World Models Survey — unifying architectures, training, and reasoning
arXiv drops a comprehensive taxonomy of world models — the learned simulators that let agents predict, plan, and reason inside their own representations.
• Architecture: representation format, dynamics, modality, learning paradigm
• Methodology: training approaches across RL, robotics, autonomous driving, video gen
• Reasoning: how models infer future states and inform planning
• Applications: concrete domains where world models now unlock performance