牛津大学和新加坡国立大学研究团队提出了「心智世界建模」(Mental World Modeling, MWM)理论框架[1]。与仅追踪物理轨迹的现有方法不同，这一框架将信念、目标、意图、情绪和社会规范等心智变量纳入世界模型的核心，使AI能够理解人类的心理状态[1]。
研究团队开发了MENTIS基线实现和Menti-Bench数据集来验证该框架的有效性[1]。在Menti-Bench数据集上的测试结果显示，8个主流世界模型在移除心智通道后无一例外地性能退化，充分说明显式建模心智状态是准确预测人类决策的必要条件[1]。
该框架的实际应用范围广泛，涵盖服务机器人、医疗辅助、协作智能体、具身协作、照护与建议、教育训练以及交互式智能体等多个人类相关场景[1]。
Researchers from the University of Oxford and the National University of Singapore have introduced Mental World Modeling (MWM), a theoretical framework that extends artificial intelligence's understanding of the world beyond physical properties to encompass mental states [1]. The framework integrates mental variables such as beliefs, goals, intentions, emotions, and social norms alongside physical state information, enabling AI systems to comprehend human psychological conditions rather than merely tracking physical movements [1].
The research team released MENTIS, a baseline implementation of the framework, along with the Menti-Bench dataset, which was used to evaluate eight mainstream world models [1]. Testing revealed that explicit modeling of mental states is necessary for predicting human decision-making, with all eight models showing performance degradation when the mental modeling component was removed [1]. The framework has potential applications in service robotics, medical assistance, collaborative agents, embodied collaboration, care and advisory systems, educational training, and interactive intelligent agents [1].
