Global Evolutionary Steering: Refining Activation Steering Control via Cross-Layer Consistency

TL;DR

This paper introduces GER-steer, a training-free activation control method for LLMs that improves robustness and generalization by leveraging the network's geometric stability, addressing noise and semantic drift issues.

Contribution

GER-steer is a novel, training-free framework that refines activation steering vectors using global geometric signals, enhancing reliability and universality in model alignment.

Findings

GER-steer outperforms existing methods in various evaluations

It effectively decouples semantic intent from artifacts

Demonstrates superior robustness without layer-specific tuning

Abstract

Activation engineering enables precise control over Large Language Models (LLMs) without the computational cost of fine-tuning. However, existing methods deriving vectors from static activation differences are susceptible to high-dimensional noise and layer-wise semantic drift, often capturing spurious correlations rather than the target intent. To address this, we propose Global Evolutionary Refined Steering (GER-steer), a training-free framework that grounded in the geometric stability of the network's representation evolution. GER-steer exploits this global signal to rectify raw steering vectors, effectively decoupling robust semantic intent from orthogonal artifacts. Extensive evaluations confirm that GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without layer-specific tuning, establishing a universal solution for reliable model alignment.