Minimizing Collateral Damage in Activation Steering
For researchers and practitioners using activation steering to control LLM behavior, this work addresses the key issue of collateral damage, offering a principled improvement over standard isotropic methods.
Activation steering in LLMs often causes unintended changes in non-target features (collateral damage). The authors formalize this problem and propose a constrained optimization method that minimizes expected squared collateral change weighted by the empirical second-moment matrix, achieving more precise control and reducing performance degradation on unrelated tasks.
Activation steering is a method for controlling Large Language Model (LLM) behavior by intervening in its internal representations to increase the alignment with a specific target feature direction. However, standard interventions, such as vector addition, often cause ``collateral damage", defined as unintended changes in the alignment of activations along other non-target feature directions. This damage occurs because standard methods implicitly assume the isotropy of non-target features. In this work, we provide a mathematical formalization of collateral damage and introduce a principled framework that models steering as a constrained optimization problem. Our method finds a new activation that minimizes the expected squared collateral change weighted by the empirical second-moment matrix of activations. This weighting encodes the nonuniform cost of the perturbation in different feature directions, in contrast to isotropic approaches that penalize changes uniformly in all feature directions. By accounting for the empirical second-moment of activations, our approach achieves more precise control while reducing the degradation of model performance on unrelated tasks.