Roleplay-doh: Enabling Domain-Experts to Create LLM-simulated Patients via Eliciting and Adhering to Principles

Personas, identity, and agents2024ACL AnthologyApproved editorial review

Authors: Ryan Louie, Ananjan Nandi, William Fang, Cheng Chang, Emma Brunskill, Diyi Yang

Keywords: LLM-simulated patients, Expert-defined principles, Human-LLM collaboration, Mental health training, Role-playing agents

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

6
Authors
8
Findings
22
Limitations
6
Evidence

Editorial summary

English

Roleplay-doh is a tool for a mental-health-support expert to construct a simulated patient from a remembered case, converse with it, and convert kudos, critiques or rewritten responses into behavioral principles. GPT-3.5 turns kudos and critiques into rules; GPT-4 explains a rewrite and derives the rule. For every dialogue turn, gpt-4-turbo-1106 receives the scenario, principles and history. A second pipeline decomposes principles into yes/no questions, adds general dialogue criteria, decides which criteria apply, evaluates the initial response and rewrites it if any applicable criterion fails. The pilot involved six 7 Cups counselors; four coauthors then conversed with four selected patients and rated 276 responses, with only one rater per response. Fifty-five, or 20%, received moderate or low satisfaction, but this estimate has no inter-rater agreement or independent expert adjudication. The main within-subject study included 25 US professionals or counselors. Everyone first held a ten-minute Scenario-Only conversation and then a thirty-minute session about the same remembered case while adding rules. They created 25 patient pairs and 123 principles. Creators rated the principle version higher on five of six scales: increases were 0.80 for authenticity, 0.76 for resemblance to the remembered case, 1.00 for challenging aspects, 0.64 for training readiness and 0.52 for recommendation; staying in role barely changed, by 0.08. The design was not counterbalanced, however: order, practice, session length and tool use changed together, while creators knew the condition and judged their own rules. Five counselors drawn from the same creator cohort rated randomized, blinded transcripts for all 25 pairs, producing 125 comparisons. A mixed model, Rating~Treatment+CreatorID+(1|AnnotatorID), estimated smaller increases for authenticity (+0.31), typical-case resemblance (+0.49), readiness (+0.39) and recommendation (+0.38); role (+0.09) and counselor challenge (+0.22) were not significant. Agreement was poor: Krippendorff alpha was 0.023-0.082 for rating differences and 0.046-0.232 for preference. The technical evaluation compared five variants with three expert counselors. On 40 selected failures, Full won, tied and lost against the initial response in 30%, 55% and 15% of overall judgments. That 30% is a win proportion, not a 30% improvement on a quality scale. On 50 random turns the result was 18%/78%/4%, and in 34 of 50 turns every method returned exactly the same response. The work therefore supports that principles are authorable and that this small sample perceived modest improvements in some transcripts; it does not establish novice learning, clinical validity, safety or transfer to real patients. Scenarios came from counselors' memories of real patients, but the paper does not document patient consent, de-identification or a privacy review. Although the project page promises code and data, its only official repository contains the static website and the Code button is commented out. The principle dataset is on Hugging Face under CC BY-NC 4.0 but requires authentication and contact sharing; the application, analysis, raw study data and model outputs needed for end-to-end reproduction are not public.

