Reason--Imagine--Act: Closed-Loop LLM Decision Making with World Models for Autonomous Driving
Researchers have developed a new closed-loop framework called Reason--Imagine--Act (RIA) to enhance the safety of large language models (LLMs) in autonomous driving. RIA integrates an LLM reasoner with an action-conditioned world model to perform online safety verification of proposed driving actions. The system iteratively proposes actions, simulates their outcomes using the world model, and selects the safest executable option, providing feedback for subsequent reasoning steps. In simulations, RIA demonstrated improved route completion and significantly reduced collision rates compared to existing methods. AI
IMPACT Introduces a novel closed-loop framework to improve LLM safety in autonomous driving by integrating world models for real-time verification.