Refusal Behavior in Large Language Models: A Nonlinear Perspective

TL;DR

This paper explores refusal behavior in large language models, revealing it as a complex nonlinear phenomenon that varies across architectures and layers, which has implications for AI safety and alignment.

Contribution

It introduces a nonlinear perspective to understanding refusal behavior in LLMs, challenging the linear assumptions and employing dimensionality reduction techniques.

Findings

Refusal behavior varies nonlinearly across models and layers.

Refusal mechanisms are multidimensional and architecture-dependent.

Nonlinear interpretability is crucial for alignment and safety improvements.

Abstract

Refusal behavior in large language models (LLMs) enables them to decline responding to harmful, unethical, or inappropriate prompts, ensuring alignment with ethical standards. This paper investigates refusal behavior across six LLMs from three architectural families. We challenge the assumption of refusal as a linear phenomenon by employing dimensionality reduction techniques, including PCA, t-SNE, and UMAP. Our results reveal that refusal mechanisms exhibit nonlinear, multidimensional characteristics that vary by model architecture and layer. These findings highlight the need for nonlinear interpretability to improve alignment research and inform safer AI deployment strategies.