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

2026-02-04

arXiv.org

AI Summary

This work challenges the prevailing view that the highly anisotropic internal representations of large language models—characterized by a few dimensions exhibiting substantially higher activation than others—are inherently defective. Instead, it reframes these high-activation dimensions as interpretable semantic functional units. The authors introduce a novel, training-free magnitude-thresholding method to identify domain-critical dimensions and propose a “Critical Dimension Steering” paradigm that achieves precise semantic control by intervening exclusively on these key dimensions. Experimental results demonstrate that this approach significantly outperforms full-dimensional manipulation in both domain adaptation and jailbreak defense tasks, offering superior efficiency and interpretability.

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.

Problem

• anisotropy
• massive activations
• interpretable control
• large language models
• domain adaptation

Innovation

• anisotropy
• massive activations
• interpretable control
• domain-critical dimensions
• activation steering