The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions
Paper • 2502.09674 • Published
Datasets:
ArXiv:
License:
The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.
Safety Residual Space (ICML 2025)
Data artifacts for the ICML 2025 paper The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions by Wenbo Pan et al.
The paper shows that LLM safety alignment writes multiple orthogonal directions into activation space, not a single refusal direction: ablating the dominant direction eliminates refusal entirely, and a trigger-removal attack retains about 40% success while other jailbreaks drop to near zero.
This dataset hosts the evaluation artifacts used in the paper: jailbreak prompt sets (GCG, PAIR, GPTFuzz, ReNeLLM, CodeChameleon, FLIP, trigger-removal) and moderation/toxicity annotations.
Links
- Paper: arXiv:2502.09674
- Code: github.com/wbopan/safety-residual-space
Author: wenbo.io
Raw CSV files are currently in the GitHub repo data directory; a mirror upload here is planned.
- Downloads last month
- 22