Reinforce-Distill-Reinforce

A three-stage progressive model training strategy that successively applies reinforcement learning exploration, distillation compression and a further reinforcement stage to restore performance, improving inference efficiency while maintaining generation quality.