Ask to Distill: Learning to Synthesize Data for On-Policy Distillation
Published in arXiv preprint, 2026
Recommended citation: Jiashuo Sun, Hao Zhang, Xueqiang Xu, Heng Wang, Yixuan Xie, Pengcheng Jiang, Zhenrui Yue, Jiajun Fan, Peiran Li, Ge Liu, Shaowen Wang, Jiawei Han. "Ask to Distill: Learning to Synthesize Data for On-Policy Distillation." arXiv preprint arXiv:2610.27450, 2026. https://arxiv.org/abs/2610.27450
Ask-to-Distill learns to write training questions for on-policy distillation without reference answers. A synthesizer initialized from the student is optimized with a distributional surrogate for question value, computed from a frozen student–teacher pair, and the filtered questions form a prompt-only bank for standard on-policy distillation. Across math, code and agentic tasks it improves on unoptimized student-generated prompts and KL-filtered real questions, and matches or exceeds distillation on a real question bank in the evaluated settings.