Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models
For LLM practitioners, it turns a known artifact into a practical interpretability and control tool, offering a simple, training-free method for targeted activation steering.
LLMs exhibit massive activations in a few feature dimensions, which are typically seen as artifacts. This work shows these dimensions are interpretable semantic detectors and proposes steering them to outperform whole-dimension steering in domain adaptation and jailbreaking.
Large Language Models (LLMs) exhibit highly anisotropic internal representations, often characterized by massive activations, a phenomenon where a small subset of feature dimensions possesses magnitudes significantly larger than the rest. While prior works view these extreme dimensions primarily as artifacts to be managed, we propose a distinct perspective: these dimensions serve as intrinsic interpretable functional units arising from domain specialization. Specifically, we propose a simple magnitude-based criterion to identify Domain-Critical Dimensions in a training-free manner. Our analyses reveal that such dimensions behave as interpretable semantic detectors for symbolic/quantitative patterns or domain-specific terms. In addition, we introduce Critical Dimension Steering, which applies activation steering exclusively to the identified dimensions. Empirical results show that this approach outperforms conventional whole-dimension steering in domain adaptation and jailbreaking scenarios.