COSMIC: Generalized Refusal Direction Identification in LLM Activations
This work addresses the problem of automated refusal direction identification for LLM safety and alignment, representing an incremental improvement over existing methods by eliminating reliance on predefined templates or manual analysis.
The paper tackles the challenge of identifying refusal behaviors in Large Language Models (LLMs) by introducing COSMIC, an automated framework that uses cosine similarity to select steering directions and target layers without relying on model outputs, achieving performance comparable to prior methods and demonstrating robustness across various alignment conditions with minimal increase in false refusals.
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.