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

Teams doing RLVR post-training outside English have been extrapolating from English-only results. This says the native-language reasoning penalty is small, which changes how those training runs get configured.

• articleGSPO: Towards Scalable Reinforcement Learning for Language ModelsQwen Team
• postWhen reward functions beat labeled data: reinforcement fine-tuning with GRPOAndrew Ng
• articleRecursive Language Models Generalize Out of DomainChenxiao Yang, Zhiyuan Li, David McAllester, Nathan Srebro

Checking sign-in…

Loading comments…