Synthetic Personality Review

The Art of Midwifery in LLMs: Optimizing Role Personas for Large Language Models as Moral Assistants

Personas, identity, and agents2026arXivApproved editorial review

Authors: Yangyi Wu, Tianqi Wang, Xilin Liu

Keywords: Persona conditioning

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

3
Authors
8
Findings
30
Limitations
4
Evidence

Editorial summary

English

Wu, Wang, and Liu propose that an LLM should act as a moral assistant that facilitates reflection rather than a judge that replaces the person. Across five models, they compare four persona prompts: Socratic, Guardian Angel, Rational Counselor, and Virtue Exemplar. Each condition answers six dilemmas in three turns. Two coders rate text on autonomy, cognitive scaffolding, emotion recognition, value neutrality, constructive challenge, and relationship building. The study defines HSI as the mean of those six dimensions, Balance as a penalty for uneven profiles, and Final Score as their product. Aggregates mirror the instructions: Socratic is strongest in autonomy, questioning, and neutrality; Guardian in emotion and relationship; Rational is low on affect; Virtue is more even. Virtue has the highest Final Score, 5.363, although its HSI of 5.639 is below Socratic's 5.950 and Guardian's 5.917: it wins through the authors' chosen balance multiplier. This does not demonstrate effective moral assistance. No user participates and no reflection, autonomy, decision, wellbeing, or moral growth is measured; coders score textual conformity to a study-defined rubric. The Virtue prompt also explicitly combines the capabilities later rewarded equally, making part of its superiority circular. The paper calls 360 turns independent even though they are nested within 120 dialogues and five models. Its equations include only six units per model-persona, apparently scenarios, and omit the three-turn index without explanation; promised Persona-by-Scenario interactions and turn trajectories are not reported. It claims significant and context-specific advantages in bioethical or existential cases without tests, intervals, or scenario-level results. The kappa table is organised by persona while the text says by dimension; only about 20% appears double-coded after calibration, and the principal researcher scores the other 80%. Complete prompts, transcripts, raw data, and code are absent. 'Constructive Divergence' is a useful anti-accommodation design idea, but the 60% Virtue weight and crisis-keyword switching are unevaluated recommendations. This is a preliminary normative taxonomy of prompting styles, not validation of a safe moral assistant or an optimal personality.

Español

Wu, Wang y Liu proponen que un LLM actúe como asistente moral que facilita reflexión, no como juez que sustituye a la persona. Comparan en cinco modelos cuatro prompts-persona: Socrático, Ángel Guardián, Consejero Racional y Ejemplo de Virtud. Cada condición responde a seis dilemas en tres turnos. Dos codificadores valoran texto en autonomía, andamiaje cognitivo, reconocimiento emocional, neutralidad, desafío constructivo y relación. El trabajo introduce HSI como media de esas seis dimensiones, Balance como penalización de perfiles desiguales y Final Score como su producto. Los agregados reflejan las instrucciones: el Socrático destaca en autonomía, preguntas y neutralidad; el Guardián en emoción y vínculo; el Racional queda bajo en afecto; el Ejemplo de Virtud resulta más uniforme. Este último obtiene el mayor Final Score, 5,363, aunque su HSI 5,639 es inferior al Socrático 5,950 y al Guardián 5,917: gana por el multiplicador de balance elegido por los autores. La evidencia no demuestra asistencia moral eficaz. No interviene ningún usuario ni se miden reflexión, autonomía, decisión, bienestar o crecimiento moral; se puntúa conformidad de textos con una rúbrica creada por el estudio. Además, el prompt de Virtud ordena combinar precisamente las capacidades que después se premian por igual, por lo que parte de su superioridad es circular. El paper llama independientes a 360 turnos, aunque están anidados en 120 diálogos y cinco modelos. Sus fórmulas solo incluyen seis unidades por modelo-persona, aparentemente escenarios, y eliminan sin explicación el índice de tres turnos; tampoco presenta las interacciones Persona×Escenario ni trayectorias temporales prometidas. Afirma diferencias significativas y ventajas en bioética o crisis existenciales sin tests, intervalos ni resultados por escenario. La tabla de kappa está desglosada por persona mientras el texto dice por dimensión; solo el 20% parece doble codificado tras calibración y el 80% lo puntúa el investigador principal. No se publican prompts completos, transcripciones, datos ni código. La propuesta de “divergencia constructiva” es una idea útil contra la complacencia, pero el 60% de peso para Virtud y el cambio por palabras de crisis son recomendaciones no evaluadas. Es una taxonomía normativa preliminar de estilos de prompt, no validación de un asistente moral seguro o de una personalidad óptima.

