Original title: Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning
Mehrnaz Mofakhami, Ananya Sahu, Alejandro R. Salamanca, Daniel D'souza +4 · September 2026
Most smart reasoning AI models think in English, even when asked a question in another language. This causes problems. Non-English speakers struggle to use them, the model can lose the user's true intent, and it misses facts better expressed in the local tongue. Researchers wanted models to do L2 reasoningThe ability of an AI model to perform reasoning tasks in a language that is not its primary trained language, matching the language of the user's prompt., which means thinking step-by-step in the same language as the prompt. They solved this by testing how to mix and schedule different training data during fine-tuning. They built a model called Tiny Aya L2-Thinker with 3.35B parameters. It succeeded, reasoning in the prompt's language over 93% of the time across 60 languages. It worked on six diverse benchmarks while staying accurate. The team found that reasoning is not tied to one language. By using good data mixtures, models can transfer reasoning skills to new languages without needing reasoning training in every single one.
Imagine you're learning to cook. You start with a cookbook that only has recipes in English. Even if you can read the instructions, you might miss out on cooking techniques or ingredients that are better known in other languages. Now, imagine you get a special cookbook that includes English recipes but also teaches you how to adapt them using ingredients and methods from many other cultures. In this paper, the AI is like that special cookbook. It learns to 'cook' (reason) using prompts in any language by carefully mixing English reasoning skills with a wide variety of other language data.
Many people don't speak English, and asking AI questions in their own language makes it more accessible. This also means AI can better understand nuances and knowledge that are specific to different cultures and languages, rather than forcing everything through an English lens.
The abstract doesn't detail specific performance differences across the 60 languages or mention how the model performs on languages not included in its training data, beyond stating that broader language coverage helps generalization.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning — Mehrnaz Mofakhami, Ananya Sahu, Alejandro R. Salamanca et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.