Expert-Aware Refusal Steering enhances LLM refusal capabilities

Arxiv CS.AI (top papers)

Paper introduces Expert-Aware Refusal Steering, a method that applies steering vectors to improve LLM refusal of harmful requests. The approach aims to maintain helpfulness while increasing safety.

llm

See 4 more sources

Safety Paradox: How Enhanced Safety Awareness Leaves LLMs Vulnerable to Posterior Attack5 days agoLong P. Hoang, Hai V. Le, Shaoyang Xu, Wei Lu, Wenxuan Zhang

When Autoregressive Consistency Hurts Safety Alignment6 days agoBochen Lyu, Yiyang Jia, Xiaohao Cai, Zhanxing Zhu

Off-Distribution Voices: Fanfiction Subgenres as Universal Vernacular Jailbreaks for Aligned LLMs7 days agoZhongze Luo, Ruihe Shi, Zhenshuai Yin, Haoyue Liu, Weixuan Wan, Xiaoying Tang

Expert-Aware Refusal Steering7 days agoAnna C. Marbut, Daniel R. Olson, Travis J. Wheeler

Arxiv CS.AI (top papers)·Kyungmin ParkTaesup Kim

7 days ago