Changwen Zheng

“Temporal GRPO keeps changes on the preceding stages close to zero and produces the largest positive improvement at md, indicating that it preserves acquired preceding behaviors and concentrates the update on the stage responsible for the rollout difference.”