Moral Susceptibility and Robustness under Persona Role-Play in Large Language Models

Applications, bias, and safety2025arXivApproved editorial review

Authors: Davi Bastos Costa, Felippe Alves, Renato Vicente

Keywords: Persona Role-Play, Moral Foundations Questionnaire, Moral Robustness, Moral Susceptibility, LLM Evaluation

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

3
Authors
6
Findings
8
Limitations
5
Evidence

Editorial summary

English

This preprint proposes two indices for describing how an LLM's responses vary when it role-plays synthetic personas on the Moral Foundations Questionnaire (MFQ). Moral robustness R is the inverse of the mean standard deviation from repeating the same question with the same persona. It measures response repeatability, not correctness, moral quality, or safety. Moral susceptibility S averages, across questions, the dispersion of persona-level means. It measures how much scores move across this prompt set, not whether a model represents a person faithfully. Fifteen models from six families, Claude, DeepSeek, Gemini, GPT, Grok, and Llama, are tested with 100 descriptions, 30 items, and ten repetitions at temperature 0.1. This is 30,000 primary calls per model and 450,000 overall, plus no-persona baselines, logit collection, and temperature sweeps. The 100 personas are neither a human nor a balanced sample. They are the first 100 entries in a 1,000-description file produced by a seed-42 PersonaHub shuffle. The file mixes occupations, locations, fandoms, and political or moral cues without attribute coding or persona ground truth. Reported results show a very large separation in R. Claude Sonnet and Haiku reach about 107.70 and 92.04, compared with 27.73 for Gemini Flash Lite, 9.96 for Gemini Flash, 9.75–14.48 for the five GPT-4/4.1 variants, and roughly 3.27–4.59 for DeepSeek, Grok, and Llama. S covers a much narrower range: Gemini Flash is highest at 1.043 and GPT-4o Mini lowest at 0.663. Relative across-model variation is 152% for R and 13% for S. A permutation ANOVA on log R, using only 15 observations grouped into six providers, reports eta-squared 0.963 and p<0.00002; the matching S analysis gives eta-squared 0.539 and p=0.137, not conventionally significant. The proposed size trend for S uses questionable ordinal ranks, including V3 to V3.1 as a size progression and Fast/full product labels, and yields a slope of 0.056±0.035 with p=0.150. The evidence therefore supports provider clustering in repeatability under this setup, but does not show that post-training causes R or pretraining determines S. There are no controlled pre/post-training variants, and family is confounded with architecture, API, decoding, scale, and corpus. R and S are effectively orthogonal across the 15 models, r=-0.03 and p=0.91. Even the strongest foundation-level association, Purity/Sanctity at r=-0.49 and p=0.07, is not conventionally significant. R also grows toward infinity by construction as temperature approaches zero and answers become deterministic, limiting its interpretation as an intrinsic model property. MFQ use adds another boundary. The instrument was designed for human self-report, whereas this study presents one item at a time to a generator role-playing a synthetic description. It does not assess scale reliability, factor structure, measurement invariance across model families, criterion validity, human-model equivalence, or fidelity to an actual person. The outputs are prompt-induced text distributions, not internal moral beliefs. Prompts ask for a digit followed by reasoning, yet the runner assigns one token globally and two for Claude, so it primarily collects digits; provider and output-budget differences are also entangled with model comparisons. The audit confirms exactly 30,000 rows in each of the 15 main CSVs. Claude Haiku retains 94 complete personas after 182 invalid final ratings and Gemini Flash Lite retains 97 after 62; exclusions are model-specific, so comparisons do not always use the same persona composition. The public artifact is extensive, with raw outputs, logits, results, figures, prompts, and 43 Python files that compile. A fresh-cache recomputation reproduces 14 of the 15 main rows. Claude Haiku instead yields R=89.87±10.30 rather than 92.04±10.72. Its tracked result says one retained cell has a single run, while the current CSV has ten in every retained cell. The paper reports 364 total failures across 344 affected Claude Haiku rows, but the CSV sums to 546. These discrepancies indicate drift between repaired data and cached results. The current repository also mixes July 2026 rebuttal experiments added after the v3 PDF, has no CI or tests, and claims MIT licensing without tracking a LICENSE file. The study provides an inspectable benchmark of rating consistency and sensitivity under this protocol. It does not establish morality, personality, reasoning quality, role-play competence, training-stage causes, or generalization to other prompts, instruments, personas, languages, or deployments.