Español

Roleplay-doh es una herramienta para que una persona experta en apoyo psicológico construya un paciente simulado a partir de un caso que recuerda, converse con él y convierta kudos, críticas o reescrituras en principios de comportamiento. GPT-3.5 transforma kudos y críticas en reglas; GPT-4 explica una reescritura y deriva la regla. Para generar cada turno, gpt-4-turbo-1106 recibe escenario, principios e historial. Una segunda tubería descompone los principios en preguntas de sí/no, añade criterios generales de diálogo, decide cuáles aplican, evalúa la respuesta inicial y la reescribe si algún criterio aplicable falla. El piloto reunió seis consejeros de 7 Cups; después cuatro coautores conversaron con cuatro pacientes seleccionados y calificaron 276 respuestas, una sola persona por respuesta. Cincuenta y cinco, el 20%, recibieron satisfacción moderada o baja, pero esta estimación no tiene acuerdo entre jueces ni evaluación experta independiente. El estudio principal, intra-sujeto, incluyó 25 profesionales o consejeros estadounidenses. Todos hicieron primero una conversación de diez minutos con Scenario-Only y luego una sesión de treinta minutos sobre el mismo caso, añadiendo reglas. Crearon 25 pares de pacientes y 123 principios. Quienes crearon los pacientes puntuaron más alto la versión con principios en cinco de seis escalas: los incrementos fueron 0,80 en autenticidad, 0,76 en parecido al caso recordado, 1,00 en aspectos desafiantes, 0,64 en preparación para entrenar y 0,52 en recomendación; mantenerse en el rol apenas cambió, +0,08. Sin embargo, no hubo contrabalanceo: orden, práctica, duración y uso de la herramienta cambiaron a la vez, y las personas conocían la condición y juzgaban sus propias reglas. Cinco consejeros tomados del mismo grupo de creadores evaluaron, a ciegas y en orden aleatorio, las transcripciones de los 25 pares, 125 comparaciones. Un modelo mixto, Rating~Treatment+CreatorID+(1|AnnotatorID), estimó incrementos menores en autenticidad (+0,31), parecido a casos típicos (+0,49), preparación (+0,39) y recomendación (+0,38); rol (+0,09) y dificultad para el consejero (+0,22) no fueron significativos. El desacuerdo fue alto: alpha de Krippendorff de 0,023-0,082 para diferencias de puntuación y 0,046-0,232 para preferencia. La evaluación técnica comparó cinco variantes con tres consejeros. En 40 fallos seleccionados, la tubería completa ganó, empató y perdió frente a la respuesta inicial en 30%, 55% y 15% de los casos en la valoración global. Ese 30% es una proporción de victorias, no una mejora del 30% en una métrica de calidad. En 50 turnos aleatorios el resultado fue 18%/78%/4%, y en 34 de 50 todas las variantes produjeron exactamente la misma respuesta. Por tanto, el trabajo respalda que las reglas son autorables y que esta pequeña muestra percibió mejoras modestas en ciertas transcripciones; no demuestra aprendizaje de novatos, validez clínica, seguridad ni transferencia a pacientes reales. Los casos procedían de recuerdos de pacientes reales, pero el paper no documenta consentimiento de esos pacientes, desidentificación o revisión de privacidad. La web promete código y datos, pero su único repositorio oficial contiene solo el sitio estático y el botón de código está comentado. El dataset de principios está en Hugging Face bajo CC BY-NC 4.0, pero exige autenticarse y compartir contacto; no se publicaron aplicación, análisis, datos crudos ni outputs para reproducir el estudio de extremo a extremo.

Research question

Can a tool convert expert feedback into principles that govern LLM-simulated patients, improve adherence to those rules, and produce transcripts that counselors perceive as more authentic and useful than a scenario without principles?

Method

The work combines iterative design, a pilot, a within-subject study with 25 creators, a blind comparison of 125 pairs by five judges from the same group, and a technical evaluation by three experts. The GPT-4 pipeline generates, formulates criteria, evaluates, and, if it detects noncompliance, rewrites. Likert scales, a mixed model for external judges, win/tie/loss preferences, and Krippendorff's alpha are reported.

Sample: Six counselors from 7 Cups in the pilot; four coauthors as conversers and sole evaluators of 276 responses; 25 US experts as creators; five of those same 25 as judges of 125 pairs; and three expert counselors in the technical evaluation. Among the 25 cases, 14 had no diagnosis, 10 described symptoms or prior diagnoses, and one a severe mental illness.

Findings

  • The 25 creators produced 123 principles, between 1 and 10 per patient, median 5; 11.4% had to be manually edited.
  • The version with principles received higher creator means on five of six measures, with increases of 0.52 to 1.00 points; staying in role changed only +0.08.
  • The five blind judges estimated increases of +0.31 in authenticity, +0.49 in similarity, +0.39 in preparation, and +0.38 in recommendation; two measures did not change significantly.
  • Agreement among judges was very low: alpha 0.023-0.082 for score differences and 0.046-0.232 for preferences.
  • In 40 selected errors, Full obtained 30% wins, 55% ties, and 15% overall losses against No Critique.
  • In 50 random turns, Full obtained 18% wins, 78% ties, and 4% losses; 34 of 50 gave the same output in all variants.
  • Scenario-Only and principle-based conversations had 13.04 and 13.64 mean turns, but outputs of 177.29 and 120.43 tokens respectively.
  • Self-perceived usability was high, although some people described difficulty in formulating, generalizing, merging, or enforcing rules.

