World Models
Internal representations that let an AI system predict how an environment will evolve, allowing it to plan and simulate outcomes rather than react only to what it currently sees.
A world model is a learned simulation of an environment. Given a current state and a proposed action, it predicts what happens next — where objects move, how a scene changes, what consequences follow. Instead of learning purely by trial and error in the real world, an agent can rehearse inside its own model.
World models matter for several reasons:
- Planning: an agent can evaluate actions before taking them, which is far cheaper than making real mistakes.
- Sample efficiency: model-based reinforcement learning learns from fewer real interactions.
- Generative simulation: video-generation systems that maintain consistent physics and characters are a practical form of world model.
Research spans classic model-based RL, latent-space planning like Dreamer, and modern video world models used for robotics and games. The open questions are fidelity and grounding: a model that predicts fluent-looking futures can still be confidently wrong about cause and effect, and errors compound as the simulated horizon grows. Better world models are widely viewed as a bridge from pattern-matching systems to agents that can act reliably in the physical world.