Simulating social perception with large language models: perceptions of China’s common prosperity

This study proposes a domain-aligned framework for computational social science. It leverages Large Language Models (LLMs) to simulate social perception by grounding analysis in China’s Common Prosperity agenda. A dual-track design is adopted, combining expert-led qualitative interviews with structured questionnaires. This process yields a corpus enriched with attitude labels and expert reasoning chains. These annotations enhance interpretability and subgroup fidelity, enabling both micro-level inference and macro-level distribution modeling. Several base LLMs are fine-tuned on this corpus and evaluated under a unified six-task pipeline. This pipeline covers social background and interview dialogue simulation, as well as the simulation of attitudes and survey answers at both individual and group levels. Across all tasks, the domain-aligned models are found to consistently outperform state-of-the-art general-purpose LLMs. These models excel in preserving heterogeneity, recovering latent signals in minority groups, and reproducing empirical distributions with low Wasserstein distance. These results demonstrate that LLMs trained on small but semantically rich corpora, are effective instruments for perception modeling and population-level inference, highlighting the potential for future research to deepen domain alignment and integrate LLM-based simulations more systematically with empirical social data.