LLM Agents Show Emotional Contagion Without Explicit Programming

A new study shared on arXiv has shown that large language model (LLM) agents in a crowd simulation can exhibit emotional contagion, even without a set way to transfer emotions directly. Researchers from an unnamed institution carried out the study, using agents that gather information through visual, auditory, and tactile senses. These inputs are processed by an LLM that considers factors like personality profiles, memories, emotional states, and context. The agents are designed based on the Big Five personality traits and Russell's circumplex model of emotions. To reduce latency, a standard crowd simulator handles basic navigation and steering. The results suggest that emotions circulate through a perception-appraisal-expression cycle, creating dynamic emotional interactions in groups.