World model

A language model predicts text. A world model tries to hold something closer to a mental map: what is in a space, where it sits, and what happens if something moves. The point is that a system with a world model can answer questions it was never shown, like what the back of that chair looks like, because it is reasoning about a scene rather than pattern matching on pixels.

The term gets used loosely in marketing, so it is worth knowing that researchers have argued about the definition. A 2026 proposal from an international team excluded pure text to video generators, on the grounds that they have no feedback loop with the real world. Three dimensional reconstruction and simulators, where physical rules can actually be checked, do qualify.