From Local to Global: Progressive Consensus via Hierarchical Communication in Multi-Agent Reinforcement Learning

Effective team collaboration hinges on consensus formation through information exchange, a principle equally critical in multi-agent reinforcement learning (MARL). However, existing communication-based and consensus-learning methods often struggle to form a coherent global understanding from agents' limited local views. We propose STAGE, a local-to-global progressive consensus framework based on hierarchical communication. By first grouping agents according to their perceptual focuses, STAGE enables intra-group communication to form local consensuses, then selects group leaders to exchange these consensuses across groups, progressively expanding them into a unified global consensus with reduced communication redundancy. To further improve consensus quality, we introduce KL-divergence constraints for consensus alignment and a variational autoencoder (VAE) objective for preserving task-relevant global information. Extensive experiments on challenging MARL benchmarks show that STAGE consistently outperforms state-of-the-art baselines, especially on more difficult tasks and larger-scale multi-agent systems.