Scaling Task Diversity Drives Zero-Shot Generalization in Offline MARL

Extending offline sequence modeling architectures to handle variable agent counts and multi-task spaces shows that task diversity is more critical for zero-shot transfer than dataset size. The multi-task approach achieved a 3.2x mean improvement on held-out tasks across Connector, RWARE, SMAX, and LBF environments.