Recent studies have revealed that AI "world models" that fail to account for human beliefs and the mental states of others are highly likely to make inaccurate behavioral predictions in the real world. This research suggests that for AI agents to possess truly practical reasoning capabilities when coexisting and collaborating with humans in complex environments, modeling the "mental states of others" is critically important, alongside an understanding of physical laws.
In this study, simulations were used to compare how predictive behavior changes when AI ignores the beliefs of others (e.g., where another person believes an object is) versus when it takes them into account. The results demonstrated that an AI unable to account for human mental states cannot infer the intentions or perceptions of people around it, leading it to select logically unnatural actions depending on the situation.
Conventional world models have primarily focused on the physical dynamics and predictability of environments. However, in interactions with human society, predicting behavior based on individual beliefs is indispensable. Equipping AI with cognitive abilities akin to the "False Belief task" represents a crucial technological milestone for enhancing AI safety and cooperativeness. Moving forward, establishing methods to integrate "theory of mind" into LLMs and generative AI architectures—thereby enabling them to accurately infer the thoughts of others—will likely become a major research and development challenge.