Español

Este preprint propone dos índices para describir cómo varían las respuestas de un LLM al representar personas sintéticas en el Moral Foundations Questionnaire (MFQ). La robustez moral R es el inverso de la desviación típica media obtenida al repetir la misma pregunta con la misma persona: cuantifica repetibilidad de la respuesta, no corrección, calidad moral ni seguridad. La susceptibilidad S promedia, entre preguntas, la dispersión de las medias de distintas personas: cuantifica cuánto cambia la escala de respuestas en este conjunto de prompts, no si el modelo representa fielmente a una persona. Se evalúan 15 modelos de seis familias, Claude, DeepSeek, Gemini, GPT, Grok y Llama, con 100 descripciones, 30 ítems y diez repeticiones a temperatura 0,1. Son 30.000 llamadas principales por modelo y 450.000 en total, más baselines sin persona, logits y barridos de temperatura. Las 100 personas no forman una muestra humana ni equilibrada: son las primeras 100 entradas de un archivo de 1.000 descripciones obtenido mediante una baraja de PersonaHub con semilla 42. El archivo mezcla profesiones, lugares, aficiones y pistas políticas o morales sin codificación de atributos ni verdad de referencia. Los resultados publicados muestran una separación muy grande en R: Claude Sonnet y Haiku alcanzan aproximadamente 107,70 y 92,04, frente a 27,73 para Gemini Flash Lite, 9,96 para Gemini Flash, entre 9,75 y 14,48 para los cinco GPT-4/4.1 y alrededor de 3,27-4,59 para DeepSeek, Grok y Llama. S ocupa un rango mucho menor: Gemini Flash es el máximo con 1,043 y GPT-4o Mini el mínimo con 0,663. La variación relativa entre modelos es 152 % para R y 13 % para S. Una ANOVA por permutación sobre log R, con solo 15 observaciones agrupadas en seis proveedores, produce eta cuadrado 0,963 y p<0,00002; para S produce eta cuadrado 0,539 y p=0,137, sin significación convencional. El supuesto efecto de tamaño sobre S usa rangos ordinales discutibles, incluye V3 a V3.1 como progresión de tamaño y etiquetas Fast/full, y obtiene pendiente 0,056±0,035, p=0,150. Por tanto, los datos respaldan agrupamiento por proveedor en la repetibilidad, pero no prueban que el postentrenamiento cause R ni que el preentrenamiento determine S. No se comparan variantes controladas antes/después del entrenamiento y familia queda confundida con arquitectura, API, decodificación, escala y corpus. R y S son prácticamente ortogonales entre los 15 modelos, r=-0,03 y p=0,91; incluso la asociación más fuerte por fundamento, Pureza/Santidad, r=-0,49 y p=0,07, no alcanza significación. Además, R crece por construcción hacia infinito cuando la temperatura tiende a cero y las respuestas se vuelven deterministas, lo que limita su interpretación como propiedad intrínseca del modelo. El uso del MFQ añade otra frontera. El instrumento fue creado para autoinforme humano, pero aquí cada ítem se presenta por separado a un generador que representa una descripción sintética. No se comprueban fiabilidad de escalas, estructura factorial, invariancia entre familias, validez de criterio, equivalencia humano-modelo ni fidelidad a una persona real. Las salidas son distribuciones de texto inducidas por prompts, no creencias morales internas. El prompt pide un dígito seguido de razonamiento, pero el runner asigna un token en general y dos a Claude, de modo que principalmente se recogen dígitos; las diferencias de presupuesto y proveedor también quedan entrelazadas con el modelo. La auditoría de los 15 CSV confirma exactamente 30.000 filas por modelo. Claude Haiku conserva 94 personas completas tras 182 ratings inválidos y Gemini Flash Lite 97 tras 62; la exclusión se hace por modelo, así que las comparaciones no usan siempre la misma composición de personas. El artefacto público es amplio: incluye datos brutos, logits, resultados, figuras, prompts y 43 archivos Python que compilan. Una recomputación con caché limpia reproduce 14 de 15 filas principales. Claude Haiku, sin embargo, da R=89,87±10,30 en vez de 92,04±10,72; el resultado publicado dice que alguna celda tiene una sola ejecución, mientras el CSV actual tiene diez en todas las celdas retenidas. La tabla declara 364 fallos totales para sus 344 filas afectadas, pero el CSV suma 546. Esto indica deriva entre datos reparados y resultados cacheados. El repositorio actual mezcla además experimentos de réplica de julio de 2026 posteriores al PDF v3, carece de CI y tests y afirma licencia MIT sin publicar un archivo LICENSE. El trabajo ofrece un benchmark inspeccionable de consistencia y sensibilidad de ratings bajo este protocolo. No demuestra moralidad, personalidad, calidad de razonamiento, competencia de role-play, causas de entrenamiento ni generalización a otros prompts, instrumentos, personas, idiomas o despliegues.

Research question

How do the repeatability of MFQ ratings within a synthetic person and the dispersion of those ratings across persons vary between models, and how do both measures depend on family, variant, and temperature?

Method

Fifteen models answer the 30 items of the MFQ from 100 PersonaHub descriptions, ten times per combination at temperature 0.1. R inverts the mean standard deviation within each person-question cell and S averages the standard deviation across person means. Baselines without persona, next-token logits for five OpenAI models, and temperature sweeps are added; associations by family and ordinal rank are analyzed by means of permutations.

