World-model data glossary
World model
Definition
A world model is a learned representation of how an environment changes over time. Depending on its design, it can predict future observations, rewards, terminal conditions, or compact latent states. An agent can use those predictions to compare possible futures before acting. The label describes a function, not one architecture: a model can operate in pixels or latent space, use deterministic or probabilistic predictions, and receive actions as inputs or learn only from passive sequences.
Why it matters
Dataset requirements follow the transition the model must learn. Passive video can teach visual continuity and recurring dynamics, but it may not identify which decision caused a change. A controllable system usually needs temporally aligned observations, actions, state, events, and outcomes, plus enough coverage to distinguish alternative futures. Evaluation should test more than visual plausibility. It should measure action sensitivity, event timing, long-horizon consistency, uncertainty, and performance on held-out environments or action sequences.
Example
Given recent game frames, selected engine variables, and a proposed controller sequence, a world model predicts the next visual states and task events. A policy can compare several candidate sequences inside those imagined rollouts before choosing an action. If the same future appears regardless of the controls, the system may be a useful video predictor but it has not demonstrated a reliable action-conditioned interface.
Primary sources
Continue
Apply this definition to dataset scope, fields, rights, and evaluation.
World-model dataset requirements