Why Sora fails the cup test: world models need theory of mind

a new research paper proposes mental world modeling (mwm), a framework extending world models to include mental variables like beliefs, intentions, and social norms, which current models like sora and genie ignore. the authors demonstrated that purely physical world models predict incorrect actions in scenarios where human beliefs diverge from reality, using an example of a moved cup to highlight the gap. to validate the theory, they released mentis, a training-free pipeline that scores actions based on mental consistency and social appropriateness, alongside the menti-bench dataset of 448 decision scenes. evaluation on eight leading llms, including versions of gpt and claude, showed that existing models struggle with these tasks, suggesting that more compute cannot replace the need for explicit mental modeling in autonomous agents.