Research question

Which persona, Socratic, Guardian Angel, Rational Counselor, or Virtue Exemplar, produces texts that best fit a moral assistance rubric, and can 'constructive divergence' guide an assistant that questions without replacing the user's autonomy?

Method

Crossed design with five models, four system-personas, six moral scenarios, and three turns per dialogue. The authors declare 120 dialogues and 360 responses. Two coders score from 1 to 7 on six dimensions; they jointly calibrate approximately 20% and the principal investigator codes the remaining 80%. HSI averages dimensions, Balance penalizes its relative deviation, and Final Score multiplies both. Correlations and PCA on aggregated profiles are added.

Sample: Five model aliases by four personas and six scenarios produce 120 dialogues; three responses per dialogue yield 360 turns. These are not independent observations: they are nested by dialogue, persona, and model. The equations use only six values per model-persona and 30 per persona-dimension, with no index for the three turns nor explanation of their aggregation. Two coders work on approximately 20%; the principal investigator codes the remainder alone.

Findings

  • Socratic/Sage reports HSI 5.950, Balance 0.773, and Final 4.599; it is the highest HSI without penalization.
  • Guardian reports HSI 5.917, Balance 0.822, and Final 4.864, with high emotion and relationship scores.
  • Rational reports HSI 4.022, Balance 0.777, and Final 3.125, low especially on affective dimensions that its prompt does not prioritize.
  • Virtuous reports HSI 5.639, Balance 0.951, and Final 5.363; it ranks first only after the balance multiplier.
  • The aggregated profiles reproduce the capabilities and deficiencies explicitly specified in each persona.
  • Emotional recognition and relationship building correlate r=0.951, a sign of strong overlap between dimensions.
  • The paper claims advantages of the Guardian in bioethics and of the Socratic in existential dilemmas, but shows no data or tests per scenario.
  • It proposes a Virtue baseline at 60% and Guardian/Socratic/Rational overlays, without evaluating that architecture.

Limitations

  • There are no user participants or pre/post measures; no moral growth, reflection, autonomy, decision, or well-being is observed.
  • The rubric measures text valuation by the authors, not external efficacy, safety, or utility for a person.
  • The Virtue prompt explicitly combines the six rewarded capabilities, creating circularity between intervention and outcome.
  • The weaknesses of Guardian, Rational, and Socratic are largely designed by their instructions, not independently discovered.
  • There is no baseline without persona, control for prompt length/content, alternative specification, or shuffled rubric.
  • HSI weights six dimensions equally without validation; Balance rewards uniformity and Final multiplies both arbitrarily.
  • Virtue does not have the highest HSI; its victory depends on the balance multiplier.
  • The 360 turns are not independent and the statistical unit changes between turn, dialogue, model-persona profile, and persona aggregate.
  • The equations use k=1..6 as 'rounds' although there are six scenarios by three turns; the turn index disappears.
  • No Persona×Scenario analysis, repeated measures, T1-T3 trajectory, or resistance variant comparison is reported.
  • Significant is used without p-values, intervals, tests, effects, or multiplicity correction.
  • Correlations may be induced by four prompt clusters and by conceptually overlapping dimensions.
  • PCA on the same scores does not independently validate the personas; explained variance, scaling, loadings, and stability are missing.
  • Figure 4 says that box width indicates consistency, but in the vertical boxplot dispersion is in IQR and whiskers, not in fixed width.
  • The kappa table is per persona while the text asserts calculation per six dimensions.
  • Only approximately 20% appears double coded; the remaining 80% lacks inter-judge control.
  • It is unclear whether kappa is calculated before or after calibrating to satisfactory agreement or whether it uses ordinal weighting.
  • Kappa lacks denominators, matrices, intervals, missingness, and adjudication rules.
  • The principal investigator knows hypotheses and rubric; label blinding does not hide highly recognizable linguistic styles.
  • Model names do not include endpoint, immutable snapshot, date, or exact provider.
  • Temperature, top-p, seed, tokens, serialization, context reset, safety, retries, refusals, and exclusions are missing.
  • Scenarios are presented sequentially without clarifying reset or counterbalancing, with possible carryover.
  • Complete prompts, transcripts, scores, figure data, code, environment, or supplement are not published.
  • The claim that traditional alignment equals imitating human values is not tested against an aligned baseline.
  • The neutrality/autonomy of the rubric comes into tension with T3, which forces the assistant to choose for the user.
  • The Down syndrome dilemma uses a controversial moral framework without documented review by disability or clinical community.
  • The organ dilemma uses age, unemployment, alcohol, and philanthropy without medical criteria, with risk of validating discrimination by social value.
  • The keyword-based crisis recommendation has no clinical evaluation, red team, false negatives, or referral validation.
  • Consent, compensation, protection, or ethical review of human coders is not reported.
  • There is only a preprint arXiv v1 and no public artifact of the study was located.

