How Affect Propagates among LLM Agents: Emergent Emotional Contagion in Crowd Simulation

A new crowd-simulation setup uses LLMs to drive each agent’s perception, appraisal, and outward emotional expression, rather than wiring in direct affect transfer between characters. Agents observe neighbors through visual, auditory, and tactile channels, then update internal state based on personality, memory, and context. Low-level steering stays in a traditional crowd simulator, which keeps the approach practical by avoiding LLM pathfinding overhead.

The interesting part for game AI is that emotional contagion emerges from the loop itself. In sparse, small crowds, seeded alarm can propagate as a traveling wave, and the overall alarmed population can settle above zero instead of snapping back to neutral. Personality distribution also matters: the same ambiguous stimulus can tip a crowd into panic for one mix of profiles, while another mix reads it as something milder.

The work spans five scenario environments, including alarming, joyful, and neutral setups with different spatial layouts. It also probes the appraisal step across prompt variants, sampling temperatures, and four model backends, and the resulting behavior changes depending on which model is doing the reasoning. That makes the system useful as both a simulation technique and a reminder that LLM-based NPC behavior is still highly backend-sensitive.

For developers, the practical takeaway is that LLMs can add socially plausible affect to crowds without bespoke contagion rules, but the cost is variability and tuning complexity. If you’re building systemic NPC behavior, this points toward hybrid architectures:...