Multilingual Steering by Design: Multilingual Sparse Autoencoders and Principled Layer Selection

In three linesSparse autoencoders (SAEs) trained on multilingual data improve language control in LLMs. Authors propose a principled layer-selection rule based on multilingual alignment and language separability, validated on LLaMA-3.1-8B and Gemma-2-9B for machine translation and cross-lingual summarization.

Summary generated by Claude — human-verified