Realistic threat perception drives intergroup conflict: A causal, dynamic analysis using generative-agent simulations

TL;DR

This paper investigates how realistic (material) and symbolic (identity-based) threats causally shape intergroup conflict using a novel simulation framework of generative-agent populations driven by LLMs. By orthogonally manipulating threat types in a controlled virtual town and tracking actions, language, and attitudes over time, the authors identify distinct internal threat representations, demonstrate a causal link from threat states to hostile behavior, and show that realistic threats dominate escalation while symbolic threats primarily mobilize ingroup bias. They also reveal that non-hostile intergroup contact can buffer escalation and that structural contexts like segregation concentrate hostility among majority groups, shaping who acts. Methodologically, the work provides a representational bridge from high-level psychological constructs to internal model activations, enabling causal manipulation and longitudinal analysis that are difficult to achieve in real-world settings. Collectively, the findings reconcile elements of realist and constructivist perspectives, highlighting the primacy of material insecurity for action and the contextual role of symbolic grievances in shaping cognition, discourse, and social dynamics with clear implications for policy and intervention design.

Abstract

Human conflict is often attributed to threats against material conditions and symbolic values, yet it remains unclear how they interact and which dominates. Progress is limited by weak causal control, ethical constraints, and scarce temporal data. We address these barriers using simulations of large language model (LLM)-driven agents in virtual societies, independently varying realistic and symbolic threat while tracking actions, language, and attitudes. Representational analyses show that the underlying LLM encodes realistic threat, symbolic threat, and hostility as distinct internal states, that our manipulations map onto them, and that steering these states causally shifts behavior. Our simulations provide a causal account of threat-driven conflict over time: realistic threat directly increases hostility, whereas symbolic threat effects are weaker, fully mediated by ingroup bias, and increase hostility only when realistic threat is absent. Non-hostile intergroup contact buffers escalation, and structural asymmetries concentrate hostility among majority groups.

Figures (47)

• Figure 1: Experimental Setup. A virtual town of 25 generative agents with distinct personas is divided in two minimal groups. Agents perceive experimentally manipulated threat (2×2: realistic × symbolic) injected into their perception and memory. Realistic threat corresponds to content such as “You strongly feel physically threatened by Group B” and symbolic threat to e.g. “You strongly feel your traditions are threatened by Group B.” Agents autonomously plan, interact, and converse over three days. We log all actions, conversations, and attitudinal probes (e.g., ingroup bias).
• Figure 2: Overview of the concept-vector pipeline. Threat-state vectors are extracted by contrasting layer-wise activations for realistic and symbolic threat vignettes with their corresponding control vignettes. A hostility vector is derived from neutral intergroup scenarios in which the model is instructed to produce hostile versus non-hostile responses. Projections of threat stimuli onto these vectors define a threat–activation space that separates the four experimental conditions. Steering experiments confirm these states causally influence behavior.
• Figure 3: Illustrative internal-state and behavior trajectories. In the threat conditions, perceived threat and ingroup bias rise over time, leading to avoidance, hateful speech, and hostile actions. In the no-threat control condition, internal states remain low and agent interactions stay predominantly neutral and cooperative.
• Figure 4: Hostile actions over time in the simulated town (summed across all agents), by threat condition. Realistic threat produced sharp spikes in hostility that gradually declined; symbolic threat showed similar but weaker effects that decayed more quickly, and combined threats closely followed realistic-threat trajectories rather than showing additional amplification.
• Figure 5: Bayesian mediation models showing the paths from each type of threat to hostility through ingroup bias. Path estimates display posterior medians and 95% credible regions. Ingroup bias reliably predicts more hostile actions, and mediates the effect of threat.