Study: Cross-lingual math reasoning transfer in LLMs is variable

A new study investigates the cross-lingual mathematical reasoning capabilities of large language models (LLMs) that are primarily trained on English data. The research found that the transfer of mathematical reasoning skills to other languages is highly variable, depending on the specific LLM, the target language, and the training methodology. Interestingly, models with stronger English mathematical reasoning did not always show more efficient transfer to other languages. The study also highlighted that incorporating even a single parallel language in training significantly boosts performance, with additional languages providing smaller but consistent gains. AI

IMPACT Highlights the need for direct multilingual evaluation of LLMs to understand their true cross-lingual reasoning abilities.

RANK_REASON Research paper published on arXiv detailing findings about LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• English
• LLMs
• Reinforcement Post-Training
• Wen Yang

• paper
• model release

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