Reasoning language models have advanced rapidly on complex tasks, but their reasoning remains overwhelmingly English-centric: models tend to reason in English no matter which language the user writes in. According to the paper, this limits accessibility for non-English speakers, risks distorting the intent of the original question, and discards knowledge that is more readily expressed in the target language.
The work focuses on L2 reasoning — the ability of a model to reason consistently in the language of the user's prompt, forming an in-language bridge between question and answer. Rather than tackling this through new architectures, the authors take a data-centric angle, studying how data composition and scheduling during supervised fine-tuning (SFT) affect reasoning generalization.
The result is Tiny Aya L2-Thinker, a 3.35B-scale model that reaches an L2 reasoning rate above 93% across 60 languages on six benchmarks covering math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning, while maintaining strong overall performance.
The paper reports that generalizing L2 reasoning to held-out languages depends on three factors: broader language coverage, readily available multilingual non-reasoning data, and a sufficient English reasoning backbone. The authors conclude that reasoning is a language-agnostic behavior that can transfer across typologically diverse languages through careful data mixing — without reasoning supervision in every target language.
Cohere Labs has released the model weights and multilingual reasoning data openly on Hugging Face, including Tiny Aya L2-Thinker, an English-reasoning counterpart (Tiny Aya En-Thinker), and a multilingual reasoning dataset covering 44 languages, to support further research on accessible, in-language reasoning.