Why learning many tasks helps then hurts continual learning?

Task diversity has a Jekyll-and-Hyde effect on reinforcement learning agents. Varying map layouts, objects, and goal hierarchies helps agents start strong on new tasks—but only temporarily. The researchers built Banyan, a new benchmark with independently controllable task dimensions, and showed that while local transfer improves with diversity, sustained learning across many sequential shifts actually deteriorates: agents forget earlier tasks and performance on longer-horizon tasks stalls. The finding reveals a hard trade-off between generalization and continual adaptation.