This paper presents Interactive World Latent (IWoL), a novel representation learning framework to facilitate team collaboration in multi-agent reinforcement learning (MARL). Building effective representations for team collaboration is challenging due to multi-agent interactions and the incomplete information resulting from local observations. IWoL directly models communication protocols to build a learnable representation space that captures inter-agent relationships and task-related world information. This representation avoids the drawbacks of explicit message passing (e.g., slow decision-making, vulnerability to malicious attacks, and sensitivity to bandwidth constraints) while maintaining fully distributed execution through implicit collaboration. In fact, the IWoL representation, as an implicit latent variable for each agent, can also be used as an explicit message for communication. We evaluate two variants of IWoL using four MARL benchmarks, demonstrating that IWoL is a simple yet powerful core for team collaboration. We also demonstrate that IWoL can be combined with existing MARL algorithms to further enhance performance.
