Long-term multi-agent simulation can teach LLMs social intelligence—agents trained on years of simulated life experience show better understanding of human-like social behavior and role-playing tasks.
Agentopia simulates 100 AI agents living together for 10 simulated years, learning from social interactions and personal growth. The framework trains language models using a 'life reward' signal based on agent well-being, showing that agents develop realistic social behaviors and that this training improves the underlying model's ability to handle social reasoning tasks.