Large language model-driven bi-level game framework for connected and automated vehicle pair at mixed unsignalized intersections

As Connected and Automated Vehicles (CAVs) become core Internet of Things (IoT) terminals in modern transportation, growing research has focused on safe driving strategy at unsignalized intersections. However, existing studies often neglect how multiple CAVs can cooperate to improve safety, and rarely address traffic fairness under stochastic Human-Driven Vehicle (HDV) behaviors. To address these issues, we propose a novel large language model (LLM)-driven bi-level game framework for CAV Pair at mixed unsignalized intersections, namely BiG-LLM. This framework combines semantic scene understanding via LLM with verifiable game-theoretic decision-making, effectively alleviating the hallucinations caused by pure LLM. A CAV Pair-based mixed platoon is introduced to exploit multi-CAV synergy. In the bi-level game strategy, the upper-level game enables the lead CAV to optimize the allocation of right-of-way. The tail CAV activates the lower-level fairness game, using the waiting anxiety model based on prospect theory to quantify frustration and discourage extreme waiting. Extensive simulations are conducted under multiple traffic flow conditions. The results demonstrate that BiG-LLM consistently achieves a high success rate, improves safety by increasing the minimum Post-Encroachment Time (PET), and reduces the maximum waiting time compared to baseline methods, while maintaining competitive efficiency. These results verify the effectiveness of BiG-LLM in balancing efficiency, safety, and system-level traffic fairness at unsignalized intersections.