Language Surgery in Multilingual Large Language Models
A technique for controlling language use in multilingual LLMs without retraining.
Project Proposal
The problem
Multilingual LLMs can drift into the wrong language or mix languages, especially when prompted in under-resourced languages.
We wanted to understand:
- How multilingual LLMs organize languages internally
- Whether we can steer which language the model generates without retraining
The approach
The team studied how multilingual LLMs organize languages in their latent space and how representations shift across languages.
They developed Inference-Time Language Control (ITLC), a method for nudging models toward more consistent language outputs without retraining.
Why this matters
Low-resource languages often see less stable outputs, more language switching, and lower quality.
ITLC offers a practical way to improve behavior in underrepresented languages while helping us understand how languages are arranged inside these models.