OPD² improves multilingual math reasoning in Qwen3 tests
The arXiv study reports gains over OPD in English, Korean and Japanese math reasoning.
Why it matters
The results point to a possible alternative to reinforcement learning for LLM post-training in multilingual reasoning tasks. They also show that multilingual training data may be important for preserving target-language outputs, not just improving scores.
The key points
- 1.OPD² beat OPD in Qwen3 multilingual math tests.
- 2.Korean and Japanese saw particularly strong improvements.
- 3.English-only OPD risked shifting outputs toward English.
Researchers studied On-Policy Distillation and On-Policy Delta Distillation, or OPD², for multilingual math reasoning in English, Korean and Japanese. In experiments with Qwen3, OPD² consistently outperformed original OPD, with especially strong improvements in Korean and Japanese, and generally narrowed the English-Korean performance gap. The study also found English-only OPD can improve Korean and Japanese performance but may shift responses toward English.
⚡ Try this today
Use multilingual post-training data when testing OPD-style methods for non-English reasoning tasks.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
Enjoyed this brief? Get the next one in your inbox.
More in Research
InternLM details Intern-S2-Mobius architecture
The model separates memory and reasoning, reporting comparable scores with less data and faster inference.
Google applies homomorphic encryption to private AI
Google says encrypted processing can help make private AI more practical.
Google advances private AI with homomorphic encryption
Google says it is making private AI more practical using homomorphic encryption.