World Model
An internal representation learned by an AI system that simulates how the environment works, enabling prediction, planning, and reasoning about consequences.
Concept
A world model allows an agent to 'imagine' the outcomes of actions before taking them. Instead of trial-and-error in the real world, the agent can plan by simulating scenarios in its internal model.
In LLMs
There's active debate about whether LLMs develop world models through next-token prediction. Evidence suggests they learn some internal representations of facts, spatial relationships, and even game states -- but the extent and nature of these representations remains contested.
Applications
Robotics (simulate physical interactions before acting), game AI (plan moves ahead), autonomous driving (predict other drivers' behavior), and video generation models (which must implicitly model physics).