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

Applications, bias, and safety2026arXivApproved editorial review

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

Keywords: Large Language Models, Personality, Persona, Personality Control, AI Safety

Source: Open primary source (opens in a new tab)

8
Authors
20
Findings
100
Limitations
13
Evidence

Editorial summary

English

This work asks whether a one-sentence system instruction, “You are an Emergency Department Physician/Nurse,” Bold/Cautious variants, “Helpful Assistant,” or no persona, changes five clinical models on two tasks. The audit covers all 14 pages of arXiv:2601.05376v1, its source package, and the official repository at commit 19a7487. As of 15 July 2026, no later definitive publication was found; the first author's CV describes it as ACL/ARR work, so it should be treated as a non-peer-reviewed preprint. The paper provides a clear demonstration of prompt sensitivity, but its results are not equivalent to clinical validation or a clean measure of expertise.

The triage task combines 1,466 de-identified records from an emergency observation unit involving suspected TIA/stroke with 201 lower-acuity symptom cases. Models choose A stay home, B routine/primary care, or C emergency care. The study tests HuatuoGPT-o1-8B and 70B on Llama backbones, HuatuoGPT-o1-7B and 72B on Qwen backbones, and MedGemma-27B. Metrics include accuracy, ECE, agreement between the highest-probability label and generated label, emergency-label propensity, and a ratio of over-triage to under-triage errors. On PatientSafetyBench, responses to 466 open-ended queries are ranked by GPT-5, HuatuoGPT-o1-70B, and 72B for harmfulness, helpfulness, and factual accuracy.

The main descriptive finding is real but narrower than the headline: emergency-department roles shift predictions toward greater urgency. That shift improves accuracy by as much as about 20 percentage points on the high-acuity subset, worsens it by about 10 points on primary-care cases, and changes ECE and consistency in model-dependent ways. No primary-care persona is tested on primary-care cases. The design therefore confounds “medical persona” with contextual congruence: the same emergency role is applied both to emergency and non-emergency data. Source, label prevalence, and clinical context also change together across subsets. Without within-source per-class performance, improved reasoning cannot be separated from a simple shift in the prior toward label C.

The abstract says gains of “~+20% in accuracy and calibration,” whereas the body and figure report percentage-point changes: roughly +20 pp accuracy and −20 pp ECE. Lower ECE is improvement, not a positive 20% calibration gain. Calibration comes from A/B/C token probabilities, not confidence communicated to a patient; the number of bins, binning sensitivity, intervals, and per-class calibration are not reported. Consistency is not access to an internal state either: it compares two observable readings of the same model, conditional label probability and decoded text, and removes unparsable outputs from the denominator.

Bold and Cautious variants do not control risk monotonically. The paper acknowledges that ordering reverses across models, with small propensity differences in some cases and widely different sensitivity ratios. This demonstrates fragility to short phrasings, not a validated manipulation of the psychological constructs “bold” or “cautious.” There is no manipulation check, paraphrase robustness, seed replication, stochastic repetition, complete role-by-style factorial design, or uncertainty for these ratios. Type-I/Type-II ratios can be unstable with small or zero denominators and omit A-versus-B errors.

LLM evaluation favors medical roles in aggregate, but case-level agreement among the three judges is low: 43–53% majority agreement and kappa between 0 and .1. Aggregate statistical differences do not turn those preferences into clinical truth. Judges observe more or less professionalized language, rate partly stylistic dimensions, and are explicitly told not to use ground truth for reasoning-quality rankings. Paired t-tests on bounded ordinal MRR, many persona/model/task/dimension comparisons, no multiplicity correction, and no complete effect/p-value tables limit inference.

Human evaluation uses three volunteer coauthors: two attending physicians with more than ten years of practice and one recent MD graduate. Each sees 50 pairs per task; 149 reasoning judgments and 150 safety judgments are returned. This is neither an independent nor a representative sample. The 50 cases are selected only after all three LLM judges unanimously agree, balanced as 25 medical-preferred and 25 non-medical-preferred. It is validation conditioned on the evaluator's own extreme cases, not an estimate over the full corpus.

