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

Researchers have developed a Mental World Modeling framework that incorporates human beliefs and intentions into the simulation process, addressing a limitation where current models like Sora and Genie only track physical interactions. By accounting for the mental states behind human actions, this approach allows smaller language models to surpass more powerful systems that lack this context. This shift suggests that understanding psychological drivers is as critical as physical realism for creating more accurate and predictable AI simulations.