World models that ignore human beliefs predict the wrong actions, new research shows
decoder ·
New research argues that current world models for AI agents, such as Sora, Genie 3, JEPA, and Marble, only simulate the physical layer of the world—objects, positions, and motion—while ignoring the beliefs, intentions, and social norms of people in that world, leading to incorrect action predictions. The authors propose a framework called "Mental World Modeling" (MWM) that extends classic world models with mental variables and couples physical and mental state transitions. They built MENTIS, a training-free pipeline that splits actions into physical and mental components, simulates both kinds of states in parallel, and scores options by physical plausibility, mental consistency, and social appropriateness. Tested on a new benchmark called Menti-Bench with 448 decision scenes, the MWM pipeline achieved an F1 score of 87.9 versus 63.3 for direct answers, while removing either the mental or physical channel caused large performance drops. The biggest remaining gap to human performance (98.5) was traced to errors in simulating state transitions, identifying that as the key bottleneck. The paper also notes that the field of world models still lacks a consensus definition, with debates involving DeepMind, Meta, and Peking University researchers.
decoder