A Multi-Agent World Model for Distributed Learning Based on Collective Predictive Coding

Abstract

Through physical and social interaction, humans form cognition and symbol systems, also sharing symbols and observational information without directly referencing others' perceptions or thoughts. The formation of internal representations through the perception–action loop is discussed under the free-energy principle and world models. To extend these ideas to the social domain, Taniguchi et al. regard symbol systems as external representations and propose a collective world model based on Collective Predictive Coding and distributed Bayesian inference. However, how each agent learns this model in a distributed way remains unexplored. This work proposes a multi-agent world model that can be trained in a distributed, end-to-end manner using the InfoNCE Naming Game to align message inference distributions. We trained agents on observation and action data from cooperative behaviors, then evaluated mutual information between formed messages and environmental states. The results confirm that messages reflecting the environment emerge as a form of collective intelligence.