LLM Bargaining for Social Interactions of Driving Agents in Mixed-Autonomy Traffic

Recent advances in Large Language Models (LLMs) provide a powerful and flexible computational mechanism to model behaviors and interactions of social agents. In this paper, we develop a new LLM bargaining framework to model social interactions of heterogeneous driving agents in mixed-autonomy traffic. We are interested in understanding how well LLM can represent social interactions of driving agents and whether the results are plausible from domain perspective. To answer these questions, we conduct controlled experiments with Gemma3-12B and examine the bargaining dynamics and outcomes in detail. The results show that: 1) the LLM bargaining framework can generate social interactions of driving agents explainable and aligned with domain knowledge; 2) physical state, agent personality, and in-context domain knowledge all influence the bargaining process and system equilibrium.