Decentralized Collective World Model for Emergent Communication and Coordination

TL;DR

This work tackles the challenge of enabling decentralized multi-agent systems to simultaneously learn shared symbol systems and coordinate under partial observability. It proposes a CPC-based, fully decentralized world model where two agents exchange bidirectional messages and optimize a collective free-energy objective that includes a collective regularization term, with CR approximated via a Product-of-Experts approach and message alignment via InfoNCE. The approach yields environment-general symbolic representations that reflect global environmental dynamics and improve coordination, especially when agents have heterogeneous perceptual capabilities. However, the coordination gap relative to a centralized or fully observable baseline highlights limitations in policy learning (demonstrated via imitation learning) and points to future work incorporating reinforcement learning, active inference, and scalability to larger agent teams. Overall, the results demonstrate that decentralized communication can foster meaningful symbol emergence while supporting coordination in dynamic, partially observable environments.

Abstract

We propose a fully decentralized multi-agent world model that enables both symbol emergence for communication and coordinated behavior through temporal extension of collective predictive coding. Unlike previous research that focuses on either communication or coordination separately, our approach achieves both simultaneously. Our method integrates world models with communication channels, enabling agents to predict environmental dynamics, estimate states from partial observations, and share critical information through bidirectional message exchange with contrastive learning for message alignment. Using a two-agent trajectory drawing task, we demonstrate that our communication-based approach outperforms non-communicative models when agents have divergent perceptual capabilities, achieving the second-best coordination after centralized models. Importantly, our decentralized approach with constraints preventing direct access to other agents' internal states facilitates the emergence of more meaningful symbol systems that accurately reflect environmental states. These findings demonstrate the effectiveness of decentralized communication for supporting coordination while developing shared representations of the environment.

Figures (7)

• Figure 1: Overview of the proposed method. Each agent perceives a partial region of the environment, while complementing knowledge of other regions through communication. This leads to the emergence of a symbol system that represents collective knowledge.
• Figure 2: The architecture of the collective world model. Bidirectional arrows indicate losses for training.
• Figure 4: Comparison of coordination achievement across learning conditions. The values represent the average of maximum cross-correlation between trajectories drawn by agents and test data. Error bars represent standard deviation.
• Figure 5: Comparison of coordination achievement with (w/ com) and without (w/o com) communication through message exchange using the EC (proposed method) model. Error bars represent standard deviation.
• Figure 6: Similarity between the structure of inferred messages when reconstructing test data observations and the structure of the actual trajectory of point $P$, as calculated by RSA. Error bars represent standard deviation.