LLM Agents Exhibit Emergent Emotional Contagion

Researchers have investigated how emotions spread among large language model (LLM) agents within a crowd simulation, revealing emergent emotional contagion. The simulation architecture does not include any explicit, hand-authored rules for transferring emotional states directly between agents. Instead, each agent perceives its neighbors through various sensory channels (visual, auditory, tactile) and then appraises these perceptions. This appraisal process, handled by an LLM, considers the agent's prompted personality profile, memory, current emotional state, and the situational context. Based on this, the agent updates its internal affective state and chooses an outward expression. The agents are designed using the Big Five personality model and Russell's circumplex model of affect. To maintain performance, low-level steering and navigation are managed by a conventional crowd simulator, separate from the LLM-based cognitive layer. The system was evaluated across five scenarios, including alarming, joyful, and neutral situations. Results showed that emotional contagion dynamics emerged with clear spatial, temporal, and personality-dependent structures in sparse, small crowds. For example, alarm spread like a traveling front, and the distribution of personality profiles influenced how ambiguous alarms were interpreted or how provocations led to anger or fear. Further controlled experiments confirmed that these dynamics are dependent on the LLM backend used.