COSMIC: Generalized Refusal Direction Identification in LLM Activations

TL;DR

COSMIC is an automated method that identifies and steers large language models towards safer refusal behaviors using cosine similarity, without relying on predefined templates or output analysis.

Contribution

It introduces a novel, output-independent framework for identifying refusal directions in LLM activations, improving robustness and safety in model steering.

Findings

Achieves comparable steering performance to prior methods.

Effectively identifies refusal directions in adversarial and weakly aligned models.

Steers models toward safer behavior with minimal false refusals.

Abstract

Large Language Models (LLMs) encode behaviors such as refusal within their activation space, yet identifying these behaviors remains a significant challenge. Existing methods often rely on predefined refusal templates detectable in output tokens or require manual analysis. We introduce \textbf{COSMIC} (Cosine Similarity Metrics for Inversion of Concepts), an automated framework for direction selection that identifies viable steering directions and target layers using cosine similarity - entirely independent of model outputs. COSMIC achieves steering performance comparable to prior methods without requiring assumptions about a model's refusal behavior, such as the presence of specific refusal tokens. It reliably identifies refusal directions in adversarial settings and weakly aligned models, and is capable of steering such models toward safer behavior with minimal increase in false refusals, demonstrating robustness across a wide range of alignment conditions.