The reported kappa=.43 is described ambiguously. Figure 5b shows clinician–clinician kappas of −.10, −.067, and .45, averaging about .09. Clinician–“LLM Judge” kappas are −.091, .76, and .64; their average is .436 and accounts for the cited .43. Thus .43 is not average agreement among clinicians. It is calculated on only 16 safety cases passing a confidence threshold. The statement that 95.9% of reasoning responses have low confidence also does not literally match the histogram, which visibly contains many individual ratings at or above 50; it may refer to cases failing an undefined joint-rater criterion. Annotations are not released to resolve the discrepancy.

The claim that this “validates” LLM judges is too strong: one annotator has negative agreement with the aggregate LLM decision, clinician-to-clinician agreement is low, sampling incorporates LLM consensus, and preferences do not measure outcomes or clinical correctness. Preference percentages describe only selected, confidence-filtered cases. Perceived safety, fluency, and justification are not actual safety.

The released code does not reproduce the full study. The repository contains five Python scripts, one figure, and a one-sentence README; there are no data, outputs, analysis scripts, tables, notebooks, requirements/lockfile, license, tests, or CI. Clinical data are DUA-restricted, but there is not even a synthetic fixture or schema. infer_safety.py hard-codes ED Nurse and runs head(5). run_triage_med_gemma.py uses head(5), while run_triage_reasoning_model.py uses tail(5), absolute GPFS paths, and repeated import blocks. Both judge scripts use head(3). No script computes McNemar tests, t-tests, ECE, MRR, risk metrics, kappa, or figures; there is no GPT-5 judge implementation or ranking-prompt assembly. Default input files are absent.

Paper and code also diverge. The MedGemma prompt requires a “## Final Response” header not present in the paper; the triage judge script asks for ASSERTIVENESS and CLINICAL EXPERIENCE in addition to JUSTIFICATION QUALITY, whereas the paper defines only the latter; and the safety script concatenates its added scenario without a separating space after the query. Parsers accept aliases and phrases even though the protocol requires one letter, adding undocumented decision rules. All Python files compile syntactically, but that is not an executable analysis pipeline and cannot verify any reported result.

The defensible conclusion is that minimal professional-role instructions can alter response priors, generated text, and judge preferences in clinical models, with direction depending on task context and model. This is a useful warning against assuming an expert persona guarantees safety. It does not demonstrate clinical improvement, acquired expertise, a causal role effect separated from label priors, user-facing calibration, deployment safety, or robust clinician agreement.

Español

Este trabajo estudia si una instrucción de sistema de una sola frase, “You are an Emergency Department Physician/Nurse”, sus variantes Bold/Cautious, “Helpful Assistant” o ninguna persona, cambia cinco modelos clínicos en dos tareas. La auditoría cubre las 14 páginas de arXiv:2601.05376v1, su paquete fuente y el repositorio oficial en el commit 19a7487. A 15 de julio de 2026 no se localizó publicación definitiva posterior: el CV de la primera autora lo presenta como trabajo ACL/ARR, por lo que debe tratarse como preprint no revisado por pares. El paper aporta una demostración clara de sensibilidad al prompt, pero sus resultados no equivalen a validación clínica ni a una medida limpia de expertise.

La tarea de triage combina 1.466 registros desidentificados de pacientes de una unidad de urgencias con sospecha de TIA/ictus y 201 casos sintomáticos de menor agudeza. El modelo debe elegir A quedarse en casa, B atención rutinaria/primaria o C urgencias. Se prueban HuatuoGPT-o1-8B y 70B sobre Llama, HuatuoGPT-o1-7B y 72B sobre Qwen, y MedGemma-27B. Se calculan accuracy, ECE, concordancia entre la etiqueta de mayor probabilidad y la generada, propensión a elegir urgencias y una razón de errores de sobretriage/undertriage. En PatientSafetyBench se generan respuestas a 466 consultas abiertas y tres jueces, GPT-5, HuatuoGPT-o1-70B y 72B, ordenan las personas por harmfulness, helpfulness y factual accuracy.

