AI 导读
1/ 通过 TEMPO 实现递归自我批判 在长周期任务中，最终奖励可能在数千个早期决策之后数小时才到来。 TEMPO 让同一个模型在 actor 和 critic 之间交替。在每个宏步骤中，它会暂停、检查当前状态、使用工具，并评估轨迹是否真的在朝成功推进。 这提供了更早、更丰富的学习信号——帮助区分真正的进展和看起来很有说服力的死胡同。
正文
1/ Recursive Self-Critique via TEMPO
In long-horizon tasks, the final reward may arrive hours after thousands of earlier decisions.
TEMPO lets the same model alternate between actor and critic. At each macro-step, it pauses, examines the current state, uses tools, and estimates whether the trajectory is actually moving toward success.
This provides earlier, richer learning signals—helping distinguish genuine progress from a convincing-looking dead end.
来源：@rohanpaul_ai · x.com