Chinese Academy of Sciences
“Temporal GRPO addresses this problem by constructing detectable task stages, aligning each rollout with stage-specific action intervals, and comparing only rollouts that have entered the same stage.”
Source→“A rollout that completes several valid stages but fails later can therefore penalize the actions that produced its earlier progress. We call this trajectory-level credit aliasing.”
Source→“Corresponding author and likely senior researcher overseeing the academic direction of the project”
Source→“The breadth of co-authors across Tsinghua, Peking University, Chinese Academy of Sciences, HKUST(GZ), and Xiaomi suggests this is a coordinated multi-institution effort, not an isolated academic project.”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.