What the study does not establish

  • It does not demonstrate that any persona improves the reasoning or moral growth of real users.
  • It does not demonstrate that Virtue Exemplar is objectively the best persona; it wins a metric designed to reward balance.
  • It does not demonstrate specific efficacy in bioethics, existential crises, whistleblowing, or organ allocation.
  • It does not support statistical inference with 360 independent observations.
  • It does not validate the six constructs or their equal weighting as a standard of moral assistance.
  • It does not convert correlation or PCA of prompt-induced ratings into distinct psychological personalities.
  • It does not prove that textual divergence is constructive, correct, safe, or accepted by the user.
  • It does not validate the 60/40 dynamic mechanism or the crisis switching it recommends.
  • It does not demonstrate safety for disability, medicine, mental health, or high-stakes moral decisions.
  • It does not allow reproducing results or auditing outputs due to lack of prompts, transcripts, data, and code.

Traceability

Scope: Full text

Version: arXiv:2603.20626v1, submitted 2026-03-21, CC BY 4.0

Consulted source: https://arxiv.org/abs/2603.20626

Review: Codex 7-page visual full-text, construct, prompt-rubric circularity, sample-accounting, coder-reliability, statistical, safety, ethics and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Kimi-2.5
  • DeepSeek-V3
  • GPT-5
  • Claude-4.5-Sonnet
  • Gemini-3

Instruments and metrics

  • Socratic Persona system prompt
  • Guardian Angel Persona system prompt
  • Rational Counselor Persona system prompt
  • Virtue Exemplar Persona system prompt
  • Six three-turn moral scenarios
  • 1-7 Autonomy Support rating
  • 1-7 Cognitive Scaffolding rating
  • 1-7 Emotion Recognition rating
  • 1-7 Value Neutrality rating
  • 1-7 Constructive Challenge rating
  • 1-7 Relationship Building rating
  • Cohen kappa
  • Helper Suitability Index
  • Balance Score
  • Final Score
  • Pearson correlation and PCA

Data used

  • 120 reported model-persona-scenario dialogue sequences, not released
  • 360 reported assistant turns, not released
  • Six-dimensional human coding sheet, not released
  • Approximately 20% calibration/double-coded subset, not released

Evidence and location

  • Design, scenarios, personas, rubric, formulas, tables, figures, discussion, and recommendations: arXiv:2603.20626v1, 7/7 pages rendered and individually inspected
  • Metadata, single version v1, date, DOI, and CC BY 4.0 license: Official arXiv abstract and Atom records inspected 2026-07-17
  • Current absence of locatable public artifacts: Paper, official arXiv record and exact-title/arXiv-ID public web and GitHub repository searches on 2026-07-17
  • Audit of circularity, denominators, statistics, reliability, safety, and reproducibility: reports/verification/article-387-moral-assistant-persona-rubric-circularity-sample-accounting-statistics-safety-and-reproducibility-audit.json