Fast-slow training pairs slow reinforcement learning weight updates with fast GEPA prompt optimization to outperform standard training on math, code, and reasoning tasks

— Approach uses less data while preserving plasticity and reducing forgetting.