New LLM variant enables Japanese reasoning but shows no performance gains

Researchers have investigated the feasibility of training large language models to reason in Japanese, aiming to improve interpretability and user experience. They developed a Japanese-reasoning variant of the Qwen-3-Swallow-8B model, which was continually pre-trained from Qwen-3-8B using GRPO. While this approach allows for reasoning-language control, the model's performance on coding, math, and science benchmarks was only on par with strong English-reasoning baselines. Furthermore, the Japanese-reasoning model did not show improved performance on Japanese cultural benchmarks, indicating that reasoning in a non-English language does not automatically translate to better performance on culturally specific tasks. AI

IMPACT Investigating non-English reasoning in LLMs could lead to more accessible and interpretable AI tools for a global user base.

RANK_REASON The cluster contains an academic paper detailing a new approach to training LLMs for non-English reasoning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• English
• GRPO
• Japanese
• Qwen 3:8b
• Qwen-3-Swallow-8B

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →