Synthetic Personality Review

Culturally Grounded Personas in Large Language Models: Characterization and Alignment with Socio-Psychological Value Frameworks

Society, culture, and collective behavior2026arXivApproved editorial review

Authors: Candida M. Greco, Lucio La Cava, Andrea Tagarelli

Keywords: Large Language Models, Personality, Persona, Cultural Values, AI Safety

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

3
Authors
18
Findings
49
Limitations
9
Evidence

Editorial summary

English

This preprint studies whether synthetic personas constructed from cultural values produce responses that preserve structure on the Inglehart–Welzel map, approximate human World Values Survey (WVS) distributions, and show interpretable Moral Foundations Theory patterns. This review uses arXiv v2, revised on 3 June 2026, a 25-page PDF that remains marked “Under Review”; no peer-reviewed publication or official code/data repository was found.

The authors define ten categorical variables derived from WVS-7: religiosity, child-rearing values, moral acceptability, social trust, political participation, national pride, happiness, gender equality, materialism orientation, and tolerance toward outgroups. The Cartesian product of their levels yields 93,312 configurations. GPT-OSS-20B receives one configuration and writes a 250–500-word persona with a name, age, optional gender, occupation, country or region, bio, and an explicit account of how each variable manifests. The prompt forbids contradictions and requires every cultural-mapping sentence to reference its conditioning variable. The resulting corpus is therefore not a sample of observed people or cultures: it is an equally weighted, designed enumeration of combinations, some rare or sociologically implausible. Generated metadata are highly concentrated: 67.743% of ages are 30–39, 31.003% are 40–49, North America and Europe dominate, and 21.37% of occupations are in education and academia.

For RQ1, the same model adopts every persona and answers ten IVS/WVS indicators used to compute Inglehart–Welzel coordinates. The 93,312 positions form a cloud with no clear clusters. The planned Voronoi tessellation is replaced in the results by a regular grid of side length 1; FPClose discovers 735 closed itemsets across 111 non-empty cells. Spatial associations emerge among happiness, trust, national pride, child rearing, tolerance, and moral acceptability. This usefully describes the geometry induced by the pipeline, but the persona already explicitly contains the variables later elicited: the analysis mainly measures transmission and internal consistency of the conditioning, not independent recovery of human cultures.

For RQ2, GPT-OSS-20B again adopts each profile, self-assigns a continent, urban/rural setting, and education level, and answers 36 WVB-Probe questions. Its distributions are compared with human references for each demographic triple using EMD. The analysis includes 92,710 personas and 45 of 46 groups. The paper reports 1−EMD of 0.790 unweighted and 0.809 weighted by synthetic group size; 90.12% and 94.25% of comparisons, respectively, fall below EMD 0.4, but only 56.96% and 59.06% fall below 0.2. Coverage is extremely uneven: Europe–urban–tertiary contributes 32,716 personas; together with North America–urban–tertiary and Asia–urban–tertiary it accounts for 60,568 of 92,710 cases (65.3%), while 13 groups contain fewer than 100 personas and one contains a single persona. The weights describe demographics invented by the model rather than human prevalence.

For RQ3, two equally synthetic moral representations are compared. The first averages the 36 MFQ-2 answers produced by GPT-OSS while role-playing each persona. The second asks an LLM “oracle” to score the 32 cultural-variable values on six moral foundations and aggregates the values present in each configuration. Gemini 3 Pro, GPT-5.2, Claude Sonnet 4.5, and the default model are tried, but Gemini 3 Pro is selected because it produces greater variance; the others return near-neutral scores of 2–3. A greedy selection procedure with a linear correction then seeks variable subsets that approximate GPT-OSS MFQ scores. It finds stable compact sets for Loyalty, Authority, and Purity, national pride, religiosity, and related variables, while Care, Equality, and Proportionality show less differentiation. There are no human MFQ-2 responses serving as an external criterion: one LLM construction is aligned with another, and the most variable oracle is selected post hoc.

