GCT-MARL: New framework cuts multi-agent RL training time via contrastive transfer

Training multi-agent systems from scratch for every new environment is expensive and impractical. GCT-MARL (Graph-Based Contrastive Transfer for Sample-Efficient Cooperative Multi-Agent Reinforcement Learning) addresses this by enabling agents to reuse skills across tasks, even when team sizes or agent types change. The framework extends the multi-view graph contrastive backbone of MAIL with an adaptively weighted alignment loss per view, ensuring robust feature representations transfer across varying populations. A two-phase training protocol first pre-trains on source tasks, then fine-tunes on target tasks with a small number of interactions.

Experiments show GCT-MARL significantly accelerates convergence compared to training from scratch in both homogeneous scenarios (varying numbers of same-type agents) and heterogeneous ones (cross-faction or mixed unit types). The method also naturally supports continual learning by sequentially chaining the transfer protocol across a series of related tasks, making it a unified solution for multi-task MARL. Accepted at the Continual RL Workshop at RLC 2026, the paper offers new insights for sample-efficient multi-agent deployment.