World models that ignore human beliefs predict the wrong actions, new research shows
Weaker models with mental state reasoning beat stronger physics-only simulators. That changes how we build world models.
Why it matters
A new capability gap has opened: world models that ignore human beliefs, intentions, and desires systematically predict the wrong actions. The Mental World Modeling framework shows that adding psychological reasoning—even to smaller models—outperforms larger models without it. This is a fundamental shift in how frontier labs will need to approach multimodal and embodied AI.
The key facts
5 to knowCurrent world models (Sora, Genie) simulate physics only; ignore beliefs, intentions, desires
Mental World Modeling framework adds mental variables (beliefs, intentions, wants) to world models
Weaker language models using Mental World Modeling outperform stronger models without mental reasoning
Key bottleneck identified: predicting how physical and mental states change together
Implication: agents and embodied systems need theory-of-mind reasoning to predict human action correctly
Go to the source
The Decoderthe-decoder.com
Publisher excerpt: Current world models like Sora or Genie only simulate physics and ignore what people think, want, or feel. The new "Mental World Modeling" framework adds mental variables like beliefs and intentions. Even weaker language models using this approach outperform stronger models without mental modeling.…