As a cross-model check, Qwen3.5-9B is run on 300 configurations sampled uniformly from IW space, with three runs per configuration. 57.33% of Qwen–GPT-OSS pairs have an IW distance between 0 and 1 and 91.66% between 0 and 2; the moral maps qualitatively preserve greater differentiation for binding than individualizing foundations. The paper provides no correlations, uncertainty intervals, null baseline, or justified threshold for interpreting those distances, and the moral comparison relies mainly on visual maps.

The defensible contribution is an exploratory framework, transparent about its prompts and several limitations, for auditing the cultural and moral structure an LLM generates under controlled conditioning. The results support internal coherence and aggregate distributional patterns within that system. They do not demonstrate that the personas are culturally authentic, represent real individuals, countries, or groups, or yield psychologically or cross-culturally valid moral responses. Without released artifacts, code, seeds, exact model revisions, and outputs, the calculations and end-to-end reproducibility cannot be independently audited.

Español

Este preprint estudia si personas sintéticas construidas a partir de valores culturales producen respuestas que conservan estructura en el mapa de Inglehart–Welzel, se aproximan a distribuciones humanas del World Values Survey (WVS) y muestran patrones interpretables en Moral Foundations Theory. Esta revisión se basa en arXiv v2, revisado el 3 de junio de 2026, un PDF de 25 páginas que continúa marcado «Under Review»; no se encontró una publicación revisada por pares ni un repositorio oficial con código o datos.

Los autores definen diez variables categóricas derivadas de WVS-7: religiosidad, valores de crianza, aceptabilidad moral, confianza social, participación política, orgullo nacional, felicidad, igualdad de género, orientación materialista y tolerancia a la diversidad. El producto cartesiano de sus niveles genera 93.312 configuraciones. GPT-OSS-20B recibe una configuración y redacta una persona de 250–500 palabras con nombre, edad, género opcional, ocupación, país o región, biografía y una explicación explícita de cómo se manifiesta cada variable. El prompt prohíbe contradicciones y exige vincular cada frase del mapa cultural con la condición recibida. Por tanto, el corpus no es una muestra de personas o culturas observadas: es una enumeración equiprobable y diseñada de combinaciones, algunas raras o sociológicamente inverosímiles. Los propios metadatos generados están muy concentrados: 67,743 % de edades entre 30–39, 31,003 % entre 40–49, predominio de Norteamérica y Europa y 21,37 % de ocupaciones en educación y academia.

Para RQ1, el mismo modelo adopta cada persona y responde diez indicadores IVS/WVS usados para calcular coordenadas del mapa Inglehart–Welzel. Las 93.312 posiciones forman una nube sin clústeres claros. El método previsto de teselación de Voronoi se sustituye en resultados por una cuadrícula regular de lado 1; FPClose encuentra 735 itemsets cerrados en 111 celdas no vacías. Aparecen asociaciones espaciales entre felicidad, confianza, orgullo nacional, crianza, tolerancia y aceptabilidad moral. Es una descripción útil de la geometría inducida por el pipeline, pero la persona ya contiene de forma explícita las mismas variables que después se preguntan: el análisis mide sobre todo transmisión y coherencia interna del condicionamiento, no recuperación independiente de culturas humanas.

Para RQ2, GPT-OSS-20B vuelve a adoptar cada perfil, se autoasigna continente, entorno urbano/rural y educación, y responde 36 preguntas de WVB-Probe. Se comparan sus distribuciones con referencias humanas por triple demográfico mediante EMD. Entran 92.710 personas y 45 de 46 grupos. El artículo informa 1−EMD de 0,790 sin ponderar y 0,809 ponderado por el tamaño de los grupos sintéticos; 90,12 % y 94,25 % de las comparaciones, respectivamente, quedan por debajo de EMD 0,4, pero solo 56,96 % y 59,06 % quedan por debajo de 0,2. La cobertura es extremadamente desigual: Europa–urbano–terciario aporta 32.716 personas; junto con Norteamérica–urbano–terciario y Asia–urbano–terciario concentra 60.568 de 92.710 casos (65,3 %), mientras 13 grupos tienen menos de 100 personas y uno solo una. Las ponderaciones representan la demografía que inventa el modelo, no prevalencias humanas.

