World Models

World models learn interactive, action-conditioned environment dynamics. While often framed rigidly as a specific architecture triad, the components are best understood as a highly useful conceptual decomposition rather than a universal law of physics.

Architecture Triad

Historically (following the Ha & Schmidhuber lineage), world modeling is decomposed into distinct parts:

• Observation vs. State: A Representation Model (V) maps high-dimensional observations (raw pixels) into a compact, abstract latent state. This forces the system to disregard noise and track core structural variables.
• Transition Model (M): A sequence model that predicts the next latent state given the current state and a user action.
• Reward/Policy Model (C): Uses the predicted states to optimize actions or estimate task success.

Prediction vs. Planning

Generic video models predict plausible pixels. In contrast, true world models enable planning. Planning involves optimizing a sequence of actions against future rewards. By predicting the consequences of actions in latent space without rendering pixels, an agent can "dream" millions of rollouts to find an optimal policy (e.g., DreamerV3). They are not faithful simulators; they hallucinate and collapse out-of-distribution.