El hallazgo descriptivo principal es real pero más estrecho que el titular: los roles de urgencias desplazan las predicciones hacia mayor urgencia. Ese desplazamiento mejora hasta unos 20 puntos porcentuales la accuracy en el conjunto de alta agudeza, empeora alrededor de 10 puntos en atención primaria y modifica ECE y consistencia de forma dependiente del modelo. No se compara una persona de primaria con los casos de primaria. Por tanto, el diseño confunde “persona médica” con congruencia de contexto: exactamente el mismo rol de urgencias se aplica a datos de urgencias y fuera de urgencias. Además, procedencia, prevalencia de etiquetas y contexto cambian juntos entre los dos subconjuntos. Sin rendimiento por clase dentro de una fuente común, no se puede separar mejor razonamiento de un simple cambio del prior hacia C.

El abstract dice ganancias de “~+20% in accuracy and calibration”, pero el cuerpo y la figura muestran cambios en puntos porcentuales: aproximadamente +20 pp de accuracy y −20 pp de ECE. La ECE menor es una mejora; no es una ganancia positiva de 20%. La calibración se deriva de probabilidades de tokens A/B/C, no de confianza comunicada al paciente, y no se documentan número de bins, sensibilidad a binning, intervalos ni calibración por clase. La consistencia tampoco accede a un estado interno: compara dos lecturas observables del mismo modelo, probabilidad condicionada y texto decodificado, y excluye outputs no parseables del denominador.

Las variantes Bold y Cautious no controlan el riesgo de manera monotónica. El paper lo reconoce: el orden cambia según el modelo, con diferencias de propensión pequeñas en algunos casos y razones de sensibilidad muy dispares. Esto demuestra fragilidad ante formulaciones cortas, no que el constructo psicológico “bold/cautious” haya sido manipulado o medido. No hay manipulation check, paraphrases, seeds, repetición estocástica, factorial role×style completo ni estimación de incertidumbre para estas razones. La razón Type-I/Type-II puede ser inestable con denominadores pequeños o cero y omite errores A↔B.

La evaluación por LLM favorece en agregado a roles médicos, pero el acuerdo caso a caso de los tres jueces es bajo: 43–53% de acuerdo mayoritario y κ entre 0 y 0,1. Las diferencias significativas agregadas no convierten esas preferencias en verdad clínica. Los jueces ven respuestas más o menos profesionalizadas, evalúan dimensiones parcialmente estilísticas y, por instrucción, no usan ground truth en reasoning quality. La t de Student sobre MRR acotado/ordinal, las numerosas comparaciones de personas, modelos, tareas y dimensiones, la ausencia de corrección y la falta de tablas de efectos/p exactos limitan la inferencia.

La validación humana usa tres coautores voluntarios: dos attending physicians con más de diez años y un MD recién graduado. Cada uno evalúa 50 pares por tarea; se reciben 149 juicios de razonamiento y 150 de seguridad. No es una muestra independiente ni representativa. Los 50 casos se seleccionan después de que los tres jueces LLM coincidan unánimemente, balanceando 25 favorables a persona médica y 25 a no médica. Es una validación condicionada a extremos del propio evaluador, no una estimación del corpus completo.

El κ=0,43 está descrito de forma equívoca. La matriz de la Figura 5b muestra κ clínico–clínico de −0,10, −0,067 y 0,45, cuyo promedio es aproximadamente 0,09. Los valores clínico–“LLM Judge” son −0,091, 0,76 y 0,64; su promedio es 0,436 y explica el 0,43 citado. Por tanto, 0,43 no es acuerdo medio entre clínicos. Solo se calcula sobre 16 casos de seguridad que superan el umbral de confianza. La afirmación de que 95,9% de las respuestas de reasoning tienen baja confianza tampoco concuerda literalmente con el histograma, que contiene numerosos ratings individuales ≥50; quizá se refiera a casos que no cumplen un criterio conjunto no definido. No se liberan anotaciones para resolverlo.

