Cohere Labs: data mixing lifts 3.35B model to 93%+ in-language reasoning across 60 languages
Cohere_Labs · x · 2026-09-11
Cohere Labs released "Building Multilingual Bridges," a data-centric study on enabling LLMs to reason in the user's language (L2 reasoning) instead of defaulting to English. Their 3.35B Tiny Aya L2-Thinker achieves >93% in-language reasoning across 60 languages on 6 benchmarks. Key findings: generalization to held-out languages requires broad language coverage, readily available multilingual non-reasoning data, and a strong English reasoning backbone.
Related event: Cohere Open-Sources Tiny Aya L2-Thinker for Native Multilingual Reasoning(4 posts)→
More from Research
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27
- Latent-Foresight: world models that learn to predict their own future — abursuc · 2026-10-27
- UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
- CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27
- arXiv advice from T. Dietterich: wrong primary category will almost surely get your paper on hold — tdietterich · 2026-10-27