训练加速1.8倍，推理开销降78%，精准筛选题目高效加速RL训练