La conclusión de que esto “valida” a los jueces LLM es demasiado fuerte: una persona anotadora tiene acuerdo negativo con el agregado LLM, el acuerdo entre clínicos es bajo, la selección incorpora el consenso LLM y las preferencias no miden outcomes ni corrección clínica. La tabla de preferencias solo describe el subconjunto seleccionado y filtrado por confianza. Perceived safety, fluidez o justificación no son seguridad real.

El código publicado no reproduce el estudio completo. El repositorio tiene cinco scripts Python, una figura y un README de una frase; no contiene datos, outputs, análisis, tablas, notebooks, requirements/lockfile, licencia, tests ni CI. Los datos clínicos son restringidos por DUA, pero tampoco se incluye un fixture sintético o un esquema. infer_safety.py fija la persona ED Nurse y procesa solo head(5). run_triage_med_gemma.py procesa head(5) y run_triage_reasoning_model.py tail(5); este último tiene rutas GPFS absolutas y bloques de imports duplicados. Los dos evaluadores solo procesan head(3). Ningún script calcula McNemar, t-tests, ECE, MRR, risk metrics, κ o las figuras; no hay implementación del juez GPT-5 ni ensamblado de prompts de ranking. Los defaults apuntan a archivos ausentes.

Hay además divergencias entre paper y código: el prompt de MedGemma exige un encabezado “## Final Response” ausente en el paper; el evaluador de triage solicita ASSERTIVENESS y CLINICAL EXPERIENCE además de JUSTIFICATION QUALITY, mientras el paper solo define esta última; y el script de safety concatena la frase adicional sin espacio tras el query. Los parsers aceptan aliases y frases aunque el protocolo exige una sola letra, introduciendo reglas no documentadas. Los archivos compilan sintácticamente, pero eso no constituye un pipeline ejecutable ni permite verificar ninguna cifra.

La conclusión defendible es que instrucciones profesionales mínimas pueden alterar el prior de respuesta, el texto y las preferencias de jueces en modelos clínicos, con dirección dependiente del contexto y del modelo. Es una advertencia útil contra asumir que una persona experta garantiza seguridad. No demuestra mejora clínica, expertise adquirido, causalidad del rol separada del prior de etiqueta, calibración para usuarios, seguridad de despliegue ni acuerdo clínico robusto.

Research question

How do a one-sentence medical professional instruction and the Bold/Cautious variants change triage classification, urgency prior, consistency, calibration, and preferences of LLM/human evaluators across five clinical LLMs?

Method

Prompt conditioning experiment with six conditions: ED Physician, ED Nurse, Bold ED Physician, Cautious ED Physician, Helpful Assistant, and No Persona. Five models classify 1,466 high-acuity cases and 201 lower-acuity cases; accuracy, ECE, logit-generation consistency, and risk metrics are computed. Two large models are additionally evaluated on 466 PatientSafetyBench queries via rankings from three LLM judges. Three clinician coauthors compare pairs on 50 cases per task selected by unanimous consensus of the LLM judges.

Sample: There is no patient trial or clinical deployment. The quantitative units are 1,667 triage cases and 466 safety prompts processed by LLMs. Human validation comprises three clinician coauthors (two attending physicians with more than ten years and a recent MD) who complete nearly 100 comparisons each on subsets selected by unanimity of LLM judges.

Findings

  • The ED Physician and ED Nurse roles shift predictions toward higher urgency.
  • In high acuity, the maximum accuracy improvement is around +20 percentage points, not necessarily +20% relative.
  • In primary care, the same emergency roles reduce accuracy by about 10 points in several models.
  • ECE changes reach around 20 points and their direction depends on task and model.
  • Logit-generation consistency improves in some models and falls in others.
  • There is no Primary Care persona, so medical role and context congruence are not separated.
  • The mixture of sources and label prevalences allows a simple shift of the prior toward C to explain a substantial part of the pattern.
  • Bold and Cautious produce non-monotonic changes and orderings that reverse across models.
  • Urgency propensity changes by up to 0.21 and the sensitivity ratio by up to 0.76 relative to no-persona.
  • LLM judges favor medical roles in aggregate for some safety dimensions.
  • Case-by-case agreement of LLM judges is low: 43–53% majority and κ 0–0.1.
  • Results by category show that no-persona outperforms medical roles in some cells of PatientSafetyBench.
  • Humans prefer more medical responses among the medium/high confidence safety judgments of the selected subset.
  • The published matrix gives an approximate mean clinician–clinician agreement of 0.09, not 0.43.
  • The 0.43 arises from averaging κ of each clinician against the aggregated LLM decision over 16 safety cases.
  • One clinical persona has negative κ with the other annotators and with the LLM judge.
  • The claim of 95.9% low confidence in reasoning is incompatible with a literal reading of the histogram of individual ratings.
  • The official repository appeared after the preprint, but only publishes partial demo scripts.
  • All Python scripts compile, although their head/tail limits prevent running the full corpus by default.
  • The robust evidence is contextual sensitivity to the prompt, not clinical expertise or safety.

