MiniToM: A Goal, Belief and Desire Predictor with Theory of Mind Modeling

Yizhe Huang, Hulingxiao He, Jinfan He, Yao Yuanhang

Keywords: Theory of Mind, Inference, POMDP, Goal, Desire, Belief

Abstract: Theory of Mind (ToM), a critical aspect of human cognitive reasoning, holds significant importance for the integration of artificial intelligence systems in social interaction scenarios. The computerization of this capability has been the subject of extensive research. We introduce MiniToM, a compact theory of mind system capable of observing an agent's action trajectory within a specific environment. MiniToM operates on the assumption that the agent possesses three mental states – goal, state belief, and desire – and conducts real-time inference on these mental states. Leveraging primarily the Bayesian method and stochastic gradient descent, MiniToM achieves inference with a minimal number of parameters. Our experimental results demonstrate the system's ability to update the agent's mental state using limited data while maintaining high accuracy. Furthermore, this framework provides a platform for exploring the accuracy, computational efficiency, and generalization capabilities of the ToM system through adjustments to the MiniToM module.

Submission Number: 230