Multilingual reasoning ability improves after careful post-training steps
What Does Post-Training Change in Multilingual Reasoning?
Computation and LanguageArtificial Intelligence
Summary
Many AI models can solve problems well only in English, limiting people who speak other languages. The authors analyzed a model called Qwen3 and found it rarely solved math problems in other languages with full reasoning shown. They studied different training updates and found that fine-tuning the model can restore reasoning in target languages but sometimes reduces accuracy or causes endless loops. Using reinforcement learning with a reward for correct, well-formed reasoning helps fix these issues and produces good results in multiple languages. This work shows a step-by-step way to make AI reasoning work better across languages.
What this means in practice
- •For multilingual ai developers: Improve language coverage in reasoning models by applying stagewise post-training with controlled supervised fine-tuning and reinforcement learning.
- •For software product teams: Deliver AI features that provide step-by-step problem-solving explanations reliably in different languages for international users.$Commercial implications: Enables commercial AI products to offer math reasoning assistance with full solutions visible in multiple user languages.
Authors
Hongyang Li, Xiao Li, Caesar Wu, Grégoire Danoy, Pascal Bouvry
Abstract
Open-source reasoning models provide unequal access to reasoning capability across languages. When a model can solve a problem but cannot deliver a complete solution in the user's language, language becomes an access barrier rather than merely a source of performance variation. We audit Qwen3 checkpoints on competition-mathematics tasks in eleven languages. Across the ten non-English languages, only 15.4-17.9% of problems receive a correct, terminating solution with visible reasoning in the requested language in any of 16 samples, compared with 92.9% in English. To identify the source of this disparity, we evaluate thirteen endpoints from one model family, spanning released checkpoints, multilingual supervised fine-tuning (SFT) at two scales, controlled SFT ablations, and three reinforcement-learning (RL) reward formulations. We jointly track correctness, language adherence, termination, and delivery efficiency. The dominant bottleneck shifts across post-training stages. Released models often reason in English. Multilingual SFT restores target-language reasoning, but accuracy declines across multilingual, English-only, and single-language SFT runs, showing that this cost is not specific to multilingual mixing; non-English reasoning traces additionally become prone to non-terminating loops. RL restores termination in both arms at no cost in accuracy, but only the arm whose reward includes a language term delivers: rewarding correctness alone returns the model to English. Together, these stages establish a constructive post-training path from English-pivoted capability to multilingual reasoning that is reliably delivered.