World model

What a world model is

A world model is a machine learning system that learns an internal representation of an environment and predicts how that environment changes in response to actions ^[1]. Rather than only mapping observations to actions, it builds something like an internal picture of how the surroundings behave, and uses that picture for planning, reasoning and acting ^[1].

The idea matters most in reinforcement learning, where an agent has to learn by interacting with an environment that may be slow, expensive or unsafe. There, a world model is a simulator learned from experience: it lets the agent simulate outcomes without trial and error in the real environment ^[1]. Every real step costs time or carries risk; a learned simulator lets the agent take many imagined steps instead.

How it works

A typical world model compresses raw observations into latent representations and predicts future representations, minimising prediction error in the embedding space ^[1]. The compression step matters: raw frames or sensor readings contain far more detail than the agent needs, so the model keeps a compact vector that captures what is relevant, and learns to predict the next such vector rather than the next frame.

Once that predictor is learned, the agent can imagine outcomes inside the learned model and plan its behaviour there, the approach used in MuZero and in the latent dynamics planning work of Hafner and colleagues ^[1].

A simple illustration: an agent that receives images of a scene pushes each image through an encoder to get a short vector. Given a candidate action, the model predicts the vector that should follow. The agent can roll several candidate action sequences forward in this space, compare the predicted vectors with what it wants, and only then act in the real environment. Nothing has to be rendered back into pixels along the way.

Where it came from

Jürgen Schmidhuber first used the term world model in 1990, using recurrent neural networks to predict future states ^[1]. In the same year, Richard Sutton proposed the Dyna architecture, which integrated learning, planning and reacting ^[1] — the same division of labour that later world-model agents rely on: learn from real experience, plan with a model, then act.

The line of work became widely known with “Recurrent World Models” by David Ha and Schmidhuber, published in 2018 ^[1]. The World Models paper, first submitted in March 2018, trains a generative neural network model of a reinforcement learning environment in an unsupervised manner to learn a compressed spatial and temporal representation ^[2]. Features extracted from that world model are then used as inputs to an agent, which lets a compact and simple policy solve the task ^[2]. The most striking part of that work was that the authors trained an agent entirely inside its own hallucinated dream generated by the world model, and transferred the resulting policy back into the actual environment ^[2].

Where it is used today

The Dreamer series by Hafner and colleagues, reported in 2019 and 2025, is one widely followed line of work in this direction ^[1]. Beyond research agents, world models have become a way to describe generative systems that produce interactive environments: DeepMind's Genie appeared in 2024, and Genie 3 offered real-time interactive worlds in August 2025 ^[1]. Meta released V-JEPA 2 in June 2025, and Waymo World Model and Nvidia Cosmos 3 followed in 2026 ^[1].

Limits

Two cautions recur. First, prediction happens in embedding space rather than by pixel reconstruction, and accuracy in video prediction does not necessarily reflect real performance ^[1]. A model can produce plausible-looking futures and still be wrong about the part that matters for a decision. Second, physical consistency remains weak: V-JEPA 2 scored close to random on IntPhys 2 ^[1]. A system that predicts well on ordinary footage may still fail at basic physical expectations.