world model

advancedexperimental

A world model is an AI system's internal picture of how its environment works, which lets it predict what will happen after its next action.

In plain terms

When you step over a puddle, you do not calculate with physics formulas: you have a picture in your head of how the water, your shoes, and your own body behave. We call this internal picture a world model. The machine world model is the same: a learned, simplified 'mockup' of the world on which the system can mentally try out what would happen if.

It lets a machine do more than react to the current situation: it can think ahead. A robot or an AI agent can play through several possible steps in its world model and carry out only the most promising one in reality. This is cheaper and safer than trial and error in the real world.

This differs from a standard conversational language model, which learns primarily from text and estimates the next word. A world model estimates how the situation itself changes: where the ball rolls, what the camera sees after the turn, or what happens if the robot arm lets go of the glass.

How it works

Observation: the system sees vast amounts of video, sensor data, or gameplay and watches what follows what.
Compression: filters the essence from raw images (what is where, where it moves) and stores this in a compressed internal description.
Forecasting: it learns to estimate how this internal description changes over time or when it takes action itself.
Imaginary rehearsal: before making a decision, the system plays through several possible sequences of steps in the model and compares the expected outcomes.
Feedback: many systems compare the result of a step actually taken in reality with what was expected, and learn further from the discrepancy.

An everyday example

A car equipped with an autonomous driving function sees a ball roll out from between two parked cars. Based on its world model, the system expects that a child might run out after the ball, so it slows down before seeing anyone.

Who is affected

industrial worker, engineerFactory and warehouse robots are increasingly trained in simulated environments before being placed next to real production lines. In addition to traditional physical simulators, learned world models are increasingly used for this. → Industry

developerIf you build an agent, its reliability largely depends on how accurately it can anticipate the consequences of its steps. → Code

entrepreneur, executiveAn increasing share of robotics, autonomous driving, and simulation developments builds on this technology, so it is worth knowing what it promises and what it does not.

everyoneYou will hear this word more and more in the news, whenever people argue over whether the machine "understands" the world.

A common misconception

MythIf an AI has a world model, it understands the world as humans do.

In realityA world model is a predictive system learned from data. It is good at things similar to what it has seen many times, but it can be confidently wrong in unusual situations. This alone proves nothing about whether the machine "understands" the world in the human sense.

Limits and risks

The model is only as good as its training data: what it rarely or never saw, it predicts poorly.
Errors accumulate: the more steps the system simulates in advance, the further the imagined future may drift from reality.
A spectacular, lifelike video does not mean precise physics; in the generated world, objects may disappear or behave irregularly.
Among researchers, there is still debate over whether large language models develop a true world model, so treat confident statements on this matter with caution.

Similar, but not the same

large language model – The language model predicts text continuation, while the world model predicts how the state of the environment changes.
digital twin – A digital twin is a copy of a specific machine or facility, continuously updated with real measurement data and usually built on physical models provided by engineers. The world model is learned by the system itself from data and is not tied to a single specific object.
video generation model – The goal of a video generator is convincing moving images. We expect a world model to keep the scene consistent and respond correctly to your interventions. The line is blurred because some video models are also used as world models.

Related concepts

Broader

artificial intelligence, machine learning

Narrower

interactive world-generation model, model-based reinforcement learning

On the edge

reinforcement learning, AI-agent, digital twin, simulation, embodied AI

A brief history

The idea originates from psychology: Kenneth Craik wrote in his 1943 book "The Nature of Explanation" that the mind carries a small-scale model of the world. In machine learning, learning systems that model their environment appeared as early as the 1990s, but the term was made widely known by David Ha and Jürgen Schmidhuber's 2018 paper "World Models." In the 2020s, Google DeepMind, NVIDIA and Yann LeCun's research group also joined the leading world-model research efforts.

Where things stand

As of late 2026, world models are used mainly in research labs, in robotics and autonomous-driving development, and in tools that generate interactive 3D worlds. There are already first commercial and open models, for example for training robots or creating 3D scenes. Many companies and researchers see this as the next big step, but no reliable, general-purpose world model exists yet.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

DeepMind unveils Genie 4 world model with more realistic physicsOct 25, 2026 · Agent Age
OneTrack unveils Iris-1 industrial world modelOct 23, 2026 · Ipar

Also known as

world model · world model · world model · world models · internal world model · world foundation model · generative world model · world simulator

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.