Predictive models are also behind many of the SOTA results in modern reinforcement learning, although they are often used to generate fictive data from which a policy is learnt.