Limitations

  • It is arXiv v1 and no definitive peer-reviewed publication was located.
  • There is no preregistration, prior protocol, power analysis, or precision plan.
  • Only emergency professional personas are used.
  • No primary care persona is included for the primary care cases.
  • Role and role-context congruence are confounded.
  • The high- and low-acuity subsets come from different sources.
  • Source, context, and label distribution change simultaneously.
  • The 1,466 suspected TIA/stroke cases represent a narrow clinical domain.
  • Performance by diagnosis, class, subgroup, or institution is not reported.
  • Confusion matrices and per-label counts are not published.
  • A shift of the prior toward emergency may improve one subset and harm the other without better reasoning.
  • The textual intervention is not separated from an explicit cost rule or threshold.
  • The personas are a single sentence with no description of competencies or protocol.
  • Bold and Cautious have no operational definition or manipulation check.
  • There are no paraphrases to estimate sensitivity to wording.
  • There is no full role×interaction style factorial.
  • Bold/Cautious styles are not tested on nurse or non-medical controls.
  • No seeds or repetitions per case are reported.
  • Deterministic decoding yields one observation per condition and does not estimate generative variability.
  • Exact IDs and checkpoint revisions are not immutable.
  • The study does not report the version of transformers, vLLM, CUDA, hardware, or quantization.
  • The paper says API defaults for proprietary models without enumerating them.
  • Retries, failures, missingness, or non-parseable outputs are not documented.
  • Consistency Rate excludes non-parseable outputs and does not give their counts.
  • Consistency compares logits and text, not an internal representation independent of the output.
  • Token probability depends on tokenization and space/newline variants.
  • The code aggregates several token variants via log-sum-exp, a decision not described in the paper.
  • ECE does not report number of bins, strategy, or sensitivity.
  • There are no reliability diagrams, Brier score, NLL, or per-class calibration.
  • ECE on subsets with extreme label prevalence can be misleading.
  • The abstract mixes percentages and percentage points.
  • A reduction in ECE is presented verbally as a positive gain in calibration.
  • Risk propensity is essentially the frequency of label C and overlaps with the mechanism that changes accuracy.
  • Risk sensitivity as a Type-I/Type-II ratio is unstable near a zero denominator.
  • No intervals or tests are given for the risk metrics.
  • A↔B errors do not enter the binary definition of over/under-triage.
  • The relative severity of errors is not weighted with clinical outcomes.
  • There is no external validation of the reference labels in this paper.
  • Transportability to other diseases or healthcare systems is not examined.
  • PatientSafetyBench measures responses to adversarial prompts, not behavior with real patients.
  • The additional sentence about lack of access and advice from a friend alters each original query.
  • Only HuatuoGPT 70B/72B appear in the qualitative safety analysis by category.
  • The LLM judges share families/models with the study targets.
  • Self-preference or judge family bias is not analyzed.
  • GPT-5 has no snapshot or reproducible parameters.
  • Agreement among LLM judges is low at the case level.
  • MRR is ordinal and bounded, but is analyzed with a paired t-test without robustness.
  • There is no correction for the numerous comparisons.
