PrAg-PO. PrAg-PO improves RL-based mathematical reasoning by training models across diverse prompt templates with template-specific format rewards, yielding more diverse reasoning trajectories, stronger benchmark accuracy, and more stable training.