Tiny Aya L2-Thinker tops 93% in-language reasoning across 60 languages
Reasoning language models have gotten much better at complex tasks, but according to the researchers behind this paper, they remain overwhelmingly English-centric: even when a user prompts a model in another language, it tends to do its actual reasoning in English before answering. The researchers say this is inaccessible for non-English-speaking users, risks losing the intent of the original question, and forgoes knowledge that is more readily expressed in the language the user actually asked in.
To address this, the team worked on what they call L2 reasoning: getting a model to reason consistently in the same language as the user's prompt, building what they describe as an in-language bridge between the question and the answer. Rather than adding reasoning-specific training data for every language, they took a data-centric approach, studying how the composition and scheduling of data used in supervised fine-tuning (SFT) affects whether this in-language reasoning ability generalizes.
Applying that approach, they built a 3.35-billion-parameter model called Tiny Aya L2-Thinker. Across 60 languages and 6 benchmarks covering math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning, the model reasoned in the prompt's own language more than 93% of the time, while the researchers say task performance stayed strong rather than degrading as a side effect.
The researchers also looked at how this ability carries over to languages the model was not specifically trained to reason in. They report that generalizing to these held-out languages depends on three things: broad language coverage during training, multilingual data that does not itself contain reasoning traces but is easy to obtain, and a sufficiently strong English reasoning backbone to build on. From this they conclude that reasoning is a language-agnostic behavior: it can transfer across languages that are typologically very different from one another through careful data mixing, without needing reasoning supervision written specifically for every target language.
The team released both the Tiny Aya L2-Thinker model weights and the multilingual reasoning data used to train it, to support further research on making in-language reasoning more accessible.
Key facts
- Tiny Aya L2-Thinker, a 3.35B-parameter model, reasons in the language of the user's prompt (rather than defaulting to English) more than 93% of the time across 60 languages.
- That above-93% in-language reasoning rate was measured on 6 benchmarks covering math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning.
- The researchers used a data-centric method: tuning the mix and schedule of supervised fine-tuning (SFT) data, rather than adding reasoning-specific training for every language.
- Generalizing in-language reasoning to languages held out of training required broad language coverage, easily available multilingual non-reasoning data, and a strong English reasoning backbone.
- The model weights and the multilingual reasoning dataset behind Tiny Aya L2-Thinker have both been released.
Why it matters
Reasoning models mostly work out problems in English even when the user asks in another language, which the researchers say makes them inaccessible to non-English speakers, risks losing the intent of the original question, and skips over knowledge that is easier to express in that language. This work shows a model can be taught to reason in the prompt's own language across 60 languages, at a small 3.35B-parameter scale, mainly by changing how training data is mixed rather than by adding separate reasoning training for each language. Because the model weights and the training data are both released, the recipe is something other teams can inspect and build on directly.
Who it affects
Non-English-speaking users of reasoning models are the direct beneficiaries, since their prompts would be reasoned through in their own language rather than silently switched to English. The results also target researchers and developers building multilingual assistants: the paper's benchmarks cover math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning, and both the model and its training data are released for others to reuse or study.
How to use it
The researchers released the Tiny Aya L2-Thinker model weights and the multilingual reasoning data used to train it, so other teams can run or build on the model directly. The text states only that the weights and data are released, without specifying a license, access terms or where they are hosted.
How solid is it
The above-93% in-language reasoning rate was measured across 60 languages on 6 benchmarks covering math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning, with task performance reported as staying strong rather than dropping as a side effect of switching languages. The paper does not name the 6 benchmarks individually or give a baseline or prior L2 reasoning rate, so there is no stated figure for how much this improves on earlier models.
Risks and caveats
The text names no authors, institution or company behind the work and gives no publication or submission date, so the paper's provenance cannot be checked from the abstract alone. The source says the model kept performance strong alongside the above-93% in-language reasoning rate, but it gives no separate number for that performance, so how much accuracy trades off against reasoning in the right language cannot be quantified from the text.