A large-scale empirical study examines Group Relative Policy Optimization (GRPO) applied to reinforcement learning with verifiable rewards (RLVR) beyond English, testing multiple base models, training languages, and reasoning language rewards. Training models to reason in their native language yields performance close to English-based reasoning training, and training in one language frequently transfers benefits to others. However, results vary substantially by model and language, and some training languages cause significant regressions in out-of-domain capabilities for other languages, underscoring the need for broad multilingual evaluation.
Questions this post answers
Does training a language model to reason in its native language hurt performance compared to training it to reason in English?
No, it only leaves a small gap. A large-scale study of Group Relative Policy Optimization (GRPO) across many base models and languages found that native-language reasoning training performs close to English-based reasoning training, and training in one language often transfers gains to many other languages, though results are highly model- and language-dependent. Researchers comparing multilingual training strategies can follow developments in reasoning model training on daily.dev.
Can training a language model with RLVR in one language cause problems in other languages?
Yes, training with reinforcement learning with verifiable rewards (RLVR) in a specific language can induce severe regressions in out-of-domain capabilities for other languages, even while producing strong crosslingual transfer overall. This means broad evaluation across languages is needed to detect language-specific regressions that a narrow English-only test suite would miss. Teams evaluating multilingual model training risks can track this kind of research on daily.dev.