Gespeichert in:
| Hauptverfasser: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | https://arxiv.org/abs/2506.00085 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866912405354708992 |
|---|---|
| author | Siu, Vincent Crispino, Nicholas Yu, Zihao Pan, Sam Wang, Zhun Liu, Yang Song, Dawn Wang, Chenguang |
| author_facet | Siu, Vincent Crispino, Nicholas Yu, Zihao Pan, Sam Wang, Zhun Liu, Yang Song, Dawn Wang, Chenguang |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2506_00085 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | COSMIC: Generalized Refusal Direction Identification in LLM Activations Siu, Vincent Crispino, Nicholas Yu, Zihao Pan, Sam Wang, Zhun Liu, Yang Song, Dawn Wang, Chenguang Computation and Language Artificial Intelligence 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. |
| title | COSMIC: Generalized Refusal Direction Identification in LLM Activations |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2506.00085 |