Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning
Researchers train Tiny Aya L2-Thinker, a 3.35B model achieving over 93% in-language reasoning across 60 languages via multilingual data mixing.
The paper addresses L2 reasoning, where models reason consistently in the language of the user's prompt rather than defaulting to English. Through data-centric SFT optimization, the 3.35B Tiny Aya L2-Thinker reaches above 93% L2 reasoning rate across 60 languages on 6 benchmarks covering math, commonsense, instruction following, open-ended generation, and cultural reasoning. The authors find that generalization to held-out languages relies on broad language coverage, multilingual non-reasoning data, and a strong English reasoning backbone, without needing reasoning supervision in every target language. Model weights and multilingual reasoning data are released.
- 3.35B Tiny Aya L2-Thinker exceeds 93% in-language reasoning across 60 languages
- Performance spans 6 benchmarks including math and cultural reasoning
- Generalization to held-out languages achieved without per-language reasoning supervision
- Released reasoning data includes multilingual non-reasoning datasets
- Findings indicate reasoning is a transferable, language-agnostic behavior
Full article220 words · extracted from huggingface.co · click to collapse
Reasoning language models have made substantial advances on a variety of complex tasks, yet their capabilities remain overwhelmingly English-centric: models primarily reason in English regardless of the language they are prompted in. This is inaccessible for non-English-speaking users, risks losing the intent of the original question, and forgoes knowledge more readily expressed in the target language. In this work, we advance L2 reasoning, the ability of a model to reason consistently in the language of the user's prompt, thus building an in-language bridge between the prompt and the answer. We approach this problem from a data-centric angle, investigating how to optimize data composition and scheduling in SFT for reasoning generalization. Building Tiny Aya L2-Thinker at 3.35B scale, we achieve an L2 reasoning rate above 93% across 60 languages on 6 benchmarks spanning math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning while keeping performance strong. We show the path to generalizing L2 reasoning to held-out languages goes through broader language coverage, readily available multilingual non-reasoning data, and a sufficient English reasoning backbone. These findings indicate that reasoning is a language-agnostic behavior that can be transferred across typologically diverse languages through careful data mixing and without requiring reasoning supervision in every target language. We release our model weights and multilingual reasoning data to support further research on accessible, in-language reasoning.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.10445