Cost of Reasoning in non-English Languages: A Case Study on Japanese

2026-07-11

AI Summary

This study investigates how to achieve efficient and interpretable reasoning capabilities in non-English languages—using Japanese as a representative case—while maintaining performance comparable to English-language models. Building upon the Japanese continuously pretrained model Qwen-3-Swallow-8B, we present the first systematic application of the GRPO algorithm for reasoning alignment training in a non-English setting. Experimental results demonstrate that this approach attains performance on par with strong English baselines across general reasoning tasks, including programming, mathematics, and science, thereby validating GRPO’s feasibility for generating controllable reasoning chains in non-English contexts. However, the model exhibits weaker performance on tasks requiring Japan-specific cultural knowledge, revealing that linguistic reasoning ability does not automatically generalize to culturally grounded understanding. This work establishes a new paradigm and empirical foundation for controllable multilingual reasoning.

Abstract

Reasoning Language Models (RLMs) achieve their strongest performance when they reason in English, the language for which reasoning-oriented training data is most abundant. However, reasoning trace is a clue for model interpretability and safety, and useful in practice for both the model users and for model developers. Thus, it is desirable to be able to develop a model that reasons in a language of the user's choice, while still maintaining strong reasoning performance. To this end, we study the feasibility of training a model that reasons in Japanese. We develop a Japanese-reasoning variant of Qwen-3-Swallow-8B, which is a Japanese LLM continually pretrained from Qwen-3-8B, with GRPO and evaluate it across coding, math, and science benchmarks. The study shows that reasoning-language control is feasible by training a Japanese continually pretrained model with GRPO. However, its performance is at best on par with strong English-reasoning baselines on several benchmarks. We also evaluate the trained model on Japanese cultural benchmarks and observe that the model's performance is worse than the baseline models, suggesting that the reasoning in Japanese does not immediately improve performance on culturally relevant tasks for free.

Problem

• Reasoning Language Models
• non-English reasoning
• Japanese language
• model performance
• reasoning trace

Innovation

• reasoning-language control
• Japanese reasoning
• GRPO
• continual pretraining
• multilingual interpretability