Multilingual AI models route languages to separate expert groups — here's how to fix the performance gap

Researchers discovered that when large language models use multiple expert subsystems (a design that lets them process information faster), they tend to use almost completely different experts for different languages. This means low-resource languages get stuck with weaker expert combinations, while high-resource languages get the best ones. A new method called RISE identifies which experts each language needs and fine-tunes only those, improving weak-language performance by up to 10% without hurting others.

Multilingual AI models have been a black box when it comes to language inequality — you could see that some languages performed worse, but you couldn't see why or fix it without retraining the entire model. This paper shows the actual mechanism: routing isolation, which means the model's internal architecture itself learns to segregate languages. RISE matters because it's a cheap fix that doesn't require starting from scratch. If this generalizes beyond the 10 languages tested, companies building multilingual systems now have a tool to guarantee performance across languages instead of accepting that some will always lag.