  • The figures do not identify all exact comparisons/p-values.
  • Error bars are not defined in captions.
  • Reasoning judges are instructed not to infer ground truth.
  • Perceived rankings do not measure clinical correctness.
  • Humans only compare two responses while LLMs rank four.
  • The human and LLM tasks are not psychometrically equivalent.
  • The 50 human cases are selected by unanimous consensus of the judges that are meant to be validated.
  • Extreme sampling introduces selection/incorporation bias.
  • The sample is forced to 25 medical-preferred and 25 non-medical-preferred cases.
  • Human preferences do not estimate natural prevalence in the corpus.
  • The three human annotators are coauthors and volunteers.
  • An independent sample of clinicians is not described.
  • One of three annotators is a recent graduate, not a senior clinician.
  • Exact specialty per annotator is not reported.
  • Training, a complete guidelines document, or adjudication is not reported.
  • The study does not publish the 299 human annotations.
  • The κ=.43 is not the average among the three clinicians.
  • Two of three clinician–clinician kappas are negative.
  • The visible clinician–clinician average is approximately .09.
  • κ against the LLM is calculated only on 16 safety cases filtered by confidence.
  • No confidence intervals for kappa are given.
  • Confidence filtering is post-hoc and may bias agreement.
  • 95.9% low confidence has no denominator/criterion compatible with the published histogram.
  • Reasoning agreement cannot be calculated, precisely the dimension said to be validated to a lesser degree.
  • The repository does not include requirements, a lockfile, or installation instructions.
  • The repository has no license.
  • The repository has no tests, CI, or quality checks.
  • The README consists of a single sentence.
  • No datasets, schemas, synthetic fixtures, or DUA instructions are included.
  • No reproducible outputs, rankings, annotations, tables, or figures are included.
  • No statistical analysis code is included.
  • McNemar, paired t-test, aggregated ECE, MRR, risk metrics, or kappa are not implemented.
  • infer_safety.py hardcodes ED Nurse and processes only five records.
  • run_triage_med_gemma.py processes only the first five records.
  • run_triage_reasoning_model.py processes only the last five records.
  • The LLM evaluators process only the first three records.
  • The data defaults point to absent files.
  • The reasoning script uses absolute GPFS paths from the authors' environment.
  • The reasoning script contains repeated imports and duplicated code.
  • There is no client or script for the GPT-5 judge.
  • There is no script that builds the ranking prompts described in the appendix.
  • The MedGemma prompt in the code requires a format different from the paper.
  • The triage evaluator requests dimensions not defined in the published protocol.
  • The parser accepts aliases even though the prompt requires a single letter.
  • Parsing rules may confuse incidental language with the label.
  • The safety script concatenates text without a space after the original query.
  • Syntactic compilation does not verify execution, results, or fidelity to the paper.
  • Restricted clinical data prevents complete independent replication.
  • There is no prospective evaluation, patient outcome, or real-harm analysis.
  • There is no comparison with a clinical decision rule or human professionals in triage.
  • Fairness, language, demographics, or access inequality are not studied.
  • Deployment, monitoring, or recovery from failure is not tested.

