Legally-Guided Automated Decision-Making System Using Language Model Agents for Autonomous Driving

Recent advances in language models have facilitated the development of agent-based systems. Despite their encouraging results in various reasoning tasks, these systems often operate as "black boxes", raising concerns about potential illegal behavior due to opaque decision-making processes. This concern is particularly critical in autonomous driving, where precise decision-making requires a thorough understanding of traffic scenes and strict adherence to established norms. In this paper, we propose a legally-guided automated decision making system (LAD) that employs language models to dynamically retrieve facts for related rules through context-based query generation while delegating decisionmaking to a symbolic solver. In our experiments, we demonstrate that this neuro-symbolic system, with a limited number of formalized traffic rules, provides a more accurate, interpretable, and traceable solution for rule-compliant decision-making compared to pure language models.