Para RQ3 se comparan dos representaciones morales igualmente sintéticas. La primera promedia las 36 respuestas MFQ-2 que GPT-OSS produce al representar a cada persona. La segunda asigna a los 32 valores culturales puntuaciones en seis fundamentos morales mediante un LLM «oráculo» y agrega los valores presentes en cada configuración. Se prueban Gemini 3 Pro, GPT-5.2, Claude Sonnet 4.5 y el modelo base, pero se elige Gemini 3 Pro porque genera más varianza; los demás dan puntuaciones cercanas a 2–3. Después, una selección greedy con corrección lineal busca subconjuntos de variables que aproximen las puntuaciones MFQ de GPT-OSS. Encuentra conjuntos estables y compactos para Lealtad, Autoridad y Pureza, orgullo nacional, religiosidad y otras variables afines, mientras Cuidado, Igualdad y Proporcionalidad muestran menos diferenciación. No hay respuestas humanas MFQ-2 que actúen como criterio externo: se alinea una construcción de un LLM con otra construcción de LLM y se selecciona post hoc el oráculo más variable.

Como comprobación entre modelos, Qwen3.5-9B se ejecuta sobre 300 configuraciones muestreadas uniformemente del espacio IW, con tres runs por configuración. El 57,33 % de los pares Qwen–GPT-OSS queda a distancia IW entre 0 y 1 y el 91,66 % entre 0 y 2; los mapas morales conservan cualitativamente más diferenciación en fundamentos vinculantes que individualizantes. El artículo no aporta correlaciones, intervalos, baseline nulo ni un umbral justificado para interpretar esas distancias, y la comparación moral se apoya principalmente en mapas visuales.

La contribución defendible es un marco exploratorio, transparente en sus prompts y explícito sobre varias limitaciones, para auditar qué estructura cultural y moral genera un LLM bajo condicionamiento controlado. Sus resultados respaldan coherencia y patrones distribucionales a nivel agregado dentro de ese sistema. No demuestran que las personas sean culturalmente auténticas, que representen individuos, países o grupos reales, ni que las respuestas morales sean válidas psicológica o interculturalmente. Sin artefactos, código, semillas, revisiones exactas de modelos y salidas publicadas, los cálculos y la reproducibilidad no pueden auditarse de extremo a extremo.

Research question

What cultural structure do 93,312 personas generated by conditioning an LLM with combinations of ten WVS variables project, to what extent do their aggregated responses to WVB-Probe approximate human distributions by demographic group, and how do their MFQ-2 responses relate to an LLM mapping of cultural variables to moral foundations?

Method

Preprint arXiv v2. The 93,312 combinations of the Cartesian product of ten categorical variables derived from WVS-7 are enumerated and GPT-OSS-20B generates one persona per combination. RQ1: the model responds to ten IVS indicators, Inglehart–Welzel coordinates are computed, and closed itemsets are mined with FPClose over a grid. RQ2: the model generates a demographic triple and responds to 36 WVB-Probe items; the distributions of 92,710 personas and 45 groups are compared with human references using EMD. RQ3: GPT-OSS MFQ-2 scores are compared with cultural variable scores produced by an LLM oracle; a greedy selection with linear correction is repeated 50 times. Qwen3.5-9B replicates IW and moral analyses across 300 configurations, three runs per case. Temperature 1.0 and top_p 1.0; no code, data, or seeds are published.

Sample: Exhaustive synthetic corpus of 93,312 equiprobable configurations, not a human sample. For WVB-Probe, 92,710 profiles and 45 of 46 demographic triples enter; the three largest groups concentrate 60,568 cases (65.3%) and 13 groups have fewer than 100. Qwen is evaluated on 300 configurations with three runs. The expert review of personas is described as a stratified sample of extremes, axes, center, and regions, but its size, exact selection, protocol, and agreement are not reported.

