Building Multilingual Bridges (cohere.com)

Recent advancements in AI language models have highlighted a significant limitation: their reasoning capabilities are primarily focused on English, which can alienate non-English speakers and obscure the intent behind prompts in other languages. Addressing this issue, researchers have introduced L2 reasoning, a method that enables models to provide answers in the same language as the user's prompt. This is exemplified by the development of the Tiny Aya L2-Thinker model, which operates at a scale of 3.35 billion parameters and demonstrates an impressive L2 reasoning rate of over 93% across 60 languages, covering diverse tasks such as math, commonsense reasoning, and cultural context. The significance of this advancement lies in the potential for more inclusive and accessible AI interactions, allowing speakers of various languages to benefit from sophisticated reasoning without losing contextual meaning. The research emphasizes the importance of optimizing data composition and scheduling during supervised fine-tuning (SFT) to enhance reasoning generalization across languages. By showcasing that reasoning can be language-agnostic and transferable through strategic data mixing, this work sets a new standard for multilingual AI applications. The release of the model weights and multilingual reasoning data supports ongoing efforts to improve in-language reasoning capabilities, making AI technology more universal.