Back to Research papers
Research paper index

On-Policy Delta Distillation for Multilingual Math Reasoning

Byeongho Heo, Jaehui Hwang, Sangdoo Yun, Dongyoon Han

arXiv:2608.05802Published August 6, 20260 citations
  • cs.CL
  • cs.LG
  • policy
  • reinforcement learning

Abstract

On-Policy Distillation (OPD) is emerging as a promising alternative to reinforcement learning for LLM post-training, yet its effectiveness in multilingual settings remains underexplored. We study OPD and its advanced variant, On-Policy Delta Distillation (OPD$^2$), for mathematical reasoning in English, Korean, and Japanese. OPD$^2$ improves OPD by using the probability gap between a post-trained teacher and its base model as the learning signal. Experiments with Qwen3 show that OPD$^2$ consistently outperforms the original OPD, with particularly strong improvements in Korean and Japanese, and generally narrows the English-Korean performance gap. We further find that English-only OPD can also increase performance for Korean and Japanese, but often shifts the responses toward English, highlighting the importance of multilingual data to preserving target-language responses.

Read the original paper

This page indexes public paper metadata. The manuscript remains with its original publisher and authors.