Active Inference Modeling of Socially Shared Cognition in Virtual Reality

This study proposes a process model for sharing ambiguous category concepts in virtual reality (VR) using an active inference framework. The model executes a dual-layer Bayesian update after observing both self and partner actions and predicts actions that minimize free energy. A disagreement in category judgment was added to the free energy as a risk term (corresponding to expected surprise in active inference). As the weight of this term, gaze synchrony measured by Dynamic Time Warping (DTW), assumed to re- flect joint attention, was used. The hypothesis was that higher weighting of gaze synchrony would improve prediction accuracy. To validate the model, an object classification task in VR including ambiguous items was created. The experiment was conducted first under a bot avatar condition, in which gaze was not synchronized and ambiguous category judgments were always incorrect, and then under a human–human pair condition. This design allowed verification of the collaborative learning process by which human pairs reached agreement. Analysis of experimental data from 14 participants showed that the model achieved high prediction accuracy for observed values as learning progressed. Introducing DTW as a model parameter further improved prediction accuracy, with optimal performance at synchrony weights of γ 0 = 0 . 5 - 0 . 9 . This approach provides a new framework for modeling shared social cognition using active inference.