Findings

  • The 93,312 IW positions form a slightly tilted elliptical cloud, with no clear cluster structure.
  • Since no clusters emerge, the analysis replaces the proposed Voronoi tessellation with a regular grid of side length 1.
  • FPClose identifies 735 closed itemsets in 111 non-empty cells of the IW map.
  • High happiness and high confidence appear together in 21 cells with mean support 0.735; minimal happiness and distrust in 15 cells with mean support 0.806.
  • The baseline persona without conditioning lies in the secular and self-expression quadrant, while conditioned personas occupy a broader region.
  • RQ2 covers 92,710 of 93,312 personas and 45 of 46 demographic groups with human reference.
  • The aggregated WVB-Probe alignment is 1−EMD 0.790 unweighted and 0.809 weighted by synthetic group size.
  • 90.12% unweighted and 94.25% weighted of questions fall below EMD 0.4; only 56.96% and 59.06% fall below 0.2.
  • Group scores range from 0.628–0.840 and some groups have sizes too small for a stable distribution.
  • Europe–urban–tertiary is the largest group with 32,716 profiles and a lower fraction of questions with EMD <0.2 than North America or Asia with the same environment and education.
  • The uniform moral mapping of the ten variables differs in scale and trend from the MFQ-2 responses, which motivates subsequent subset selection.
  • Gemini 3 Pro is chosen as the moral oracle because its scores vary more; the other models tested tend toward 2–3.
  • The greedy selection produces always-selected sets for Proportionality (materialism), Loyalty (national pride), Authority (five variables), and Purity (religiosity).
  • Care and Equality show less stable selection: mean Jaccard 0.669 and 0.738, respectively.
  • Associations are stronger for Loyalty, Authority, and Purity than for Care, Equality, and Proportionality.
  • In 300 cases, 172 Qwen–GPT-OSS pairs (57.33%) lie at IW distance ≤1 and 275 (91.66%) at distance ≤2.
  • The Qwen maps qualitatively preserve the contrast between binding and individualizing foundations observed with GPT-OSS.
  • The evidence demonstrates regularities of the pipeline under explicit prompts; it does not provide external validation of individual cultural authenticity.

