Question Regarding On-Policy Distillation in Qwen3

[caiyuchen-ustc]

Dear Qwen Team,

In the Qwen3 technical report, Table 21 mentions the use of on-policy distillation. During my reproduction experiments, I found that when no sentence-level importance sampling or clipping is applied, the student model’s errors are progressively amplified, eventually leading to training collapse.

Could you clarify whether your on-policy distillation implementation includes any stabilization mechanisms, such as sentence-level importance weighting, ratio clipping, or other constraints? If so, I would appreciate it if you could share which techniques were used in practice.

Thank you for your help.

Best regards,
Cai Yuchen