GRPO Beyond English: A Large-Scale Study of GRPO in Non-English and Multilingual Settings

Training a model to reason in a non-English language can improve its performance in other languages, but the gains are uneven. This study of multilingual reinforcement learning examines training with checkable rewards across different base models, training languages and incentives for the language used during reasoning.

The authors report that native-language reasoning often comes close to English reasoning, and that training in one language frequently helps others. Some combinations, however, severely reduce performance on other kinds of tasks in other languages. A gain in the training language therefore gives an incomplete picture of what multilingual users will experience.