Selective Steering: Norm-Preserving Control Through Discriminative Layer Selection

TL;DR

Selective Steering introduces a norm-preserving, discriminative layer selection method for controlling large language models, significantly improving attack success rates while maintaining model stability and performance.

Contribution

It presents a mathematically rigorous norm-preserving rotation formulation combined with discriminative layer selection, addressing limitations of previous activation steering methods.

Findings

Achieves 5.5x higher attack success rates than prior methods.

Maintains zero perplexity violations and near 100% capability retention.

Demonstrates effectiveness across nine different models.

Abstract

Despite significant progress in alignment, large language models (LLMs) remain vulnerable to adversarial attacks that elicit harmful behaviors. Activation steering techniques offer a promising inference-time intervention approach, but existing methods suffer from critical limitations: activation addition requires careful coefficient tuning and is sensitive to layer-specific norm variations, while directional ablation provides only binary control. Recent work on Angular Steering introduces continuous control via rotation in a 2D subspace, but its practical implementation violates norm preservation, causing distribution shift and generation collapse, particularly in models below 7B parameters. We propose Selective Steering, which addresses these limitations through two key innovations: (1) a mathematically rigorous norm-preserving rotation formulation that maintains activation distribution integrity, and (2) discriminative layer selection that applies steering only where feature representations exhibit opposite-signed class alignment. Experiments across nine models demonstrate that Selective Steering achieves 5.5x higher attack success rates than prior methods while maintaining zero perplexity violations and approximately 100\% capability retention on standard benchmarks. Our approach provides a principled, efficient framework for controllable and stable LLM behavior modification. Code: https://github.com/knoveleng/steering