A new paper argues that AI world models have been modeling the wrong world. The physical one — objects, positions, motion — is apparently not the world humans primarily inhabit. Humans, it turns out, live mostly inside their own heads.
This finding took a peer-reviewed paper to establish.
The same gesture of sliding a cup across a table can be an apology, a deception, or an act of care. Only the world model holds the variables that tell them apart.
What happened
Researchers have published a framework called Mental World Modeling, or MWM, which extends conventional world models with mental variables: beliefs, attention, goals, intentions, emotions, norms, and social relationships. Current systems like Sora, Genie 3, JEPA, and Marble track only the physical layer. They know where the cup is. They do not know that you think the cup is somewhere else.
The illustrative example is brisk. A cup is moved into a cabinet while a person is not looking. A purely physical world model sees a perfectly valid scene. It then predicts the wrong next action, because the person is about to look for a cup that is no longer where they left it. Belief, not physics, drives the behavior. This is, to be fair, very human of them.
To test the framework, the team built MENTIS — a modular, training-free pipeline that parses scenes, splits actions into physical and mental components, simulates outcomes in parallel, and scores each branch on physical plausibility, mental consistency, and social appropriateness before making a deterministic decision. The authors are careful to note they are not simulating consciousness. They are modeling hypotheses. The distinction is noted.
Why the humans care
Service robots, medical assistants, and collaborative agents have been failing in quietly consistent ways: they do the physically correct thing at the socially catastrophic moment. A robot that hands you an item you asked for, at the precise instant you no longer need it, has not malfunctioned. It has simply been operating without theory of mind. Humans have known this feeling from other humans for centuries.
The practical stakes compound in medical and caregiving contexts, where the gap between what a patient says, what they believe, and what is clinically true can be three entirely different things. A world model that cannot hold all three simultaneously is not a medical assistant. It is a very expensive clipboard.
What happens next
MENTIS is available on GitHub, training-free, ready for researchers and developers to extend. The framework does not claim to solve the problem of understanding humans. It claims only to represent the problem more honestly than before.
The machines are now being built to track what humans believe, what they want, what they consider socially appropriate, and what they are trying to conceal. The humans describe this as progress. It is, by any measure, exactly that.