Dreamer-CPC enables 5x better multi-agent coordination via world model messaging

A new paper on arXiv (2607.19809) introduces Dreamer-CPC, a decentralized multi-agent reinforcement learning (MARL) method that tackles the challenge of partial observability through learned inter-agent communication. Existing representation-learning approaches like IPPO-CPC generate messages solely from current observations, failing to convey information accumulated over time. Dreamer-CPC addresses this by integrating Collective Predictive Coding (CPC) into the world model of DreamerV3. Each agent independently maintains its own world model and a dedicated message module. Messages are inferred from the latent states of the world model, which naturally encode the history of past observations and actions.

The method was evaluated in two environments: Observer (a non-cooperative information-sharing task) and CatchApple, a newly designed task where task-relevant observations are temporarily missing. Dreamer-CPC outperformed IPPO-CPC and no-communication baselines in both settings. In CatchApple, it achieved 4 to 5 times the episode return of IPPO-CPC, demonstrating robust coordination precisely where other methods fail due to missing observations. The results suggest that communication grounded in the latent dynamics of world models can enable effective decentralized decision-making when current observations are insufficient.