WORLD MODELS NEED HUMAN BELIEFS TO PREDICT ACTIONS

A new "Mental World Modeling" framework addresses a critical gap in how AI systems understand the world. Existing models focus exclusively on physics and spatial dynamics, missing the human element entirely. Researchers found that even weaker language models incorporating mental variables—beliefs, desires, and intentions—outperform stronger models without this capability. The approach recognizes that predicting human actions requires understanding what people think and want, not just physical laws. The framework's main challenge lies in modeling how physical and mental states interact and evolve together. This coupling problem represents the primary bottleneck preventing wider adoption. The findings suggest future AI systems need explicit representations of human cognition to accurately simulate real-world scenarios. This shift could improve applications ranging from robotics to video generation, where predicting human behavior is essential.