Minimizing Collateral Damage in Activation Steering

• URL: http://arxiv.org/abs/2605.01167v1
• Date: Fri, 01 May 2026 23:52:54 GMT
• Title: Minimizing Collateral Damage in Activation Steering
• Abstract summary: 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.<n>Standard interventions, such as vector addition, often cause unintended changes in the alignment of activations along other non-target feature directions.<n>In this work, we provide a mathematical formalization of collateral damage and introduce a principled framework that models steering as a constrained optimization problem.
• Score: 26.274685653829792
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• GEOSTEER: Geodesic Optimization for Activation Steering in Large Language Models [9.438320824151157]
  We propose GeoSteer, an optimization-based method for norm-preserving activation steering.<n>We show that GeoSteer consistently improves over state-of-the-art activation steering baselines.<n>These results suggest that norm-preserving steering can be made more effective by replacing predefined one-step edits with adaptive, geometry-aware optimization.
  arXiv  Detail & Related papers  (2026-09-09T16:55:52Z)
• Safety Cost of Steering Vectors Is Separable and Reducible [15.133262969401004]
  Steering vectors can unintentionally compromise a model's safety mechanisms and increase compliance with harmful requests.<n>We show that this safety degradation arises from a separable component in the vector that disrupts the model's safety mechanisms but contributes little to the steering objective.<n>Our method substantially reduces steering-induced safety degradation while preserving the original steering effect with minimal impact on false refusal.
  arXiv  Detail & Related papers  (2026-08-09T00:29:54Z)
• Activation Steering Induces Emergent Misalignment: A More Comprehensive Evaluation [74.17379276939599]
  We show that activation steering can induce broad misalignment, even in the recent Qwen-3.5 series.<n>We characterize properties of AS-induced EM by analyzing key steering-specific factors, including steering magnitude, the low-rank structure of the steering subspace, and the number of epochs during steering-vector construction.
  arXiv  Detail & Related papers  (2026-06-07T15:34:59Z)
• Principled Steering via Null-space Projection for Jailbreak Defense in Vision-Language Models [71.11171576590134]
  Vision-language models can be easily induced by visual jailbreak attacks to generate harmful content.<n>We propose NullSteer, a null-space projected activation defense framework.<n>We show that NullSteer significantly reduces harmful outputs under various jailbreak attacks.
  arXiv  Detail & Related papers  (2026-03-23T15:23:23Z)
• Weight Updates as Activation Shifts: A Principled Framework for Steering [54.70188910511715]
  Activation steering promises to be an extremely parameter-efficient form of adaptation, but its effectiveness depends on critical design choices.<n>We establish a first-order equivalence between activation-space interventions and weight-space updates, deriving the conditions under which activation steering can replicate fine-tuning behavior.<n>This equivalence yields a principled framework for steering design and identifies the post-block output as a theoretically-backed and highly expressive intervention site.
  arXiv  Detail & Related papers  (2026-02-28T02:50:04Z)
• Spherical Steering: Geometry-Aware Activation Rotation for Language Models [15.078810641141295]
  Inference-time steering has emerged as a promising paradigm for controlling language models (LMs) without the cost of retraining.<n>In this work, we explore Spherical Steering, a training-free primitive that resolves this trade-off through activation rotation.<n>Our method rotates activations along a geodesic toward a target direction, guiding the activation toward the target concept while preserving the integrity of the signal.
  arXiv  Detail & Related papers  (2026-02-09T00:15:47Z)
• MirrorLA: Reflecting Feature Map for Vision Linear Attention [49.41670925034762]
  Linear attention significantly reduces the computational complexity of Transformers from quadratic to linear, yet it consistently lags behind softmax-based attention in performance.<n>We propose MirrorLA, a geometric framework that substitutes passive truncation with active reorientation.<n>MirrorLA achieves state-of-the-art performance across standard benchmarks, demonstrating that strictly linear efficiency can be achieved without compromising representational fidelity.
  arXiv  Detail & Related papers  (2026-02-04T09:14:09Z)
• BLOCK-EM: Preventing Emergent Misalignment by Blocking Causal Features [6.495737609776765]
  Emergent misalignment can arise when a language model is fine-tuned on a narrowly scoped supervised objective.<n>We investigate a mechanistic approach to preventing emergent misalignment by identifying a small set of internal features that reliably control the misaligned behavior.
  arXiv  Detail & Related papers  (2026-01-31T15:11:05Z)
• Selective Steering: Norm-Preserving Control Through Discriminative Layer Selection [1.7802147489386628]
  Large language models (LLMs) remain vulnerable to adversarial attacks that elicit harmful behaviors.<n>We propose Selective Steering, which addresses these limitations through two key innovations.<n> Experiments across nine models demonstrate that Selective Steering achieves 5.5x higher attack success rates than prior methods.
  arXiv  Detail & Related papers  (2026-01-27T08:56:25Z)
• Unifying Perplexing Behaviors in Modified BP Attributions through Alignment Perspective [61.5509267439999]
  We present a unified theoretical framework for methods like GBP, RectGrad, LRP, and DTD.<n>We demonstrate that they achieve input alignment by combining the weights of activated neurons.<n>This alignment improves the visualization quality and reduces sensitivity to weight randomization.
  arXiv  Detail & Related papers  (2025-03-14T07:58:26Z)
• A Perceptual Distortion Reduction Framework for Adversarial Perturbation Generation [58.6157191438473]
  We propose a perceptual distortion reduction framework to tackle this problem from two perspectives. We propose a perceptual distortion constraint and add it into the objective function of adversarial attack to jointly optimize the perceptual distortions and attack success rate.
  arXiv  Detail & Related papers  (2021-05-01T15:08:10Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.