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

Jul 16, 2026

Upload by Liam

Reasoning Language Models perform best when producing reasoning in English, where training data is most abundant, yet multilingual reasoning traces are valuable for interpretability, safety, and user accessibility. The study investigates training a model to reason in Japanese without losing reasoning strength. A Japanese-reasoning variant of Qwen-3-Swallow-8B is built via continual pretraining and GRPO, then evaluated on coding, math, and science benchmarks, plus Japanese cultural benchmarks. Results show feasible Japanese reasoning control, but performance does not consistently exceed strong English baselines on culturally relevant tasks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

Abstract

Introduction

Motivation: English dominance in reasoning

Goal: reasoning traces in a user’s language

Case study: Japanese-reasoning model design

Cost of Reasoning in non-English Languages: A Case Study on Japanese Yuu Jinnai CyberAgent / Tokyo, Japan [jinnai_yu@cyberagent.co.jp](jinnai_yu@cyberagent.co.jp) Models: [https://huggingface.co/collections/cyberagent/cat-thinking](https://huggingface.co/collections/cyberagent/cat-thinking) arXiv :2607 . 10 1 14v 1 [ cs .CL] 11 Jul 2026 Abstract Reasoning Language Models (RLMs) achieve their strongest performance when they reason in English, the language for which reasoningoriented 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 Englishreasoning 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. 1 Introduction English is the most resource-rich language by far, especially for reasoning-oriented training data. Thus, most of the Reasoning Language Models (RLMs) are trained to reason in English (Yang et al., 2025 ; Guo et al., 2025 ; Team et al., 2025 ; Agarwal et al., 2025 ; Muennighoff et al., 2025) . Even multilingual LLMs tend to reason best in English: across many benchmarks, including multilingual ones, reasoning in English outperforms reasoning in the question’s own language (Saji et al., 2026 ; Yong et al., 2025) . Yet, reasoning trace is a useful signal for the model users and for model developers that would ideally be available in the user’s or developers language of choice. For example, reasoning trace is useful for model interpretability (Wei Jie et al., 2024) and safety (Jiang et al., 2025 ; Guan et al., 2025), and can be used to improve model performance (Wei et al., 2023 ; Seo et al., 2026) . Such reasoning trace is also useful for model developers to understand the model’s reasoning process and to debug the model’s behavior. Users may also prefer to see reasoning trace in their own language for better accessibility. The question is thus how to achieve a reasoning trace in a language of one’s choice with minimum degradation of the model’s performance. To answer this question anecdotally, we develop a Japanese-reasoning language model as a case study. We build upon Qwen-3-Swallow-8B (Fujii et al., 2024 ; Ma et al., 2025), a Japanese LLM continually pretrained from Qwen-3-8B (Yang et al., 2025), and train it with supervised warm-start and two-stage GRPO to produce reasoning traces in Japanese. The resulting model achieves 100% Japanese reasoning traces on all evaluated benchmarks, including those whose instructions are entirely in English. Its performance remains competitive with the English reasoning models on most benchmarks, but falls short on several tasks, suggesting that reasoning-language control is feasible but not free of cost without further improvements to the methodology. We observe two ingredients that were necessary for the development of our Japanese-reasoning model. First, we find that continual pretraining on Japanese data significantly helps the model to reason in Japanese. Qwen-3-Swallow-8B learns to reason within hundreds of GRPO steps, while Qwen-3-8B (with SFT warmstart) shows no sign of generating

FAQs

Why do reasoning language models tend to perform best in English?

Because reasoning-oriented training data is far more abundant in English, and existing RLMs frequently show higher scores when reasoning in English even on multilingual settings.

How is the Japanese-reasoning model trained in this case study?

It builds a Japanese-reasoning variant of Qwen-3-Swallow-8B using continual pretraining on Japanese data and GRPO, with supervised warm-start and a two-stage GRPO setup to generate Japanese reasoning traces.

What do the benchmark results imply about the “cost” of reasoning in Japanese?

Japanese reasoning-language control is feasible, but performance is often only on par with strong English reasoning baselines and can be worse on Japanese cultural benchmarks, indicating that improving reasoning language does not automatically improve culturally relevant task performance.

Jul 16, 2026

Upload by Liam