LLM Agents Simulate Human Behavior from Self-Reports

A recent study investigates the ability of large language models (LLMs) to generate general-purpose simulations of individuals, referred to as generative agents, based on self-reported data. Researchers utilized a varied national sample of 1,052 Americans, constructing agents through two-hour semi-structured interviews (American Voices Project), structured surveys (General Social Survey and Big Five personality inventory), or a combination of both methods. The accuracy of these agents on withheld General Social Survey items was 83% (interview only), 82% (surveys only), and 86% (combined), demonstrating superior performance compared to agents relying solely on demographic data. These results indicate that LLMs can forecast human behavior across various domains without needing extensive structured datasets for each specific outcome, presenting a more versatile method for behavioral simulation.