Strengthening Target-Language Features: SAE-Based Steering for Multilingual Inference

Hongsheng Wang, Philipp Koehn

Multilingual large language models exhibit substantial performance differences across languages, while existing adaptation methods often require parameter updates and considerable…