Large Language Models for Lane Change Decisions in Mixed Traffic for Autonomous Vehicles
TL;DR
A novel lane-change assistance framework for AVs that leverages large language models (LLMs) to enable context-aware, human-like driving decisions and demonstrates that the LLM-based approach significantly enhances operational efficiency and adaptability.
Abstract
Autonomous vehicles (AVs) face significant challenges as they become increasingly integrated into real-world transportation systems, particularly in achieving reliable and adaptive decision-making within dynamic mixed traffic environments—an essential prerequisite for their widespread and safe deployment. This paper proposes a novel lane-change assistance framework for AVs that leverages large language models (LLMs) to enable context-aware, human-like driving decisions. Unlike traditional rule-based or heuristic approaches, the proposed system employs natural language prompts—including structured traffic rules, scenario descriptors, and few-shot examples—to guide real-time decision-making. Integrated into a high-fidelity PTV VISSIM microscopic traffic simulation environment, the system dynamically interprets evolving traffic conditions and outputs lane-change decisions, which are translated into executable vehicle control actions. Multiple LLM architectures and prompting strategies are evaluated against key performance indicators, including average delay, stop frequency, collision count, response time, and inference cost. The results demonstrate that the LLM-based approach significantly enhances operational efficiency and adaptability, with one of the tested models achieving the most favorable balance between safety, performance, and computational feasibility. This study highlights the potential of language-driven intelligence as a scalable and interpretable decision layer for next-generation autonomous driving systems.