General Social Agents

Presents an approach for building AI “general” agents that can predict human behavior in novel settings without requiring extensive setting-specific training data. The agents use theory-grounded natural language instructions combined with existing empirical data and knowledge from language model pretraining. Demonstrates that these agents can predict initial human play across 883,320 novel games better than cognitive hierarchy models, game-theoretic equilibria, and out-of-the-box agents, sometimes outperforming even published human behavioral data.