基于这些关键发现，研究团队进一步提出了一个 training-free 的并行推理控制算法 Parallel-Probe，能够在不牺牲核心准确率的前提下，显著减少无效计算，将推理延迟降低 35.8%，总 token 成本降低 25.8%。

Isaacman also has been driven by something else that took place the day he was sworn in: the release of an executive order on space policy. That directive calls on NASA to return humans to the moon by ...