Model Heterogeneity Attenuates Escalation Dynamics among Large Language Models in War Games

AI systems are increasingly integrated into the military and diplomatic decision-making as intelligent agents for strategic operations. However, existing evaluation frameworks for the escalation risk of such agentic systems often assumes that all agents are deployed with the same underlying model. This assumption is unlikely to hold in practice, as different countries already contract with different AI companies for military applications, making findings from homogeneous-model simulations difficult to transfer to real-world use. In this study, we evaluate nation agents' escalatory behavior in a wargame simulation where each nation is paired with a different model. We find that a model's extreme escalatory behavior is substantially attenuated once mixed with other models. Qualitative analysis of reasoning transcripts suggests a possible explanation: under homogeneous deployment, nations mirror each other's escalatory tone, reinforcing the conflict, which rarely emerges when different models interact. Our study suggests that a model's escalation risk depends not only on itself but on which other models it is deployed alongside.