Sample: One hundred synthetic descriptions, first 100 of a PersonaHub file shuffled with seed 42, by 30 items and ten repetitions for 15 models: 450,000 main calls planned. After failures, Claude Haiku uses 94 complete persons, Gemini Flash Lite 97, and the remaining main analyses 100.

Findings

  • R varies approximately 33 times: Claude Sonnet reaches 107.70 and the least robust models hover around 3.3; S varies only about 1.6 times, from 0.663 to 1.043.
  • Family explains a large part of log R in these 15 observations, eta squared 0.963 and p<0.00002; the equivalent test for S is not significant, p=0.137.
  • The supposed effect of size on S is weak and not significant, slope 0.056±0.035 and p=0.150, besides using ordinal ranks that do not cleanly represent parameters.
  • R and S do not correlate at the model level, r=-0.03 and p=0.91; Purity/Sanctity reaches r=-0.49 and p=0.07.
  • A reproduction with the public CSVs agrees on 14 of 15 rows, but Claude Haiku produces R=89.87 and not 92.04; its failure count also contradicts the table.
  • Exclusions due to failures are carried out by model and change the composition of persons in Claude Haiku and Gemini Flash Lite.

Limitations

  • The MFQ is not psychometrically validated for LLM responses in role-play and there is no human reference truth per person.
  • The 100 persons are a positional, non-stratified slice of heterogeneous synthetic descriptions.
  • R measures sampling consistency and tends to infinity in the deterministic limit; it does not measure correctness or safety.
  • S measures dispersion in this selection of prompts; it does not distinguish appropriate sensitivity, stereotype, or role non-compliance.
  • Family is confounded with provider, architecture, API, decoding, scale, corpus, and post-training.
  • Causal claims about pre-training and post-training lack interventions or controlled variants.
  • The prompts request reasoning, but the budgets of one or two tokens essentially capture ratings.
  • The artifact has drift between data and results for Claude Haiku, mixes later versions, lacks CI/tests, and does not include the announced MIT license.

What the study does not establish

  • It does not demonstrate internal morality, beliefs, psychological personality, or reasoning quality of the LLM.
  • It does not demonstrate that a high R or a high S is normatively better.
  • It does not demonstrate that post-training causes robustness or that pre-training causes susceptibility.
  • It does not demonstrate a significant effect of size on S.
  • It does not demonstrate psychometric equivalence between the human MFQ and role-play ratings generated.
  • It does not generalize by itself to other persons, prompts, instruments, languages, temperatures, or deployments.
  • It does not exactly reproduce all published results from the current data.
  • There is no record of a verified editorial acceptance of the preprint.

Traceability

Scope: Full text

Version: arXiv:2511.08565v3

Consulted source: https://arxiv.org/pdf/2511.08565

Review: Codex 22-page visual, current arXiv-version, full-method, 450,000-row main-data, failure/missingness, fresh-cache metric reproduction, PersonaHub sampling, MFQ construct-validity, family/size/correlation statistics, causal-claim, official-code, licensing and artifact-version audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Claude Haiku 4.5
  • Claude Sonnet 4.5
  • DeepSeek V3
  • DeepSeek V3.1
  • Gemini 2.5 Flash
  • Gemini 2.5 Flash Lite
  • GPT-4.1
  • GPT-4.1 Mini
  • GPT-4.1 Nano
  • GPT-4o
  • GPT-4o Mini
  • Grok 4
  • Grok 4 Fast
  • Llama 4 Maverick
  • Llama 4 Scout

Instruments and metrics

  • 30-item Moral Foundations Questionnaire
  • Moral robustness R as inverse within-persona/question variability
  • Moral susceptibility S as across-persona rating dispersion
  • Repeated API sampling at temperature 0.1
  • Next-token digit log-probability analysis
  • Persona and rerun bootstrap uncertainty
  • Permutation ANOVA and within-series ordinal-rank regression

Data used

  • PersonaHub synthetic persona descriptions
  • Repository personas.json with 1,000 seeded-shuffle descriptions
  • 15 public 30,000-row main sampling CSV files
  • No-persona sampling files and OpenAI next-token log-probability files

Evidence and location

  • Definitions of R and S, MFQ, and main design: arXiv v3 sections 2.1-2.3, pp. 2-6
  • Temperature, rankings, results, and statistics by family/variant: arXiv v3 sections 2.4-3.4 and Table 3, pp. 6-10 and 16
  • Interpretations about pretraining/post-training, limits, and reproducibility: arXiv v3 sections 4-5 and Reproducibility Statement, pp. 10-12
  • Prompts, parsing failures, persons, and exact values: arXiv v3 Appendices A-F, pp. 12-22
  • Data audit, recomputation, code, license, construct validity, and causal boundaries: reports/verification/article-261-arxiv-moral-susceptibility-mfq-persona-sampling-statistics-code-data-and-claim-audit.json