Limitations

  • The creator study had a fixed order: Scenario-Only always before Scenario+ExpertPrinciples.
  • The phases lasted 10 and 30 minutes and also differed in practice, familiarity, and use of the tool; the causal effect of the principles is not isolated.
  • Creators knew the condition, had invested effort in their rules, and judged similarity to a case that only they remembered.
  • The five external judges came from the same group of 25 creators, not from an independent sample.
  • There are only five annotator clusters and power was simulated with parameters estimated from early data from the same study and an assumed effect of 0.52.
  • Agreement among judges is extremely low, so mean differences do not equate to stable expert consensus.
  • Six outcomes per group are tested without reported multiple comparison correction or confidence intervals.
  • The paper does not specify the test or model used for the significance stars in the creator study.
  • The 30% is the percentage of wins in 40 cases chosen because they already contained errors, with 55% ties; it is not a percentage improvement in quality nor prevalence in normal use.
  • In the random sample, ties predominate and 68% of cases produce identical outputs in all variants.
  • The pilot figure of 20% uses selected patients, coauthor conversations, and a single evaluator per response.
  • GPT-4 generates, creates criteria, evaluates, and rewrites; an independent model judge is not tested.
  • No latency, cost, total tokens, retries, JSON failures, or operational load from the multiple calls are reported.
  • The system assumes internal consistency of the principles and does not demonstrate automatic detection or resolution of conflicts.
  • The cases recall real patients, but patient consent, deidentification, redaction, minimization, or reidentification risk assessment is not described.
  • Only perceptions about transcripts are measured, not performance or learning of novice counselors.
  • Clinical correctness, crisis, self-harm, bias, safety, patient outcomes, or harm in deployment are not evaluated.
  • There is only textual dialogue and a brief session; nonverbal cues, longitudinal consistency, and follow-up are missing.
  • The sample is US-based and contains only one case of severe mental illness.
  • Dependence on gpt-4-turbo-1106 limits temporal reproduction and generalization to current or open models.
  • The dataset is gated and there is no application code, raw data, outputs, or public analyses to reproduce end to end.
  • The paper, the dataset, the website, and the phrase 'research use only' present distinct licenses and there is no pipeline code to which to apply a clear license.

What the study does not establish

  • It does not demonstrate that novice counselors learn better or transfer skills to real patients.
  • It does not clinically validate the simulated patients or their suitability for unsupervised training.
  • It does not causally isolate the effect of the principles from order, time, practice, and effort.
  • It does not demonstrate strong expert consensus on authenticity.
  • It does not demonstrate a 30% quality improvement in ordinary use.
  • It does not prevent harmful, absurd, or derogatory outputs.
  • It does not test privacy or consent of the recalled real patients.
  • It does not generalize to severe illness, other cultures, other domains, multimodality, or longitudinal sessions.
  • It does not allow integral reproduction with the public artifacts.
  • It does not test the result with current, open, or non-OpenAI models.

Traceability

Scope: Full text

Version: EMNLP 2024, pages 10570-10603; arXiv:2407.00870v2, project repository and gated dataset audited separately

Consulted source: https://aclanthology.org/2024.emnlp-main.591/

Review: Codex 34-page visual, official-ACL, arXiv-v2, full-method, prompt, participant-flow, fixed-order, mixed-model, power, agreement, selected-vs-random-case, privacy, artifact, license, reproducibility and claim-boundary audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • gpt-4-turbo-1106
  • GPT-3.5, snapshot not specified for principle elicitation

Instruments and metrics

  • Roleplay-doh expert-feedback interface
  • Six seven-point creator patient-rating items
  • Four seven-point tool-experience items
  • Six seven-point third-party transcript-rating items
  • Linear mixed-effects model: Rating~Treatment+CreatorID+(1|AnnotatorID)
  • Simulation-based power analysis with R simr over 300 trials
  • Win/tie/loss majority vote from three expert rankings
  • Krippendorff alpha for rating differences, preferences and ranked outputs

Data used

  • Twenty-five remembered real-patient scenarios and 25 paired simulated patients
  • 123 expert-defined principles, gated at SALT-NLP/roleplay-doh_principles
  • Sixteen pilot conversations with 276 singly rated responses
  • 125 blinded third-party patient-pair comparisons
  • Forty selected error turns and fifty random turns for pipeline evaluation

Evidence and location

  • Metadata, DOI, pages, authors, license, and exact abstract: ACL Anthology 2024.emnlp-main.591
  • System, prompts, participants, tables, statistics, limitations, ethics, and appendices: EMNLP 2024 PDF, 34 pages, sha256 3dfcae6a4597fa50ebe8fff89095aea6bf63a48ad7ecd0c63dc8067381059feb
  • History and version of the preprint: arXiv:2407.00870v2
  • Actual availability of code and repository content: roleplay-doh/roleplay-doh.github.io at commit d87eae73826c371df47ee5136eeb0f37c3dcb468; checked 2026-07-16
  • Gate, files, license, and access of the dataset: SALT-NLP/roleplay-doh_principles metadata and unauthenticated 401 GatedRepo response; checked 2026-07-16
  • Audit of design, statistics, privacy, artifacts, and claim boundaries: reports/verification/article-247-emnlp-roleplay-doh-study-design-statistics-privacy-artifact-and-claim-audit.json