Minimizing Collateral Damage in Activation Steering

2026-05-01

AI Summary

This work addresses the collateral damage often incurred in large language model activation interventions due to the isotropic assumption, which indiscriminately perturbs non-target feature directions. For the first time, the paper formally characterizes this collateral damage mathematically and introduces a non-isotropic optimization framework grounded in empirical second-moment matrices of activations. The approach formulates activation steering as a constrained optimization problem, where perturbation costs are differentially weighted across feature directions. This enables precise alignment with target attributes while substantially mitigating adverse effects on unrelated task performance. The proposed method thus achieves more accurate and safer control over model behavior.

Abstract

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.

Problem

• activation steering
• collateral damage
• feature direction
• large language models
• model alignment

Innovation

• activation steering
• collateral damage
• constrained optimization
• second-moment matrix
• feature alignment