2026-08-21
Across 59 countries and nine open models, LLM value simulation is more accurate for richer countries. Extra context improves accuracy and equality; preference alignment does not.
LLM personas are now a common shortcut for studying public values across countries. Mean accuracy is the wrong summary if North America, Europe, and Australia are simulated well while other regions are pulled toward a U.S.-like answer. Downstream opinion experiments then flatten underrepresented voices.
Researchers at Tsinghua, Fudan, Nankai, and Shanghai Jiao Tong turn that imbalance into a comparable quantity: representational equality. The question is not whether groups should give identical answers, but whether simulation accuracy is evenly distributed. The panel covers Wave 7 of the World Values Survey: 59 countries, 2,420 country–demographic cells, and 160 shared multiple-choice items.
Simulation uses first-token probabilities. For each item the model returns the top-20 first-position tokens; valid option letters are kept and normalized into a predicted response distribution. Accuracy is 1 minus Jensen–Shannon divergence from the human distribution, averaged first within each of 11 value dimensions and then across dimensions so question-heavy domains cannot dominate.
The headline equality index is EqCV, the coefficient of variation of country-level accuracies; lower is more even. Max–min gap, min–max ratio, and Gini are reported in the appendix and agree. A composite AE score is the geometric mean of mean JSD and EqCV, so a model must be both accurate and even to score well.
Two intervention families are compared. Contextual adaptation uses native-language prompts in Chinese, Arabic, and Spanish, or retrieves three non-target WVS distributions as historical memory. Parametric modification evaluates SambaLingo continued post-training on six languages, and DPO/GRPO on human versus GPT-annotated HH-RLHF. Nine open models run from ChatGLM3-6B to Qwen2.5-72B.
Mean country accuracy ranges from 0.538 to 0.739. GLM-4-9B leads; the two small Qwen models sit at the bottom. EqCV ranges from 0.0286 (ChatGLM3-6B) to 0.0546 (Qwen2.5-72B). GLM-4-9B remains unequal at 0.0486. AE reverses the ranking: ChatGLM3-6B 0.096, GLM-4-9B 0.113, Mistral-7B 0.116. The most accurate model is not the most even.
Accuracy correlates positively with GDP per capita, internet use, the Global Innovation Index, and governance indicators, negatively with power distance, and positively with individualism and indulgence. Australia, North America, and Europe are simulated better than Egypt, Jordan, and Iraq. On family trust, more than 95% of Iraqi respondents choose “trust completely” versus about 60% in the United States; GLM-4-9B predicts around 40% for both, and for more than half the items the Iraqi-profile output is closer to U.S. humans than to Iraqi humans.
Native-language prompts often raise accuracy, with model-dependent failures. On Jordan’s freedom-versus-security item, 90% of locals choose security; Qwen2.5-72B under English assigns 100% to freedom and under Arabic 79.2% to security. When gains concentrate in one language group, inequality is relocated rather than reduced. Extra information is more consistent: EqCV falls for almost every model, and GLM-4-9B under BM25 drops from 0.0486 to 0.0255. Language-specific continued training raises averages in all six groups, with Arabic ranging from −1.2% in Jordan to +3.4% in Morocco. Preference alignment yields no systematic gain; human annotations preserve accuracy better than GPT annotations.
Cross-country simulation and cultural-alignment evals that report only mean accuracy hide a structural bias. This paper turns evenness into an index and splits prompt, retrieval, continued training, and alignment. If the goal is a more even global simulator tomorrow, adding subgroup context beats another round of preference optimization. It is an evaluation paper, not a new trainer.
The main evidence is fixed-option surveys and does not transfer to open discourse, moral reasoning, or behavior. Country as the unit masks within-country heterogeneity and cross-border linguistic minorities. WVS dimensions differ in item count, abstraction, translation, and social-desirability pressure. First-token probabilities are cheaper than sampled text and match it in the appendix, but they remain an approximation inside a constrained option set. Alignment uses off-the-shelf preference data, not labels collected for value simulation.