Language Steering for Multilingual In-Context Learning
Published in arXiv preprint, 2026
arXiv preprint, 2026
A training-free method that adds language vectors, built from activation differences between parallel English and target-language inputs, to the residual stream at inference. Improves multilingual in-context learning across 19 languages (up to +4.9 on MGSM) without translation or extra tokens.
