GRPO Beyond English: large-scale study finds strong crosslingual transfer but hidden regressions
May_F1_ · x · 2026-08-17
A new arXiv paper, GRPO Beyond English: A Large-Scale Study of GRPO in Non-English and Multilingual Settings, empirically studies RLVR/GRPO across many base models, training languages, and reasoning-language rewards.
Key findings:
- Training models to reason in their native language often leaves only a small gap vs. English-reasoning training;
- Strong crosslingual transfer: training in one language frequently improves many others;
- Effects are highly model- and language-dependent, and training in a particular language can sometimes cause severe regressions on out-of-domain capabilities in other languages.
The authors conclude RLVR beyond English can deliver broad crosslingual gains, but requires broad evaluation to detect language-specific regressions.
More from Research
- Microsoft Research: discarding singleton proteins in PLM training is likely wrong — KevinKaichuang · 2026-08-18
- Scott Aaronson on AI Math Breakthroughs: 'Enough with Milestones' — dilipkay · 2026-08-17
- Microsoft Research: 4B model tuned with SocialRL out-negotiates GPT-5 family — dair_ai · 2026-08-17
- Math costs collapse, enabling theorems to replace experiments in transformer research — DimitrisPapail · 2026-08-17
- AI + Data Science: Moving towards real-time predictive insights — mdancho84 · 2026-08-17
- Recursive Self-Improvement Harder Than Estimated, Novel Ideas Rare — sarahcat21 · 2026-08-17