World models that ignore human beliefs predict the wrong actions, new research shows

New research from MIT and the MIT-IBM Watson AI Lab shows that world models like Sora or Genie, which only simulate physics, fail to predict human actions because they ignore beliefs, intentions, and other mental states. The proposed 'Mental World Modeling' framework, which incorporates mental variables, enables even weaker language models to outperform stronger models without mental modeling, with the key challenge being the joint prediction of physical and mental state changes.