Generative Agent Simulations of 1,000 People

https://arxiv.org/abs/2411.10109

Agents from two-hour interviews with 1,052 U.S. participants. The arXiv abstract as fetched (later version, retitled LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals) gives normalized GSS accuracy of 83% interview-only, 82% survey-only, 86% combined, 74% demographics-only; the Stanford HAI brief reports 85%. Reduced accuracy disparities across racial and ideological groups.

LLM models of individual human choice