Efficacy of Autonomous Vehicle’s Adaptive Decision-Making Based on Large Language Models Across Multiple Driving Scenarios

Understanding how large language models (LLMs) generalize across diverse traffic scenarios is critical for advancing autonomous driving systems. While previous studies have validated LLMs’ potential in specific driving tasks, evaluations of their scenario adaptability remain limited. This research adopts the Dilu framework as a case study, with the objective of investigating the generalisation performance of LLMs in five typical scenarios: basic highway sections, highway merge area, intersection, racetrack, and roundabout, with varying traffic parameters. Through extensive experiments with 17 configurations in scenarios metioned above, we employ success rate (SR) and success steps (SS) as metrics to quantify LLMs’ generalization capabilities in different driving scenarios. The results reveal significant scenario-dependent performance variations: the LLM achieves a peak SR of 99% at 30 m/s in low-speed merges but declines to 69% at 60 m/s. In intersection scenarios, the LLM outperforms traditional reinforcement learning methods (DQN, PPO) by about three times (61% SR vs. 24% SR). Furthermore, expanding memory entries from 2-shot to 5-shot enhances median SS by 114% in roundabouts and 69% in intersections, highlighting the role of experience accumulation in dynamic environments. These findings provide empirical evidence for LLMs’ scenario-aware generalization capabilities and offer actionable insights for optimizing their deployment in real-world autonomous driving systems.