Internal representations that allow an intelligent system to simulate how the world works, predict outcomes, and plan actions before executing them. A world model encodes relationships between objects, causes, and consequences, enabling an agent to imagine possible futures and choose strategies rather than reacting blindly to immediate inputs. In both neuroscience and AI research, the concept refers to the capacity to build a mental or computational map of reality that supports reasoning, foresight, and goal-directed behavior.
In machine learning, world models are often associated with simulation-based or model-based reinforcement learning, where an agent learns a compact internal environment to test actions cheaply and safely. The idea was popularized in AI research by work such as David Ha and Jürgen Schmidhuber, who demonstrated agents that learn latent simulations of their environments to improve planning. In contemporary discussions of advanced or agentic AI, world models are considered a foundational step toward systems that behave more like humans, since prediction and imagination are prerequisites for autonomy, strategy, and common sense.