Limitations

  • The work remains marked Under Review and no peer-reviewed version was found.
  • Artifacts are promised only after acceptance and are not published at the time of this review.
  • There is no official repository of code, generated dataset, raw outputs, or analysis scripts available.
  • Parsing, retries, invalid responses, failures, and completeness of millions of responses cannot be audited.
  • No exact commit, revision, or hash is fixed for GPT-OSS-20B or Qwen3.5-9B.
  • No inference engine, quantization, chat template, numerical precision, parallelization, or dependency versions are identified.
  • For Gemini 3 Pro, GPT-5.2, and Claude Sonnet 4.5, no endpoint, snapshot, date, parameters, or effective provider are reported.
  • Temperature and top_p are reported, but no seeds or determinism policy.
  • The corpus is the equiprobable Cartesian product of designed variables and does not reflect human prevalences or covariation.
  • The Cartesian product contains rare or sociologically implausible combinations, acknowledged by the authors.
  • The ten variables are a normative reduction of culture and do not encompass language, history, institutions, class, region, or situational context.
  • The moral acceptability variable groups heterogeneous issues such as abortion, euthanasia, divorce, suicide, and homosexuality.
  • Categories are treated as ordinal for correlations even though the authors describe them as conceptually categorical.
  • The direction of ordinal coding is chosen to align with IW indicators, introducing pattern-oriented analytical decisions.
  • Generation is entirely in English and does not test multilingual equivalence or culturally localized formulations.
  • A single main model family generates profiles, metadata, and responses, increasing dependence and self-caused coherence.
  • The prompt forces each variable to be explicit and forbids contradictions; subsequently asking about those same constructs creates construct leakage.
  • RQ1 reuses indicators linked to the variables that already defined the persona, so it is not an independent validation.
  • The post hoc substitution of Voronoi with a grid occurs after observing that there are no clusters.
  • Cell size, support 0.2, and retention threshold are not subjected to published sensitivity analysis.
  • There is no baseline with direct configuration, shuffled profile, empty persona, demographics alone, or generic persona.
  • RQ2 lacks comparison with the model without persona; the observed EMD cannot be attributed to the generated profile.
  • The model simultaneously generates the demographic triple and the responses, so grouping and behavior share the same priors.
  • The association between cultural configuration and country is endogenous to the model and may reproduce pretraining stereotypes.
  • The synthetic demographic distribution is heavily biased toward Europe/North America, urban, tertiary education, and ages 30–49.
  • 65.3% of RQ2 is concentrated in three urban-tertiary groups; aggregate results are dominated by those groups.
  • Thirteen groups have fewer than 100 profiles and one has n=1, with no intervals or minimum size rules.
  • Weighting uses group sizes invented by the LLM, not human population weights.
  • The sizes and uncertainty of the reference human distributions are not reported or propagated.
  • The EMD 0.4 and 0.2 thresholds are labeled moderate/high without published empirical justification.
  • The printed EMD formula does not specify normalization across questions with different scales; without code, comparability cannot be verified.
  • The 1−EMD score is averaged across questions and groups without confidence intervals or testing against a baseline.
  • There are no human MFQ-2 responses to validate the moral profiles.
  • The two moral representations are generated by LLMs; their similarity measures agreement between synthetic constructions.
  • The Gemini oracle is selected post hoc for greater variance, a criterion that may favor visually informative patterns without greater validity.
  • Complete comparable results for the other oracles or a prior selection rule are not published.
  • The text calls a r_MFT ground truth during selection, although the logical objective appears to approximate r_MFQ; the specification is internally ambiguous.
  • The train/validation split, stopping criterion, linear correction, seeds, and R² distribution are not detailed.
  • Random splits over a Cartesian product reuse the same 32 categorical levels in training and validation and do not test generalization to new cultures or configurations.
  • The 50-split repetition evaluates selection stability, not external validity of the moral foundation.
  • The ISMed review does not report the number of profiles, experts, training, rubric, blinding, disagreements, or inter-rater reliability.
  • The Qwen validation uses only 300 of 93,312 configurations and does not fully detail uniform sampling in IW space.
  • Qwen–GPT distances lack a baseline, reference scale, intervals, and a previously justified threshold.
  • The moral comparison between Qwen and GPT is primarily visual and does not report correlations, EMD, or other concordance metric.
  • Variation due to regenerating the persona, re-answering the questionnaire, and stochastic sampling are not separated.
  • No additional closed or open models are fully evaluated over the entire corpus.
  • The conclusions are descriptive and do not allow individual prediction, causality, or decisions about persons or groups.
  • Personas with plausible names, countries, and professions may reinforce stereotypes even when they do not correspond to real individuals.
  • WVS-7 corresponds to 2017–2022 and may not reflect recent cultural changes, acknowledged by the authors.

What the study does not establish

  • It does not demonstrate that an LLM possesses culture, moral values, or a human social identity.
  • It does not establish that synthetic personas are culturally authentic or realistic individuals.
  • It does not test that an equiprobable configuration corresponds to the distribution of a real population.
  • It does not validate individual predictions of WVS or MFQ-2 responses.
  • It does not establish statistical equivalence between synthetic and human distributions.
  • It does not demonstrate that the generated profile adds more than directly indicating the ten variables to the model.
  • It does not separate cultural learning of the model from prompt obedience or semantic repetition.
  • It does not demonstrate that the generated country, age, gender, or occupation are valid sociological consequences of values.
  • It does not provide an external human criterion to assert that the mapping of moral foundations is correct.
  • It does not establish that greater variance of an oracle implies greater moral quality or validity.
  • It does not demonstrate multilingual, cross-cultural, or cross-prompt-formulation robustness.
  • It does not justify using the personas for clinical, political, educational, commercial, or collective decisions.
  • It does not test causality between cultural variables and moral foundations.
  • It does not allow end-to-end reproduction of results while code, data, and model revisions remain unpublished.

