HOW AFFECT PROPAGATES AMONG LLM AGENTS: EMERGENT EMOTIONAL CONTAGION IN CROWD SIMULATION

This study introduces a multi-agent simulation architecture where individual agents are governed by an LLM-based cognitive layer. Instead of using hard-coded rules to dictate how emotions spread through a crowd, each agent perceives its neighbors, appraises the situation based on its unique personality profile (Big Five) and current affective state (Russell’s circumplex model), and then selects its own outward expressions. This creates a reciprocal feedback loop where an agent's behavior—such as screaming in fear or moving agitatedly—influences the perceptions and subsequent appraisals of nearby agents.

The research evaluates this architecture across various scenarios, including alarming, joyful, and neutral environments. The findings confirm that emotional contagion emerges naturally from these local interactions. For instance, in alarming situations, fear spreads from seeded agents as a traveling front, and the crowd's overall state often settles into a stable, non-zero plateau of alarm. The system also captures nuanced social phenomena: the distribution of personality traits within the crowd determines whether an ambiguous stimulus triggers collective panic or remains contained, and whether a provocation is interpreted as anger or fear.

As LLM-driven agents become standard for simulating human societies, understanding how affect propagates is critical for both social science and safety-critical applications. By replacing rigid, hand-authored transfer rules with LLM-based appraisal, this framework allows for more flexible and psychologically grounded simulations. This approach provides a testbed for studying complex social dynamics—such as evacuation behavior or the stability of multi-agent systems—in environments that would be logistically or ethically impossible to study with human participants.