Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning
A 3.35B model is reported to keep its reasoning in the prompt language more than 93% of the time across 60 languages.
The paper frames “L2 reasoning” as making models reason consistently in the user’s language instead of defaulting to English. Its Tiny Aya L2-Thinker is trained through supervised fine-tuning with careful data composition and scheduling. The authors say generalization to unseen languages depends on broad language coverage, multilingual non-reasoning data, and a strong English reasoning base. They release model weights and multilingual reasoning data for further work. ArXiv · AI/CL/LG's note
The paper frames “L2 reasoning” as making models reason consistently in the user’s language instead of defaulting to English. Its Tiny Aya L2-Thinker is trained through supervised fine-tuning with careful data composition and scheduling. The authors say generalization to unseen languages depends on broad language coverage, multilingual non-reasoning data, and a strong English reasoning base. They release model weights and multilingual reasoning data for further work. ArXiv · AI/CL/LG's note
score 5