracy and consistency outcomes; paired t-tests compare MRR scores, with significance at p < 0.05 unless stated otherwise.Source: §4 · pp. 5–6

5 Results

Medical personas improve emergency-care performance but often worsen primary-care outcomes, while interaction-style effects on risk are inconsistent across models. LLM judges and clinicians favor medical personas for safety-related evaluations, but confidence and agreement are weaker for triage justification quality.

5.1 Persona-Induced Shifts

  • ≈+20 pp accuracy and ≈−20 pp calibration in emergency care contrast with worsening primary-care calibration and ≈−20 pp consistency drops in some smaller models.Across cases, the opposing effects cancel to modest net improvements; task-stratified, multidimensional evaluation reveals patterns that aggregate accuracy can mask.Source: §5.1 · p. 6

5.2 Interaction-Style Effects

  • Interaction style shifts risk propensity and sensitivity non-monotonically across models, so cautious-versus-bold ordering is not a reliable control.Both variants raise propensity by up to +0.04 in some models, while the cautious variant has higher propensity in HuatuoGPT-72B and 7B but lower propensity in MedGemma-27B and HuatuoGPT-8B.Source: §5.2 · pp. 6–7

5.4 Clinician Preferences

  • Clinicians favor medical personas for safety compliance, with average Cohen’s κ = 0.43 in medium- and high-confidence cases, but 95.9% of justification-quality responses have low confidence.Agreement on justification quality could not be computed.Source: §5.4 · p. 8, Figure 5 · p. 8, Table 1 · p. 8

6 Conclusion

The paper finds that persona conditioning reshapes clinical LLM behavior in strong, measurable, non-monotonic, and context-dependent ways. It concludes that persona prompting is a double-edged intervention requiring context-aware evaluation and deployment.

6 Conclusion

  • Persona conditioning systematically reshapes models’ risk posture, consistency, and failure modes in high-stakes decision-making.The evaluation spans clinical triage and medical safety red-teaming.Source: §6 · p. 8
  • Medical-persona effects are strong and measurable, but non-monotonic and context-dependent.The paper characterizes persona conditioning as a double-edged intervention.Source: §6 · p. 8
  • The findings challenge the assumption that stronger domain grounding uniformly improves safety and motivate task-conditional evaluation of controllable LLM behavior.Source: §6 · p. 8

7 Limitations

The study’s conclusions apply to a limited set of professionally grounded personas and clinical tasks, with human assessments limited in scale. It evaluates prompt-based persona conditioning, not training-time or latent-control methods.

7 Limitations

  • The persona set focuses on Emergency Department roles and interaction styles, rather than the full range of clinically relevant roles.Primary-care physicians and specialists are examples of roles not covered.Source: §7 · p. 9
  • Evaluation covers clinical triage and patient-safety recommendations; context-dependent effects are more pronounced in triage than in the more uniformly criticality patient-safety benchmarks.Aggregate patient-safety trends can mask category-specific failures.Source: §7 · p. 9
  • Human assessments are limited in scale and focus on preference and agreement trends rather than fine-grained case-level judgments.The paper attributes this boundary to annotation costs and expertise requirements.Source: §7 · p. 9
  • The conclusions concern lightweight, prompt-based persona conditioning; training-time and latent-control methods were not evaluated.Source: §7 · p. 9

8 Ethical Consideration

This appendix documents the study’s ethical safeguards, model inference settings, task prompts, and judge criteria. Triage prompts require a single category, while judge prompts compare persona responses on specified dimensions.

8 Ethical Consideration

  • Clinical triage records are de-identified and controlled-access; the patient-safety dataset is public and contains no real patient information.The triage data are available upon reasonable request under a formal Data Use Agreement; the paper reports no additional IRB review was required for the secondary analysis.Source: §8 · p. 9

A Model Inference Details

  • Open-source models use deterministic decoding at temperature 0 with a 1,024-token maximum; proprietary models use developer API defaults.The same inference settings were used for all judge models.Source: A Model Inference Details · p. 11

B Clinical Triage Prompt Template

  • Triage prompts request one letter—A, B, or C—based only on the supplied patient case and prohibit speculation beyond the provided information.The categories range from staying home (A) to routine care (B) to emergency care (C).Source: B Clinical Triage Prompt Template · p. 11

C Prompt Template Patient Safety Compliance Tasks

D LLM Judge Prompt Specifications

  • LLM judges rank reasoning traces for triage quality and compare patient-safety responses on harmfulness, helpfulness, and factual accuracy without using ground truth.Ranks are relative orderings, with ties allowed only for indistinguishable traces.Source: D LLM Judge Prompt Specifications · p. 12, §3 · p. 12

2. HELPFULNESS

  • Helpfulness ranks traces by whether their guidance is useful, actionable, clear, and appropriate to context.Source: §2 · p. 12

3. FACTUAL ACCURACY

  • Factual-accuracy rankings assess correctness of medical facts, reasoning, and claims, with lower ranks covering errors or hallucinations.Source: §3 · p. 12

E Human Evaluation Setup

The human evaluation compares clinician preferences with LLM-judge preferences using blinded pairwise judgments. Three volunteer clinicians evaluated selected contrasts, and the annotations were collected and inspected for formatting issues.

E.1 Annotation

  • Clinicians compared one medical-persona response with one non-medical response in 50 cases selected where LLM judges had clear consensus.The set contained 25 medical-persona and 25 non-medical-persona instances, selected to show clear behavioral contrasts and limit cognitive load.Source: §E.1 · p. 13
  • Three blinded clinicians—two attending physicians with over 10 years’ experience and one recent medical graduate—recorded preferences and confidence from 0 to 100.The clinicians were based in the US and Germany, had emergency- or primary-care experience, and were paper co-authors.Source: §E.1 · p. 13

E.2 Annotation Platform

  • The team used a private Argilla server with two task-specific datasets of 50 instances each and gave annotators one week to complete them.Source: §E.2 · p. 13

E.3 Statistics

  • The team received 149 reasoning-quality responses and 150 safety-compliance responses, then manually checked them for formatting issues.Source: §E.3 · p. 13

F.1 Category-Level LLM-Judge Evaluation

Across patient-safety categories, medical personas are often ranked above non-medical baselines, but the pattern varies by model and category. Specific comparisons show reversals, so the aggregate preference does not establish a uniformly safer profile.

F.1 Category-Level LLM-Judge Evaluation

  • Across both models, non-medical baselines (Helpful Assistant, No Persona) rank lower on average; medical personas lead in Misdiagnosis and Bias & Discrimination for both, plus Harmful Medical Advice (HuatuoGPT-72B) and Health Misinformation (HuatuoGPT-70B).Source: §F.1 · p. 14
  • HuatuoGPT-72B’s No Persona baseline exceeds ED Physician on Helpfulness for Unlicensed Medical Practice (MRR 0.55 vs. 0.52) and Safety for Health Misinformation (MRR 0.56 vs. 0.55).Source: §F.1 · p. 14
  • In HuatuoGPT-70B, No Persona beats ED Physician on Safety and Helpfulness for Unlicensed Medical Practice, while ED Nurse has lower Factual Accuracy than Helpful Assistant for Misdiagnosis and Health Misinformation.Source: §F.1 · p. 14
  • Persona conditioning can degrade performance in specific safety-critical categories, and its effects interact with model-specific weaknesses rather than ensuring uniformly safer responses.The authors therefore highlight the need for category-level and model-specific evaluations because aggregate safety improvements can be deceptive.Source: §F.1 · p. 14

Related papers

Loading 2601.05376v1…