arxivcs.LG2026-07-31
Adaptive FastOPD: Progress-Aware Rollout Horizon Expansion for Efficient On-Policy Distillation
Qian Tan, Huaifei Liang, Xuanyu Zhu, Lei Jiang, Yuqiang Li
On-policy distillation (OPD) provides dense teacher supervision along student-generated trajectories, but its online rollout process incurs substantial computational cost, particularly when a few long responses delay batch completion. Existing acceleration methods typically contr…