Cohere has shown a training method that makes an AI reason in the same language as the question, and its results.
Cohere researchers trained a model to reason in the prompt language rather than switching to English, across 60 languages. The model is Tiny Aya L2-Thinker (3.35B parameters). Training mixed 1.7 million English step-by-step solutions, about 5,000 translated solutions per language across 44 languages, and question-answer pairs to generalize to unseen languages. The model reasons in the prompt language over 93% of the time across 60 languages, and accuracy dropped by at most 2 to 3 points on five of six benchmarks.
In-language reasoning held even in low-resource languages. A comparable larger model fell from 72% to 5% in one case, while this model stayed at 94 to 99%. However, accuracy fell more on competition-level mathematics, and the method depends on the quality of the translated training data.
Whether further training closes the gap on competition math, and how well in-language reasoning holds when translation quality differs by language. These two points set how far the method can be applied.