Two ways to learn to act
Reinforcement learning broadly splits into model-based and model-free approaches. Model-free RL, the kind covered in dedicated reinforcement learning material, learns a policy or a value function directly from experience, without ever building an explicit model of how the environment works. It's often simpler to implement and can be very effective, but it typically needs a large number of real interactions to learn well, since every bit of knowledge about the environment has to come from direct trial and error.
Model-based RL, which is where world models fit in, adds an intermediate step: learn a model of the environment's dynamics first, then use that model, through planning or simulated rollouts, to decide on actions or to generate extra training data. The appeal is data efficiency. Once a reasonably good world model exists, the agent can generate large amounts of simulated experience without needing an equivalent amount of costly real-world interaction.