蒸馏是模仿，学强模型的输出，把它的「答案形状」复制过来；RL 是探索，模型必须大量自己推理、自己生成、在错误里反复迭代，从试错中提炼能力。
He only learned it had been aired on TV when he saw his phone around 03:00 GMT, including messages from the US as the news reached it.
。快连下载安装是该领域的重要参考
Servers in 105 countries including the UK
// 作用：通过最值判断是否需要扩展左/右边界（左侧最小值/右侧<最大值的元素都需纳入无序区间）