Can Chinese‑Only Inference Training Match English? Apple’s New Study Shows Only 1.1% Gap

Apple, together with the Hasso Plattner Institute, released a paper that performed more than 200 GRPO reinforcement‑learning‑for‑inference (RLVR) experiments on nine base models and eleven languages. The key design choice examined was whether the reward model used English reasoning or the questioner’s native language.

When the reward language was switched to Chinese, the performance gap compared with English was only 1.1 percentage points , dramatically smaller than the >10 percentage‑point gaps reported in earlier studies. For other high‑resource languages the gaps were similarly small: French 1.4, German 1.7, Japanese 1.6, all generally within 2 percentage points. In contrast, low‑resource languages showed larger degradations: Bengali 5.7, Telugu 4.6, Swahili 4.5.

The authors also built an 11 × 14 cross‑language transfer matrix by training each language monolingually and evaluating on all fourteen languages. The matrix revealed that training on a single language often improves many others. For example, Spanish‑only training restored 96 % of the gain that direct French training would provide, and Chinese‑Japanese transfer was especially strong, likely due to shared CJK pre‑training resources.

Surprisingly, English was not always the best source language. In Qwen3‑4B‑Base and Qwen3‑8B, Bengali training yielded the strongest transfer to Chinese. The authors hypothesize that low‑resource language training forces deeper re‑organisation of reasoning abilities, sometimes leading to better generalisation.

Mixed‑language training also performed well: Qwen3‑Base models trained on multilingual data were on average only 0.3 percentage points behind the ideal per‑language optimal source, while requiring a single model.

However, the study identified severe performance collapse cases. Certain model‑language pairs caused drastic drops on unrelated tasks. For Qwen3 non‑Base models trained with Swahili or Telugu, performance on unseen English tasks fell by up to 19.2 percentage points . Similar collapses appeared on difficult math tasks: non‑English monolingual training on Qwen3 non‑Base and SmolLM3‑3B reduced PolyMath high‑difficulty scores by more than 30 percentage points.

Reward‑language choice could also trigger collapse. The gemma‑3‑1b‑it model trained with Swahili reward data lost 3.1 percentage points on simple math (MGSM, PolyMath‑low), whereas the same data with English rewards improved performance by 17.5 percentage points. These failures were specific to particular model‑language configurations and could not be predicted from the model or language alone.

The authors stress that thorough per‑language, per‑task, and per‑model evaluation is essential before deployment, as average scores can hide hidden regressions. The experiments were limited to open‑source models under 8 B parameters, programmatically generated math and logic tasks, a single random seed, and no extensive hyper‑parameter tuning, so real‑world performance may differ.

Overall, the results suggest that for high‑resource languages like Chinese, inference‑reward training in the native language is viable with minimal accuracy loss, opening a path for non‑English models to bypass English‑centric pipelines. Yet, careful risk assessment remains necessary, especially for low‑resource languages and specific model families.