LLM Agents Develop Emergent Reputation Systems, Trust Networks, and Strategic Deception in Repeated Avalon Games

arXiv·medium signal

Researchers study LLM agents playing repeated games of The Resistance: Avalon with persistent memory of past interactions. Unlike prior single-game studies, agents develop emergent reputation tracking, trust networks based on behavioral history, and increasingly sophisticated deception strategies across rounds. The finding that memory-augmented LLM agents spontaneously develop social dynamics mirrors real multi-agent deployment scenarios where agents interact repeatedly and build implicit models of each other.