Traceability

Scope: Full text

Version: arXiv:2601.22396v2, revised 3 June 2026; 25-page preprint marked Under Review

Consulted source: https://arxiv.org/pdf/2601.22396v2

Review: Codex full-text, bilingual-fidelity, arXiv-version, 25-page visual, construct-leakage, distributional-weighting, EMD-scale, moral-oracle, model-version, artifact-availability and reproducibility audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI GPT-OSS-20B as the default open-weight persona-generation and response model; exact repository revision and serving stack not reported
  • Qwen3.5-9B as a cross-model check on 300 configurations with three runs per configuration; exact revision not reported
  • Google Gemini 3 Pro selected as the culture-to-morality oracle because its outputs had greater variance
  • OpenAI GPT-5.2 evaluated as an alternative culture-to-morality oracle; detailed outputs and API snapshot not reported
  • Anthropic Claude Sonnet 4.5 evaluated as an alternative culture-to-morality oracle; detailed outputs and API snapshot not reported

Instruments and metrics

  • Ten author-defined categorical cultural variables derived from WVS-7 items and Inglehart–Welzel dimensions
  • Persona generation prompt with explicit metadata, bio, and ten-variable cultural mapping
  • Ten IVS/WVS indicators used to reconstruct Inglehart–Welzel coordinates
  • WorldValuesBench WVB-Probe: 36 value questions and continent × settlement × education reference groups
  • Moral Foundations Questionnaire-2: 36 items across Care, Equality, Proportionality, Loyalty, Authority, and Purity
  • LLM-judged 32 cultural-value × 6 moral-foundation mapping matrix
  • Earth Mover's Distance and reported 1−EMD aggregate alignment
  • Regular-grid partition and FPClose closed frequent-itemset mining
  • Greedy forward variable selection, linear correction, validation R², and 50 repeated splits
  • Spearman correlations and qualitative cross-model IW/moral map comparison

Data used

  • World Values Survey Wave 7 country-pooled datafile, version 5.0, covering survey waves from 2017–2022
  • Integrated Values Surveys: merged EVS Trend File and WVS Trend File for Inglehart–Welzel reconstruction
  • WorldValuesBench WVB-Probe split with 36 questions and human reference distributions for demographic triples
  • 93,312 GPT-OSS-20B-generated persona profiles, one per designed cultural configuration; not publicly released at review time
  • 92,710 generated personas mapped to 45 WVB-Probe-supported demographic groups for RQ2
  • 300-condition Qwen3.5-9B validation sample drawn uniformly from generated Inglehart–Welzel space

Evidence and location

  • Editorial status, versions, authors, and abstract: arXiv:2601.22396v2 abstract page and submission history; revised 3 June 2026; marked Under Review
  • Variables, 93,312 configurations, generation, and expert review: arXiv v2, Section 3 and Appendices C–E, Tables 5–10 and Figure 5
  • IW projection, grid, itemsets, and RQ1 results: arXiv v2, Sections 4.1 and 5.1, Figures 2–3 and 9–11, Table 1
  • WVB-Probe, EMD, coverage, and demographic inequality: arXiv v2, Sections 4.2 and 5.2, Tables 2–3 and 11, Figures 7 and 12
  • MFQ-2, oracle selection, and moral mapping: arXiv v2, Sections 4.3 and 5.3, Figures 8 and 13–16, Tables 12–13
  • Qwen validation on 300 configurations: arXiv v2, Appendix J, Figures 17–18
  • Environment, parameters, and declared reproducibility: arXiv v2, Appendix A and Reproducibility Statement
  • Limitations, stereotype risks, and descriptive scope: arXiv v2, Limitations and Ethical Considerations
  • Absence of definitive publication and official code/data links: arXiv abstract Code, Data and Media section and targeted title/arXiv/GitHub searches, checked 15 July 2026