OPD2 improves multilingual math reasoning in Qwen3 tests
The distillation method beat OPD in English, Korean and Japanese math reasoning experiments.
Why it matters
The results add evidence that distillation-based post-training can improve multilingual reasoning without relying solely on reinforcement learning. They also show that language coverage in post-training data matters for preserving target-language behavior.
The key points
- 1.OPD2 uses the teacher-base probability gap as its learning signal.
- 2.Qwen3 tests showed stronger gains in Korean and Japanese.
- 3.English-only OPD improved scores but could push outputs toward English.
A new arXiv paper studies On-Policy Distillation and On-Policy Delta Distillation, or OPD2, for multilingual mathematical reasoning in English, Korean and Japanese. In experiments with Qwen3, OPD2 consistently outperformed the original OPD method, with especially strong gains in Korean and Japanese, and generally narrowed the English-Korean performance gap. The authors also found that English-only OPD can improve Korean and Japanese performance but may shift responses toward English.
⚡ Try this today
Use multilingual post-training data when applying OPD-style methods to non-English reasoning tasks.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
- arXiv cs.LGOn-Policy Delta Distillation for Multilingual Math ReasoningAug 7, 12:00 PM↗
- arXiv cs.CLOn-Policy Delta Distillation for Multilingual Math ReasoningAug 7, 12:00 PM↗
- HF Daily PapersOn-Policy Delta Distillation for Multilingual Math ReasoningAug 6, 4:00 AM↗
- arXiv cs.AIWhen Teachers Mislead: Spurious-Signal-Aware On-Policy DistillationAug 5, 12:00 PM↗
- arXiv cs.CLLanguage-Specialized Multi-Teacher On-Policy Distillation for Multilingual LLM-Based ASRAug 5, 12:00 PM↗
Enjoyed this brief? Get the next one in your inbox.
More in Research
Google advances private AI with homomorphic encryption
Google says it is making private AI more practical using homomorphic encryption.
Google says it is making private AI practical
A Google Security post drew Hacker News discussion alongside broader AI critiques.
Researchers release LittleLearner and LittleCurriculum
The 5B-parameter model is trained on an 88B-token corpus capped at U.S. Grade 5 material.