Representational Equality in Cross-country Value Simulation: A Systematic Analysis of Large Language Models

2026-08-08

AI Summary

This study addresses significant representational inequities in large language models (LLMs) when simulating values across nations, revealing that models more accurately reflect the values of individuals from wealthier and technologically advanced countries—a bias that risks amplifying societal inequities. The work presents the first systematic evaluation of value simulation accuracy across 59 countries and compares two intervention strategies: contextual adaptation (e.g., native-language prompting and auxiliary information injection) and parameter modification (e.g., language-specific continued pretraining and preference alignment). Findings indicate that improving overall or target-group average accuracy does not necessarily enhance representational equity. Among the approaches tested, injecting auxiliary information grounded in human-annotated preference data most effectively boosts both accuracy and equity, whereas preference alignment yields no systematic benefits.

Abstract

Traditional methods for studying human opinions often struggle to support representative and scalable research across countries. Large language models (LLMs) can serve as scalable proxies for simulating human opinions, enabling more efficient opinion analysis. However, this use of LLMs requires not only high average accuracy but also representational equality, that is, comparable simulation accuracy across populations. Uneven simulation accuracy may reproduce or amplify societal biases in downstream applications. This study systematically investigates country-level representational equality across 59 countries and finds substantial, systematic inequality. Populations from wealthier and more technologically advanced countries are simulated more accurately. We further compare two foundational intervention pathways, contextual adaptation and parametric modification, and show that improvements in average or target-group accuracy do not necessarily translate into greater representational equality. For contextual adaptation, native-language prompting generally improves accuracy but remains model-dependent, whereas additional information more often improves both accuracy and equality. For parametric modification, language-specific continued post-training improves accuracy for targeted language groups but unevenly, while preference alignment yields no systematic gains in accuracy or equality. Human-annotated preference data generally preserve accuracy better than AI-annotated data. These findings highlight the need for representational equality alongside accuracy and offer guidance for more inclusive, socially responsible LLM-based simulations.

Problem

• representational equality
• cross-country simulation
• large language models
• societal bias
• opinion simulation

Innovation

• representational equality
• cross-country simulation
• large language models
• contextual adaptation
• parametric modification