Tiny Aya L2-Thinker enables reasoning in user's native language
Reasoning models have historically defaulted to English regardless of input language, creating accessibility barriers and losing semantic nuance for non-English speakers. This work tackles in-language reasoning through data-centric optimization, exploring how training data composition and scheduling can enable models to reason natively in the user's language. The team built Tiny Aya L2-Thinker, a 3.35B parameter model designed to bridge prompts and responses within the same linguistic context. The approach signals a shift toward multilingual reasoning as a core capability rather than a post-hoc translation problem, with implications for model inclusivity and knowledge preservation across language families.
Modelwire context
ExplainerThe paper's core claim rests on data scheduling and composition as the primary lever for in-language reasoning, not architectural novelty. What's absent from the summary: whether this approach actually preserves reasoning quality across language families or merely shifts where errors occur.
This connects directly to the IdeaAMBIG work from the same day. That benchmark flagged how research methods are often too vague for faithful implementation. Here, the Aya team had to make concrete choices about data mixing and scheduling that likely aren't fully specified in the paper. The real test isn't whether the idea is sound, but whether another team can reproduce the exact data pipeline and get comparable results on non-English reasoning tasks. If implementation details are sparse, this becomes another case study in the specification gap.
If independent teams reproduce Tiny Aya L2-Thinker's in-language reasoning performance within 6 months using only the paper's method description, the work has genuine clarity. If reproduction attempts stall on data composition choices or scheduling specifics, it confirms that even well-intentioned multilingual research can fall into the documentation trap that IdeaAMBIG identified.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsTiny Aya L2-Thinker · Aya
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.