Probing Multimodal LLMs as World Models for Driving

Dataset Analysis

MLLMs like GPT-4o struggle to understand dynamic driving scenes across frames, despite excelling at single images, highlighting gaps in world model capabilities.