Tag
LocUS is a novel method for targeted activation steering in large language models that restricts interventions to specific attention heads and a subspace of the unembedding matrix, improving performance while preserving general capabilities.