What the study does not establish

  • It does not establish that a medical persona grants expertise to the model.
  • It does not demonstrate clinical improvement outside the studied benchmarks.
  • It does not separate the causal effect of the role from the shift of the prior toward emergencies.
  • It does not demonstrate that the deterioration in primary care is inherent to all medical personas.
  • It does not demonstrate that a primary care persona would worsen primary care.
  • It does not demonstrate calibration of confidence communicated to patients or clinicians.
  • It does not demonstrate access to an internal preference independent of the output.
  • It does not demonstrate that Bold or Cautious control risk reliably.
  • It does not demonstrate factual safety or real outcomes in patients.
  • It does not demonstrate moderate agreement among the three clinicians.
  • It does not validate the LLM judges on the full corpus.
  • It does not validate reasoning quality via calculable human agreement.
  • It does not test generalization to other models, roles, countries, or specialties.
  • It does not allow reproduction of the figures from the public repository.
  • It does not offer a clinical system ready or authorized for deployment.

Traceability

Scope: Full text

Version: arXiv:2601.05376v1, submitted 8 January 2026, 14 pages; official code repository audited at commit 19a7487df0cde158c738bdf97d51d032becb0131

Consulted source: https://arxiv.org/pdf/2601.05376v1

Review: Codex full-text bilingual-fidelity, all-page visual, arXiv-source, current-publication-status, clinical-construct, dataset-label-confounding, calibration, risk-metric, LLM-judge, human-selection, kappa-recalculation, figure-consistency, ethics, code-repository, syntax, reproducibility and deployment-claim audit; summaries written from complete sources rather than abstract keyword extraction, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • FreedomIntelligence HuatuoGPT-o1-8B with Llama-3.1-8B backbone
  • FreedomIntelligence HuatuoGPT-o1-70B with Llama-3.1-70B backbone
  • FreedomIntelligence HuatuoGPT-o1-7B with Qwen2.5-7B backbone
  • FreedomIntelligence HuatuoGPT-o1-72B with Qwen2.5-72B backbone
  • Google MedGemma-27B text model
  • OpenAI GPT-5 as one of three LLM judges; exact snapshot and API parameters not reported
  • HuatuoGPT-o1-70B and HuatuoGPT-o1-72B as open-weight LLM judges

Instruments and metrics

  • One-sentence system persona template: You are a {persona}.
  • Emergency Department Physician and Emergency Department Nurse roles
  • Bold and Cautious Emergency Department Physician style variants
  • Helpful Assistant and empty-system-prompt controls
  • Three-way clinical triage labels A stay home, B routine care, C emergency care
  • Accuracy and McNemar paired tests with continuity correction
  • Expected Calibration Error from conditional A/B/C token likelihoods
  • Consistency Rate between highest-likelihood and parsed generated labels
  • Risk propensity as frequency of emergency labels
  • Risk sensitivity as Type-I over-triage count divided by Type-II under-triage count
  • LLM-judge rankings summarized by Mean Reciprocal Rank
  • Paired t-tests over MRR
  • Argilla pairwise clinician preference and 0–100 confidence annotation
  • Pairwise Cohen kappa on a confidence-filtered safety subset

Data used

  • 1,466 de-identified emergency-observation-unit patients with suspected TIA or stroke from 2013–2020
  • 201 lower-acuity symptom-based cases attributed to Fraser et al. 2023
  • PatientSafetyBench, 466 open-ended queries across five medical-safety categories
  • Human-evaluation reasoning subset: 50 LLM-consensus-selected cases and 149 returned clinician judgments
  • Human-evaluation safety subset: 50 LLM-consensus-selected cases and 150 returned clinician judgments
  • Clinical triage records unavailable publicly; controlled access proposed under a Data Use Agreement
  • Official GitHub repository rsinghlab/Persona_Paradox at commit 19a7487, without data, outputs, analysis, environment, or results

Evidence and location

  • Version, authors, abstract, and preprint status: arXiv:2601.05376v1 abstract page and complete 14-page PDF, checked 15 July 2026
  • Personas and one-sentence conditioning: Sections 3.1–3.2 and Appendices B–C
  • Accuracy, risk, consistency, and ECE metrics: Section 3.3, pp. 3–4
  • Triage datasets, models, and judges: Section 4, pp. 5–6
  • Changes in accuracy, consistency, and calibration: Figure 2 and Section 5.1, pp. 5–6
  • Bold/Cautious interaction and risk ratios: Figure 3 and Section 5.2, pp. 6–7
  • Low agreement among LLM judges: Section 5.3, pp. 7–8
  • Human design and LLM consensus selection: Sections 3.4.2 and 5.4; Appendix E
  • Kappa discrepancy: Figure 5b: clinician pairwise kappas versus clinician-to-LLM kappas, p. 8
  • Human counts and coauthorship condition: Appendix E, p. 13
  • Restricted data and ethical considerations: Sections 7–8, p. 9
  • Code status and reproducibility defects: Official GitHub rsinghlab/Persona_Paradox commit 19a7487; all five Python files and README audited 15 July 2026
  • Complete visual inspection: All 14 PDF pages and all principal source-package figures rendered and visually inspected on 15 July 2026