GRPO Beyond English: A Large-Scale Study of GRPO in Non-English and Multilingual Settings
Uncovering the potential and hidden performance pitfalls of multilingual reinforcement learning for LLMs.
Reinforcement Learning with Verifiable Rewards (RLVR), often optimized with Group Relative Policy Optimization (GRPO), has become a central recipe for improving the reasoning capabilities of pretrained language models but current studies remain heavily English-centric. We conduct...
ArXiv CS.CL
Source
Generated insight by OnAirToday AI
This research reveals that while reinforcement learning with verifiable rewards provides strong cross-lingual reasoning gains, it introduces significant risks of performance regression in non-target languages. Establishing these multilingual scaling laws is essential for developers aiming to build truly global, robust reasoning models.
Comments
Keep comments relevant and respectful. No links, advertising, self-promotion, or spam.