Cohere's Tiny Aya L2-Thinker reasons in any language

Cohere Labs published Building Multilingual Bridges, research on a gap most reasoning models have: they answer in your language but think in English. The post argues this loses nuance through internal translation, leaves language specific cultural knowledge unused, and shuts non English speakers out of reading and engaging with the reasoning trace itself. Reasoning in the user’s language rather than only answering in it is what the authors call L2 reasoning.

The method is data mixing rather than a new architecture, and the post reduces it to three pillars. Broader language coverage in reasoning training data improves generalization to languages that have no reasoning supervision at all. A small share of multilingual non reasoning data boosts both accuracy and reasoning in the prompt language. English reasoning data provides the task solving backbone the model still needs.

The result is Tiny Aya L2-Thinker, a 3.35B model that reasons in the prompt language over 93% of the time across 60 languages. Cohere reports minimal accuracy loss to get there, that it holds up on low resource languages, and that it uses fewer reasoning tokens than comparable models.

The most transferable finding is the last one: a model’s ability to reason and the language it reasons in are separate properties, so reasoning in the user’s own language does not have to be built language by language. That is what puts it within reach where multilingual reasoning data is scarce. Cohere is releasing the model weights and the multilingual reasoning data behind the work.

(Original source: Cohere)