World Models

In the field of artificial intelligence, a "world model" is a model that can characterize the state of an environment or the world and predict transitions between states. This model enables agents to learn in a simulated environment and transfer learned strategies to the real world, thereby improving learning efficiency and reducing risks.

In 2018, Jürgen Schmidhuber and David Ha published a paper titledRecurrent World Models Facilitate Policy Evolution"The world model is mentioned in the article. It has the ability to understand and simulate the environment, learn behavioral strategies, and transfer learned knowledge to new situations. It can predict future sensory data based on current motor actions.

Yann Lecun defined the “world model” on the X platform in February 2024 as follows: A world model is a prediction system based on sequence data that processes observations, previous states, actions, and latent variables through encoders and predictors to predict the next state. The autoregressive generative model is a simplified form of it, using an identity encoder and discrete states without considering the encoder collapse problem. In March of the same year, Lecun’s research team published a paper titled “Learning and Leveraging World Models in Visual Representation Learning" introduced the concept of "Image World Models (IWM)", which is based on the JEPA architecture and extends the potential restoration technology and photometric transformation.

With the continuous advancement of artificial intelligence technology, "world model" is expected to become a key tool for intelligent agents to understand complex environments, predict future events, learn effective strategies and apply them to the real world.