World models have long served as predictive tools for planning and action, but a new paper from Hugging Face argues that they miss a critical dimension: the hidden mental states that drive human behavior. The paper, titled "Mental World Modeling," proposes a framework that treats mental variables—beliefs, intentions, emotions, and social norms—as core components of a world model, rather than after-the-fact explanations.
The authors illustrate the problem with a simple observation: two physically identical scenes can lead to completely different actions if the people involved have different beliefs or goals. A world model that only tracks physical objects and their dynamics may predict the wrong action for a scene that looks right. To address this, they introduce MWM, which maintains a coupled physical-mental world state and simulates how candidate actions update both components.
To test their framework, the researchers built MENTIS, a training-free and fully inspectable baseline that decomposes decision prediction into explicit stages: state parsing, target-observation generation, action decomposition, coupled transition simulation, and branch-level evaluation. They created a quality-controlled dataset of situated decision scenarios across text, images, and videos, and evaluated eight modern LLM-based world models.
The results show that explicitly modeling mental states is essential for predicting human decisions. Full MWM achieved the strongest performance across all models, while removing the mental channel degraded performance. The gains were largest in interpersonal situations, where hidden beliefs and social norms play a decisive role. The authors also identified transition simulation—predicting how an action changes the coupled physical-mental world—as the biggest remaining bottleneck.
The paper emphasizes that a physically possible action is not necessarily the one a person will take. The same movement can represent help, pressure, deception, or trust, depending on the mental and social context. This distinction is crucial for embodied assistants, human-AI collaboration, education, and any AI system operating around people. The authors note that MWM does not claim direct access to private consciousness; mental states are treated as approximate, task-relevant hypotheses that must remain inspectable and responsibly used.