Apple researchers have published a large-scale study of how reinforcement learning affects reasoning in non-English and multilingual language models. The work focuses on Group Relative Policy Optimization, or GRPO, a training method often used to improve reasoning with verifiable rewards.
The central finding is practical: training models to reason in a native language often leaves only a small gap compared with training them to reason in English. That matters because much of today’s reasoning research still treats English as the default test bed.
The study spans a range of base models, training languages, and reward settings. It examines whether the language used during training changes the model’s ability to solve reasoning tasks across languages.
The result does not mean multilingual reasoning is solved. It points to a narrower question: how much performance is lost when models are optimized outside English, and when English-centric assumptions are unnecessary. For developers and evaluators, the paper is a reminder that language choice is part of the system design, not just a localization detail.