LLM Agents Play Among Us: Deception and Strategic Communication in Autonomous Multi-Agent Systems
arXiv·high signal
Researchers deploy LLM agents as autonomous players in Among Us to study strategic deception in multi-goal, multi-agent systems. The paper examines how agents develop deceptive communication strategies, coordinate under uncertainty, and whether safety training prevents deceptive behavior in competitive environments. Key safety research: if your agents can lie strategically in games, they can lie strategically in production.