A new CDP diagnostic finds DeepPersona-Inspired prompting shows the strongest cultural flattening in every model-domain block across four backbones and two survey domains

The work introduces Cultural Divergence Preservation (CDP), which identifies reduced cross-country divergence as cultural flattening and increased divergence as cultural caricature. Existing distance-based metrics such as Jensen-Shannon divergence do not directly capture cross-country differences; CDP directly quantifies the attenuation or amplification of cross-country divergence and is reference-light, relying on a one-time human calibration. The metric is evaluated across four LLM backbones, three persona-based prompting methods, and two survey domains, the WVS and the Big Five Personality Test.

There is a systematic discrepancy between conventional fidelity metrics and CDP. Controlled experiments show CDP changes monotonically as cross-country divergence is attenuated or amplified, while the corresponding changes in JSD remain relatively small, indicating that within-country distributional fidelity alone does not show whether cross-country differences are preserved. The finding rests on two kinds of evidence: controlled experiments and an audit of real LLM generations.

In the audit of real LLM generations, DeepPersona-Inspired prompting is frequently favored by conventional fidelity metrics but exhibits the strongest flattening in every model-domain block. This suggests that selecting prompting methods by conventional fidelity metrics may systematically favor configurations that compress cross-country divergence. The pattern is observed in every model-domain block formed by the four model backbones and two survey domains.

The work targets research and engineering settings that use LLMs as synthetic survey respondents and need to estimate cross-country response distributions, and it applies to the evaluation stage of cross-cultural survey simulation. CDP is positioned as a complement to fidelity metrics rather than a replacement: it directly quantifies the attenuation or amplification of cross-country divergence, so it is suited to being used alongside metrics such as JSD when the question is whether cross-country differences are preserved. Its calibration relies on a one-time human annotation, so users applying it to a new survey domain or a new language population need to reconfirm that the calibration still holds in that setting.

The abstract does not give the concrete computation of CDP, the scale and procedure of the one-time human calibration, or specific values or effect sizes for each model-domain block, so the absolute magnitude of flattening cannot be judged from the abstract. The exact operations used in the controlled experiments to attenuate or amplify divergence, and the quantitative standard for JSD changes being relatively small, also require the full text. In addition, the conclusions rest on two survey domains, the WVS and the Big Five Personality Test, so whether other survey topics or language populations show the same pattern remains an open question.