On-Policy Delta Distillation for Multilingual Math Reasoning
A delta signal from the teacher’s post-training appears to help multilingual math models more than plain on-policy distillation.
The paper tests OPD and OPD² on English, Korean, and Japanese math reasoning with Qwen3. OPD² uses the probability gap between a post-trained teacher and its base model as the training signal, and it consistently beats the original OPD in the reported experiments. The gains are strongest for Korean and Japanese, and the method generally reduces the English-Korean performance gap. English-only OPD can help Korean and Japanese scores, but it often pushes answers toward English, making multilingual training data important for keeping responses in the target language. HF Daily Papers' note
The paper tests OPD and OPD² on English, Korean, and Japanese math reasoning with Qwen3. OPD² uses the probability gap between a post-trained teacher and its base model as the training signal, and it consistently beats the original OPD in the reported experiments. The gains are strongest for Korean and Japanese, and the method generally reduces the English-Korean performance gap. English-only OPD can help Korean and Japanese scores, but it often pushes answers toward English, making multilingual training data important for keeping responses in the target language. HF Daily Papers' note
score 4