
My team's write-up is out on the @Cohere_Labs community blog.
TLDR: binary benchmarks miss a third category. When multilingual models get confused by code-mixed inputs, they don't refuse or comply. They drift (40+% times)
Thanks to @cohere, @hongyucharlie and my team for support
Cohere Labs@Cohere_Labs
Expedition Aya team captain, @tanav2202, tells the story of how his team came together to identify a critical safety gap in multilingual models when fine-tuned on harmful data and probed with code-mixed inputs. 🚧 Read the full post at labscommunity.cohere.com/blog/2026/mix-…
English

























