Large-scale GRPO study: native-language reasoning nearly matches English

Reinforcement Learning with Verifiable Rewards (RLVR), typically optimized using Group Relative Policy Optimization (GRPO), has become a key method for improving language model reasoning. But nearly all research has focused on English, leaving open questions about how well these techniques work for other languages. In a new large-scale study, Konstantin Dobler and colleagues from multiple institutions examined GRPO across a wide range of base models, training languages, and reasoning-language reward configurations, aiming to quantify the trade-offs of moving beyond English.

The results are encouraging for multilingual AI. Training models to reason in a native language leaves only a small performance gap compared to training for English reasoning, suggesting that high-quality reasoning can be achieved without forcing English as a lingua franca. The study also found strong crosslingual transfer: training on one language often boosts performance across many others, meaning investments in a single language can pay off broadly. However, the authors caution that these trends are highly model- and language-dependent. In certain cases, training in a specific language caused severe regressions on out-of-domain capabilities in other languages—a hidden cost that standard evaluation would miss. The paper concludes that RLVR beyond English offers broad crosslingual gains, but only if researchers and practitioners adopt comprehensive evaluation suites to catch language-specific failures.