Language models can now be steered to specific languages using only monolingual data

Researchers found a way to identify which neurons in a language model control which language, using only single-language text and random tokens instead of expensive parallel datasets. This means companies building multilingual AI systems can now control which language the model outputs without needing to collect paired translations of the same content.

Multilingual AI systems have always been hard to control — they'd drift between languages unpredictably because nobody had a reliable map of where language identity lives in the model. This research shows language identity is sparse and localized, which means you can steer it cheaply. The practical effect: smaller teams and lower-budget projects can now build multilingual systems that actually do what they're told.