Understanding World Models?

Robots that tidy up the kitchen and fetch the medicine box are back on the horizon. Language models translate instructions into motion, a "digital twin" of the room helps them "reason in space" — and somewhere along the way the term world model is used. But what does it actually mean?

This seminar paper, written for the seminar Weltmodelle für Agentic AI: Repräsentation, Autonomie und Implementation, starts from Yann LeCun's claim that world models are the key to human-like common sense and understanding in machines. Two questions follow at once: what must something satisfy to deserve the name model, and when can we speak of understanding?

For the first, it draws on Herbert Stachowiak's general theory of models, with its features of mapping, reduction and pragmatics, which forces the question for whom something is a model and what for. For the second, it uses Christoph Baumberger's types of understanding and his notion of grasping. Against these criteria it examines the architectures of Ha & Schmidhuber and LeCun's (H-)JEPA.

The result is no simple yes or no. H-JEPA plausibly counts as a predictive surrogate of its environment, and its Mode-2 planning specifies genuinely counterfactual use of the model. What the design does not specify is a way for the dependencies it relies on to become available to the agent itself as reasons it can explain and justify. The paper then discusses three objections: that this function could simply be added, that it is not needed at all, and that humans are just as opaque to themselves.

The full seminar paper (in German) can be downloaded here.