Refusal Behavior in Large Language Models: A Nonlinear Perspective
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866909456477978624 |
|---|---|
| author | Hildebrandt, Fabian Maier, Andreas Krauss, Patrick Schilling, Achim |
| author_facet | Hildebrandt, Fabian Maier, Andreas Krauss, Patrick Schilling, Achim |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2501_08145 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Refusal Behavior in Large Language Models: A Nonlinear Perspective Hildebrandt, Fabian Maier, Andreas Krauss, Patrick Schilling, Achim Computation and Language Artificial Intelligence 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. |
| title | Refusal Behavior in Large Language Models: A Nonlinear Perspective |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2501.08145 |