arxivcs.AIcs.CL2026-07-07
TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training
Yuhang Zhou, Kai Zheng, Haoling Li, Dengyun Peng, Can Xu, Jingjing Chen
On-policy distillation (OPD) trains a student policy by matching a stronger teacher on the student's own trajectories, offering a promising framework for language agent training. However, its application to long-horizon agentic tasks remains insufficiently explored. We identify t…