World models of environment, agent and joint agent-environment systems

arXiv:2608.20401v1 Announce Type: new Abstract: World models are a central component of modelbased reinforcement learning. They are usually discussed in terms of what variables they predict, such as observations, rewards, states, latent or information states.