A world model is an AI system’s internal simulator: given the current situation and a possible action, it predicts the next situation. That lets an agent “imagine” outcomes before acting, and even practise inside its own predictions.
The idea was popularised by David Ha and Jürgen Schmidhuber’s 2018 paper “World Models”. You now meet it in robotics and video generation, including “world-action models” that predict both the next action and the resulting scene.