Reproduce “Persona-Based Simulation of Human Opinion at Population Scale” on OpenResearch

About this paper

Mao Li, Frederick G.Conrad

Researchers from the University of Michigan developed SPIRIT, a framework enabling large language models to simulate human opinion at population scale by inferring multi-dimensional personas from social media and integrating them with probability-based sampling. This approach yields simulations that are more accurate, heterogeneous, and representative of population-level distributions compared to demographic-only methods.

Key insights

• Psychologically rich, multi-dimensional personas derived from social media traces significantly enhance LLM simulation accuracy and the reproduction of human-like heterogeneity in responses, outperforming demographic-only profiles.
• Combining persona inference with probability-based sampling and weighting allows LLM-based simulations to achieve population-level generalizability, enabling valid social scientific inference.
• LLM agents exhibit biases, such as 'deliberation bias' (constructing detailed causal chains for complex questions) and a tendency to default to normative or factual judgments over nuanced human belief states, especially in morally loaded contexts.

Summary by alphaXiv.