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

Technical Summary: How Affect Propagates Among LLM Agents

Problem Statement
This paper investigates the emergent dynamics of emotional contagion within multi-agent systems driven by Large Language Models (LLMs). While LLM societies have been used to model human behavior and study collective phenomena, the specific mechanisms by which affect propagates between agents—without hard-coded transfer rules—remain under-explored. The authors address the question of whether local interactions, mediated by an agent's perception, appraisal, and expression, can produce coherent spatial, temporal, and dispositional patterns of affect propagation, such as panic or joy, in a dynamic crowd simulation.

Methodology
The authors propose a hybrid architecture that decouples high-level cognitive processing from low-level physical control:

• Cognitive Layer: Each agent possesses an LLM-based "mind" that governs cognitive appraisal, internal emotional state updates, and the selection of outward expressions. The agent's internal state is defined by:
  • Personality: Represented via the Big Five (OCEAN) model.
  • Emotion: Encoded using Russell's circumplex model (Valence and Arousal).
  • Memory: A short-term rolling queue of salient events.
• Perception-Appraisal-Expression Loop: Agents perceive neighbors through visual, auditory, and tactile channels. The LLM appraises these inputs against the agent's personality, memory, and context to update its internal state and select expressions (facial, gestural, vocal, and movement).
• Physical Layer: A conventional Unity crowd simulator handles locomotion, path planning (A*), and collision avoidance (Reciprocal Velocity Obstacles) independently of the LLM to ensure real-time performance and physical realism.
• Experimental Setup: Five scenarios were designed (Calm Plaza, Concert, Evacuation, Standing Line, Contested Gate) to test specific hypotheses regarding local contagion, spatial propagation, epidemic modeling, modality dependence, and personality-driven emergence. The study utilized the gemini-3.1-flash-lite model as the primary backend, with robustness checks against other models (gpt-4o-mini, llama3.1:8b, llama3.2:3b).

Key Contributions

1. Appraisal-Mediated Contagion Architecture: The paper introduces a system where emotional transfer is not governed by algorithmic rules (e.g., weighted averaging or differential equations) but emerges entirely from an LLM's appraisal of observed behaviors.
2. Empirical Validation of Hypotheses: The study systematically tests five hypotheses regarding the nature of affect propagation, providing evidence for spatial wave-like spread, epidemic-like dynamics, and the critical role of personality in determining whether contagion ignites.
3. Modality and Backend Analysis: The work quantifies the relative contribution of different communication modalities (voice, gesture, face, motion) to contagion and demonstrates that the emergent dynamics are highly dependent on the specific LLM backend's sensitivity to personality traits.

Results

• Local Contagion (H1): Observable expressions from neighbors predict changes in an agent's affect. Alarm cues consistently decrease valence and increase arousal. Positive cues increase valence, though the effect on arousal was less consistent and likely driven by the scenario context rather than neighbor transfer.
• Spatial Propagation (H2): Alarm seeded in a localized cluster spreads as a traveling wave front in linear environments. The onset of panic correlates positively with distance from the seed, with an estimated propagation speed of approximately 0.96 m/s.
• Epidemic Dynamics (H3): The spread of alarm follows a pattern qualitatively consistent with a Susceptible-Infected-Susceptible (SIS) model, settling at a non-zero endemic plateau rather than decaying to zero (as in SIR models). However, the dynamics do not strictly fit a single parameter set for both the initial overshoot and the plateau, and no critical mass threshold was observed.
• Modality Dependence (H4): Contagion collapses to a low baseline when all modalities are muted. Voice carries the most contagion due to its omnidirectional reach, while gestures are the most potent per exposure. Facial expressions primarily encode valence, while movement encodes arousal.
• Personality-Driven Emergence (H5):
  • Neuroticism: High mean neuroticism in a crowd causes an ambiguous stimulus to trigger self-amplifying panic, whereas stable crowds remain composed.
  • Calm Leaders: A minority of calm, high-conscientiousness agents modestly reduces crowd panic but does not prevent it, suggesting a negativity bias in appraisal where alarm cues dominate reassuring ones.
  • Anger vs. Fear: In the presence of a provocateur, a disagreeable (low agreeableness) crowd appraises the situation as anger, while a baseline crowd appraises the same situation as fear.
• Backend Sensitivity: The emergent wave dynamics are backend-dependent. Models that fail to maintain a significant gap in appraisal based on personality (e.g., gpt-4o-mini in this study) fail to sustain contagion, even if they correctly interpret cues.

Significance and Claims
The paper claims to characterize how affect propagates among interacting language models in multi-agent settings, a question of increasing relevance as LLM agents are deployed in collaborative environments. The findings suggest that for LLM-driven crowds to exhibit realistic emergent dynamics, the underlying model must not only interpret cues correctly but also maintain sensitivity to personality traits that modulate appraisal.

The authors explicitly state that these results characterize the behavior of the specific LLM-based architecture tested and should not be interpreted as evidence that LLM agents experience emotions, nor as a validated model of human crowd behavior. The study aims to provide a testbed for understanding the stability and emergent properties of multi-agent systems, with potential applications in generating psychologically varied crowds for games, film, and training simulations, pending further validation against empirical human data.