World Model
A learned simulator of how an environment evolves — imagine outcomes before acting in the real world.
A world model learns the dynamics of an environment — given a state and an action, predict what happens next — so an agent can plan by imagination instead of trial and error. Long central to robotics and RL, world models became a headline research direction as video-generation models showed they implicitly learn physics, and as agent builders sought something beyond next-token prediction.
Worked example: a model’s learned internal representation of how an environment works — its dynamics — so it can predict the consequence of an action and plan ahead rather than only react. Gotcha: whether large models hold a genuine world model (vs. surface pattern-matching) is genuinely debated; the practical test is generalization to unseen situations — a real world model transfers, a memorized correlation breaks off-distribution, which is why held-out and counterfactual evals matter.