Teaching world models to read human minds

Show a modern AI a closed door, and it can predict the exact force needed to push it open. Ask it why the person on the other side is refusing to unlock it, and the system falls apart completely.

That is the fundamental blind spot of today's world models. They track pixels, rigid bodies, and trajectories with surgical precision, but they are totally blind to what actually drives human behavior: beliefs, desires, emotions, and unspoken social rules. A model tracking only the physical scene ends up predicting the wrong action for the right-looking environment.

Researchers from the University of Oxford and the National University of Singapore published a paper proposing Mental World Modeling (MWM), a new framework that treats human mental states as first-class citizens in world models.

Giving AI an inner psychological engine

Traditional systems only care where the coffee mug landed. MWM argues that an AI must also track where each person thinks the mug is, what they want to do with it, and how they feel about it.

The framework couples physical states with five core mental variables: beliefs, goals, intentions, emotions, and social permissibility. Instead of coming up with psychological explanations after an action happens, MWM simulates the physical and mental updates simultaneously before deciding what happens next.

To make this concrete, the team built MENTIS, a training-free baseline that breaks down decision-making into transparent steps. It parses the world state, creates target-specific partial observations, breaks down candidate actions, updates physical and mental components together, and evaluates the outcomes branch by branch. Because it requires no extra model pre-training, its internal logic is completely inspectable.

The cost of stripping out empathy

To put MWM to the test, researchers built a quality-controlled benchmark featuring 448 situated decision scenarios across text, images, and sounding videos, running tests across 8 LLM-based world models.

Equipped with full mental modeling, the systems hit an average accuracy of 87.9% in predicting human decisions.

Then came the reality check. When researchers switched off the mental modeling channel and forced models to rely purely on physical trajectories, performance across every single model tanked, shedding an average of 12.1 points. In heavily interpersonal scenarios, explicit mental modeling delivered a massive 26.4-point performance leap.

Why it matters

For the past two years, the AI world has treated "world models" as synonymous with physical video simulators like Sora. But if AI agents are ever going to leave the screen and interact with people, understanding Newton's laws is only half the battle.

For developers building tutoring agents, companion robots, or interactive game NPCs, MWM offers an immediate blueprint. Because MENTIS is training-free, you do not need to wait for a hypothetical frontier foundation model to start testing agentic theory-of-mind. You can hook structured mental tracking directly into existing prompt and inference pipelines.

We think the cleverest move here was making mental variables explicit and inspectable rather than relying on an LLM's black-box intuition. You can actually see which mistaken belief led to which predicted behavior.

The risk, however, lies in scaling. Simulated scenarios with 448 clean cases are a far cry from the chaotic, irrational, and deceitful social dynamics of daily life. Tracking five structured variables works well on paper, but keeping those mental representations grounded in messy, multi-agent environments will be the real hurdle to clear.