Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models

TL;DR

This paper explores the intrinsic interpretability of massive activation dimensions in Large Language Models, proposing a simple criterion to identify domain-critical dimensions and demonstrating their effectiveness in domain adaptation and safety scenarios.

Contribution

It introduces a training-free method to identify interpretable, domain-critical dimensions and a targeted activation steering technique that improves domain adaptation and safety interventions.

Findings

Identified domain-critical dimensions behave as semantic detectors.

Activation steering on these dimensions outperforms whole-dimension methods.

The approach enhances domain adaptation and safety in LLMs.

Abstract

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.