Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

World Model

nounEmerging AI Researchalso world models in AI, learned environment model, predictive model of the world

In one line

A world model is an internal model, learned or built by an AI system, of how an environment works and changes, which lets the system predict the consequences of actions, imagine possible futures and plan before acting.

A world model is an internal representation that an intelligent system uses to predict how its environment will change, especially in response to its own actions. People rely on such models constantly. You know that a dropped glass will fall and break, and you can imagine the route to a store without walking it. In artificial intelligence, a world model is a learned or engineered component that lets a system simulate possible futures in its head, so that it can choose better actions. The idea connects reinforcement learning, embodied AI, robotics and AI reasoning.

Definition and basic idea

A world model typically answers a question of the form: given what I currently observe and an action I might take, what will I observe next, and what reward or outcome will follow? Once a system has this predictive ability, it can plan by running imagined rollouts instead of trying every option in the real world. This matters because real experiments can be slow, costly or dangerous, as in driving or surgery.

The concept has roots in psychology, where Kenneth Craik proposed in 1943 that the mind builds small-scale models of reality to anticipate events. In control theory and robotics, explicit models of dynamics have long been used for planning. What distinguishes modern world models is that they are often learned from data, using deep learning, rather than written by engineers.

History and key milestones

  • Model-based reinforcement learning. Early work in the 1990s, such as Richard Sutton’s Dyna architecture, combined real experience with simulated experience from a learned model.
  • World Models, 2018. David Ha and Jurgen Schmidhuber showed that an agent could learn a compact model of a game environment from pixels and train a controller largely inside its own imagined version of the world [1].
  • Dreamer and related agents. Danijar Hafner and colleagues developed agents that learn a latent world model and plan by imagining outcomes, and later versions mastered many different tasks with a single set of settings [2].
  • MuZero, 2019 to 2020. DeepMind’s system learned a model that predicted only quantities useful for planning, such as value and reward, and reached strong play in Go, chess and Atari games without being told the rules.
  • Video and interactive world models. From the 2020s, large generative models of video and games have been proposed as general simulators that respond to actions, giving rise to the idea of world models as a basis for robotics and virtual environments.

How world models work

### Learning a latent state

Raw observations such as camera images are large and noisy. Many world models first compress them into a compact latent state that keeps the information needed for prediction, using an encoder network, for example a variational autoencoder.

### Predicting dynamics

A second component predicts how the latent state changes over time given an action. Recurrent networks, transformers and state space models are common choices. The model may predict one step or long sequences.

### Predicting rewards and outcomes

For decision making, the model also predicts rewards, costs or goal progress, so that imagined trajectories can be scored.

### Planning and learning in imagination

The agent can plan by searching over imagined action sequences, or it can train its policy on simulated experience, which reduces the amount of real interaction needed. See planning for related methods.

### Joint embedding approaches

Yann LeCun has argued that learning world models is a key missing ingredient for human-level AI, and has proposed the Joint Embedding Predictive Architecture, which predicts abstract representations rather than every pixel [3]. The aim is to focus on predictable structure and ignore unpredictable detail.

World models and large language models

Large language models learn from text and can describe how things work, including physical and social situations, and some researchers suggest that predicting text forces them to build partial models of the world. Studies of models trained on board game moves, for instance, found internal representations of the board state. Critics argue that language models lack grounded, reliable models of physical causality and that their errors in simple spatial and physical reasoning show limits; see large language models. The debate shapes research on spatial intelligence, physical AI and multimodal reasoning.

Applications

  • Robotics: planning grasps and motions, and training in simulation before real deployment; see embodied AI.
  • Autonomous driving: predicting the behavior of other road users and generating rare scenarios for testing; see AI in transportation.
  • Games and virtual worlds: generating interactive environments and non-player characters; see AI in entertainment.
  • Scientific simulation: learned emulators of weather, fluids and molecules; see AI in climate science.
  • Decision support: forecasting the effects of interventions in business and policy settings.

Benefits

  • Sample efficiency, because imagined experience supplements real experience.
  • Safer exploration, since risky actions can be tried in simulation.
  • Support for long-term planning and counterfactual thinking.
  • Reuse across tasks, since one model can serve many goals.

Challenges and limitations

  • Compounding errors. Small prediction errors build up over long imagined rollouts, so plans may drift from reality.
  • Model exploitation. A planner may find actions that fool an imperfect model but fail in the real world.
  • Partial observability and rare events. Models trained on limited data may miss unusual situations; see model drift.
  • Evaluation. There is no agreed test of whether a system has a good world model, and visually realistic video does not prove an understanding of physics.
  • Compute cost. Training large generative world models requires heavy resources.
  • Safety and verification. Decisions based on an inaccurate model can cause harm in physical settings; see AI safety.

A worked example

Imagine a warehouse robot that must place boxes on a shelf without tipping the stack. It first learns from camera data how boxes move when it pushes or lifts them, building a latent model of the scene. When given a new pile, it imagines several placement sequences, and the model predicts that placing the heavy box on top of the small one would topple the stack. The robot picks a sequence that the model scores as stable, carries it out and compares the real result with its prediction. When the outcome differs, for example because a box is slippery, the error is used to update the model. Over time the robot needs fewer real trials and avoids many failures.

FAQs

What is a world model in AI? It is an internal model that lets an AI system predict how its environment will change, so it can plan and choose actions.

Do large language models have world models? They appear to learn some structure about the world from text, but experts disagree about how reliable or grounded it is.

Why are world models important for robots? They let robots practice and plan in imagination, which is cheaper and safer than trial and error in the real world.

What is JEPA? It is Yann LeCun’s proposed architecture that predicts abstract representations of future states rather than raw pixels.

Are video generators world models? Some researchers see them as steps toward world simulators, but realistic video does not guarantee correct physics or reliable control.

Key facts

  • David Ha and Jurgen Schmidhuber published a 2018 paper titled World Models showing that an agent could learn a compact model of an environment and train a controller largely inside its own imagined dream1.
  • The Dreamer family of agents, introduced by Danijar Hafner and colleagues, learns a world model from images and plans by imagining future outcomes in a compact latent space2.
  • Yann LeCun has argued since 2022 that learning world models is a key missing ingredient for human-level AI and has proposed the Joint Embedding Predictive Architecture, or JEPA, as an approach3.

Related concepts

Quick checkWhat is a world model in AI?Show answer

A world model is an internal model, learned or built by an AI system, of how an environment works and changes, which lets the system predict the consequences of actions, imagine possible futures and plan before acting.

Sources

  1. Ha and Schmidhuber, World Models, arXiv, https://arxiv.org/abs/1803.10122
  2. Hafner et al., Mastering Diverse Domains through World Models, arXiv, https://arxiv.org/abs/2301.04104
  3. LeCun, A Path Towards Autonomous Machine Intelligence, OpenReview, https://openreview.net/forum?id=BZ5a1r-kVsf

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue