World model

A world model is an internal representation that an AI system builds and uses to simulate how its environment works. It captures the rules, dynamics, and causal relationships that determine how states change in response to actions. Instead of relying only on direct interaction with the real world, an agent with a world model can mentally "roll out" hypothetical sequences of events. It can predict what would happen if it took a given action before acting. This makes world models central to model-based reinforcement learning, where a learned model of the environment is used to plan, generate synthetic training data, or both.

In practice, world models are typically learned from experience. The agent observes transitions in its environment and trains a neural network or other function approximator to predict future observations, rewards, or latent states. Recurrent neural networks, variational autoencoders, and transformer-based sequence models have all been used to build world models that compress high-dimensional sensory input into compact, predictive representations. Systems such as Dreamer and Ha & Schmidhuber's "World Models" paper showed that agents could learn entirely inside their own imagination, training a policy against the world model rather than the real environment.

World models are useful because real-world interaction is often expensive, slow, or dangerous. A robot learning to manipulate objects, an autonomous vehicle navigating traffic, or a game-playing agent exploring a complex environment can all simulate consequences internally before acting. World models also connect to broader ideas in AI safety and alignment. An agent that accurately models its environment can reason about long-term consequences and avoid unintended side effects.

Beyond reinforcement learning, world models have become a unifying concept in discussions about general intelligence. Researchers argue that robust generalization, the ability to adapt to novel situations, requires something like a world model rather than purely reactive, stimulus-response behavior. This framing connects machine learning to cognitive science, where internal models of the environment are considered fundamental to how biological agents plan, imagine, and reason about counterfactuals.