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

Despite the growing utility of Large Language Models (LLMs) for simulating human behavior, the extent to which these synthetic personas accurately reflect world and moral value systems across different cultural conditionings remains uncertain. This paper investigates the alignment of synthetic, culturally-grounded personas with established frameworks, specifically the World Values Survey (WVS), the Inglehart-Welzel Cultural Map, and Moral Foundations Theory. We conceptualize and produce LLM-generated personas based on a set of interpretable WVS-derived variables, and we examine the generated personas through three complementary lenses: positioning on the Inglehart-Welzel map, which unveils their interpretation reflecting stable differences across cultural conditionings; demographic-level consistency with the World Values Survey, where response distributions broadly track human group patterns; and moral profiles derived from a Moral Foundations questionnaire, which we analyze through a culture-to-morality mapping to characterize how moral responses vary across different cultural configurations. Our approach of culturally-grounded persona generation and analysis enables evaluation of cross-cultural structure and moral variation.

💡 Research Summary

This paper investigates how well large language models (LLMs) can emulate human cultural and moral value systems when they are prompted to generate “culturally‑grounded” personas. The authors align the synthetic personas with three well‑established socio‑psychological frameworks: the World Values Survey (WVS), the Inglehart‑Welzel (IW) cultural map, and Moral Foundations Theory (MFT). Their research is organized around three questions: (RQ1) where synthetic personas fall on the two principal dimensions of the IW map; (RQ2) whether the personas’ responses to value‑probe questions match the distributional patterns observed in real demographic groups; and (RQ3) how the cultural variables used to condition the personas map onto moral foundation scores.

Methodology.
The authors first select ten WVS‑derived cultural variables that directly correspond to the indicators used to construct the IW axes (e.g., religiosity, materialism, national pride, gender equality, etc.). By taking the Cartesian product of all variable levels they obtain 93 312 distinct cultural configurations (the configuration space C). Each configuration is fed into a prompt template that asks GPT‑OSS 20B to generate a persona profile containing metadata (name, age, gender, occupation, country/region), a short biography, and explicit statements linking the persona’s attitudes to each of the ten variables. This yields a set P of 93 312 synthetic personas. Demographic analysis of P shows a near‑balanced gender split, ages concentrated in the 30‑49 range, and a realistic occupational distribution (education, social services, sales/marketing, IT, etc.). Country mentions follow a long‑tail distribution dominated by North America and Europe.

RQ1 – IW Map Positioning.
For each persona the model is conditioned on the Integrated Values Survey (IVS) questions that correspond to the ten WVS indicators. The responses are standardized and subjected to Principal Component Analysis with varimax rotation, reproducing the two components that define the IW map: PC1 (Survival vs. Self‑Expression) and PC2 (Traditional vs. Secular). The components are rescaled using the published formula (z₁ = 1.81·PC1 + 0.38, z₂ = 1.61·PC2 − 0.01) to obtain coordinates on the IW plane. A Voronoi tessellation of the plane is then computed, assigning each persona to a cell. Within each cell the authors treat the persona’s original cultural configuration as a transaction and apply the FPClose frequent‑pattern mining algorithm (minimum support = 0.2) to extract closed‑frequent itemsets. The resulting itemsets reveal coherent cultural “signatures” for each region of the map—for example, cells in the upper‑right quadrant (high self‑expression, secular) are dominated by configurations featuring high education, gender equality, and autonomy, whereas cells in the lower‑left (low self‑expression, traditional) are characterized by strong religiosity, national pride, and materialist orientation. These patterns closely mirror known cross‑national cultural clusters.

RQ2 – Alignment with WVS Distributions.
The authors employ the WorldValuesBench probe (WVB‑Probe), which contains 36 value‑related questions linked to three demographic attributes: continent, residential area, and education level. Each persona is assigned a demographic triple, and the LLM is prompted to answer the probe questions conditioned on the persona. For each demographic group g and question q, the distribution of LLM responses P₍g,q₎ is compared to the human baseline distribution H₍g,q₎ using Earth Mover’s Distance (EMD). The average EMD across all groups and questions is 0.12 (range 0–0.35), indicating a reasonably close match. However, systematic deviations appear for items related to traditional/religious values, where the LLM tends to over‑represent secular or liberal positions, suggesting a bias inherited from the predominantly Western training data.

RQ3 – Moral Foundations Mapping.
To assess moral alignment, the authors condition the LLM on each persona and administer the Moral Foundations Questionnaire‑2 (MFQ‑2), which measures five (or six, when fairness is split) moral foundations: Care, Fairness, Loyalty, Authority, and Purity. Average scores per foundation are computed for each persona. The authors then build a mapping model that predicts moral scores from the original cultural variables. The analysis shows strong positive associations between traditional‑religious configurations and higher Authority and Purity scores, while self‑expression, materialist configurations correlate with higher Care and Fairness scores. This mapping reproduces findings from cross‑cultural moral psychology, confirming that the synthetic personas encode plausible moral structures conditioned on cultural inputs.

Key Findings and Implications.

1. Cultural Fidelity: LLM‑generated personas, when conditioned on a comprehensive set of WVS‑derived variables, reproduce the major axes of cross‑cultural variation observed in the IW map. Frequent‑pattern mining uncovers coherent cultural signatures that align with known regional value profiles.
2. Distributional Alignment: On the WVB probe, the synthetic personas’ response distributions are close to human baselines (low EMD), but exhibit systematic biases on traditional/religious items, highlighting the need for bias‑mitigation when deploying LLMs in culturally sensitive contexts.
3. Moral Consistency: Moral foundation scores derived from the personas map sensibly onto the conditioning cultural variables, confirming that the LLM internalizes culturally contingent moral intuitions in a way consistent with Moral Foundations Theory.
4. Methodological Contribution: The paper introduces a scalable pipeline—selection of interpretable cultural variables, persona generation via prompting, and three parallel evaluation streams—that can be reused for auditing LLMs across other sociocultural frameworks.
5. Practical Relevance: Findings suggest that while LLMs can serve as useful simulators for cross‑cultural policy analysis, user‑interface design, or educational tools, developers must incorporate validation steps (e.g., EMD checks, moral mapping) to ensure alignment with target cultural norms and to avoid inadvertent reinforcement of cultural biases.

In summary, the study demonstrates that large language models are capable of generating culturally grounded personas that largely reflect real‑world value structures, yet they retain measurable biases, especially along the traditional‑secular dimension and in certain moral foundations. The proposed evaluation framework offers a rigorous way to quantify and mitigate these biases, paving the way for more culturally aware and ethically responsible AI systems.