World models and simulation concept
What is a world model?
ReferenceWhat is a world model? It is a learned model of an environment that helps an agent estimate what may happen next, including how an action could change what it observes. An agent can use those predictions to consider possible futures before acting in the environment.
Reviewed 2026-10-09
A learned picture of an environment
In reinforcement-learning research, the “world” is the environment an agent observes and acts in; the model is a learned representation of relevant parts of that environment. It can summarize past observations and make predictions about what may happen next. The representation may be compact and hidden, rather than a detailed picture of every object or event. Some world models also predict rewards, which are signals used to evaluate outcomes. The details vary by system, but the central idea is to learn from experience so an agent can reason about possible outcomes.
Learn, imagine, act, and learn again
An agent collects experience by observing the environment and taking actions. A learned model uses that experience to estimate how the environment changes, and may also estimate the reward associated with a predicted state. The agent can run possible action sequences through the model and use the predicted outcomes to help choose what to do. In some systems, these imagined sequences are represented as compact internal states rather than generated images. The agent then acts in the real environment, gathers new experience, and can update its model. A model’s predictions are estimates; they depend on what the system learned from experience.
The model supports planning or learning a policy
A world model is part of a model-based approach to reinforcement learning: it predicts aspects of the environment, while a planner or policy uses those predictions to select actions. A policy is a rule for choosing an action from what the agent has observed. One approach searches through possible future action sequences and revises its plan as new observations arrive. Another learns behavior by imagining trajectories in the model’s compact state space. In either case, the model and the action chooser have related but distinct jobs: the model predicts possible outcomes; the planner or policy chooses how to act.
Questions people ask
Does a world model have to generate images?
No. Some research systems learn compact internal states and predict within that representation; they do not need to generate a full image for every imagined step.
How is a world model different from a language model?
They are described by different prediction tasks. For example, OpenAI’s 2019 GPT-2 page describes a language model trained to predict the next word from previous words. A world model in reinforcement learning predicts aspects of an environment, often conditioned on actions, so an agent can estimate possible outcomes.
Does a world model choose the action?
Not necessarily. The model predicts possible environment outcomes; a planner or learned policy can use those predictions to choose an action.
Is a world model a complete copy of the real world?
No. The cited reinforcement-learning systems learn representations and predictions for an agent’s tasks. A compact representation is intended to support prediction or action selection, not to reproduce every detail of the environment.
Learn more
- World Models
Read an early research example of a learned environment representation used by an agent.
- Dream to Control: Learning Behaviors by Latent Imagination
See how an agent can learn behavior by predicting trajectories in a compact world-model state space.
What changed
Sources
Part of World models and simulation: Learned models for prediction and planningSimulation, models, and digital twins