This preprint, described in the manuscript as under review at VALE 2026, asks whether cultural survey responses can construct textual personas that improve aggregate prediction of other World Values Survey (WVS) questions. The descriptive result is clear within that design: across ten deliberately selected countries and 35 held-out questions, value-based personas obtain lower MAE than generic, country, and sociodemographic prompts. However, the phrase “cross-cultural simulation” needs an important boundary: the system starts from real responses from the same population, and its calibration also learns from human distributions in the target country. It is more accurately survey-assisted population imputation than autonomous reconstruction of a culture.
The study uses WVS Trend File 1981-2022 v4.1.0. For each respondent, ten answers underlying the Inglehart-Welzel cultural map are converted by GPT-5.4 or Claude Sonnet 4.6 into persona sentences. Each evaluated model receives these descriptors, answers a held-out question, and supplies probabilities over admissible integer options; distributions are averaged across personas. Nine countries on the cultural-map convex hull, Moldova, Taiwan, Japan, Iceland, Sweden, Puerto Rico, Colombia, Ghana, and Jordan, are tested alongside the United Kingdom as a WEIRD comparator. This is a deliberately extreme, not representative, country sample.
The appendix heatmap reports mean MAE of 0.164 for generic prompting, 0.154 for country prompting, 0.138 for sociodemographic prompting, and 0.108 for value personas. Ghana changes descriptively from 0.210 to 0.121 relative to generic prompting and Jordan from 0.213 to 0.125, while the cross-country range narrows. This supports the narrower finding that the ten observed items contain useful predictive signal for other WVS attitudes and that textual descriptors help these models use it. It does not show that Ghana or Jordan are less represented in model training: the paper measures no exposure, language, tokenization, or corpus coverage. Mitigation of inherited training bias is therefore a causal hypothesis rather than an observed mechanism.
Uncalibrated distributions are under-dispersed. The paper raises temperature and applies an exponential tilt whose beta preserves the original mean exactly. Temperature is fit leave-one-question-out from human distributions on the other questions in the same analysis and then applied to the held-out question. Lower Wasserstein distance and closer marginal variance are useful, but “without sacrificing accuracy” follows mathematically from constraining the expectation: mean-response MAE cannot change. The transform reweights existing support; it cannot create zero-probability answers or validate cross-question correlations, subgroups, individual consistency, or realistic multimodality.
The result is not zero-shot for a new population. Personas contain ten real answers from target-country individuals, and calibration needs human labels for other questions in that country. The design tests item-wise generalization within one survey corpus, not a country with no survey data. Nor does it predict each respondent: the paper explicitly limits its claim to aggregate means and distributions. WVS provides both predictors and outcomes, and the Inglehart-Welzel associations are well known and potentially present in pretraining.
Uncertainty reporting is incomplete. The paper omits human N, complete-case and persona counts by country, survey weights, wave composition, detailed missing-data handling, seeds, model repetitions, sample-curve resample count, and the method behind its bands or ranges. The heatmap applies Wilcoxon p<0.05 annotations to many comparisons without documented multiplicity correction; descriptor Mann-Whitney tests pool repeated question, country, and model cells without a clear observational unit. No detected difference between GPT-5.4 and Claude is not an equivalence result.
There is also internal model-version drift: methods list Qwen 3.5-36B and Ministral 3-27B, while figures label Qwen3.6-35B and Ministral-3-14B. IDs/revisions, providers, API dates, log-probability extraction, templates, quantization, and run settings are absent. No repository, code, data, outputs, per-persona tables, fitted temperatures, environment, tests, CI, or license was found; the paper promises release after publication. The numbers are therefore not currently reproducible. The faithful conclusion is that personas derived from real WVS values descriptively improve aggregate imputation of other WVS items and that human-supervised calibration corrects marginal dispersion. The work does not establish faithful cultural persons, genuine beliefs, individual prediction, a training-data-bias mechanism, human joint structure, or generalization to data-free countries, open responses, behavior, or high-stakes decisions.