Dynamic Bargaining of LLM Driving Agents in Mixed-Autonomy Traffic

Mixed-autonomy traffic consists of heterogeneous driving agents, who compete for road space with interdependent decisions driven by AI. Understanding the dynamics of agent interactions is fundamental to design of efficient autonomous driving systems. Previous studies show that collective rationality (CR) can emerge in such systems even when all driving agents are self-interested. However, these results either assume the agents are perfectly rational or endowed with an explicit reward function. In this paper, we relax the assumptions and ask whether CR can emerge when the interactions of driving agents are driven by large language models (LLMs), considering implicitly defined objectives and information asymmetry. To answer this question, we propose a new LLM-driven multi-agent model, conduct controlled experiments, and examine the bargaining dynamics and outcomes in detail. The results show that: 1) CR still emerges from the interactions of LLM driving agents, aligned with findings from game-theoretic and deep reinforcement learning models; 2) bargaining dynamics and observed equilibrium are influenced by physical state, agent personality, and in-context domain knowledge.