Paper presents geometric account of activation steering via angle-norm decomposition

Arxiv CS.AI (top papers)

The paper proposes a spherical steering paradigm and analyzes activation steering through angle-norm decomposition, addressing limitations of additive interventions. It offers a geometric framework for understanding and improving steering effectiveness.

Activation steering

See 9 more sources

Steer Where It Matters: Token-Level Visual-Sensitivity Steering for LVLMs Hallucination Mitigation5 days agoRuipeng Zhang, Zhihao Li, C. L. Philip Chen, Tong Zhang

DOG-DPO:Dynamic Optimization in Geometry for Safety Alignment5 days agoYi Nian, Tiankai Yang, Yudi Zhang, Qi Pan, Zelong Xu, Shenzhe Zhu, Qingqing Luan, Yue Huang, Xiangliang Zhang, Yue Zhao

IRAF: Interference-Resilient Adaptive Fusion for Noise-Robust End-to-End Full-Duplex Spoken Dialogue Systems6 days agoTao Zhong, Jiajun Deng, Nikita Kuzmin, Yinke Zhu, Tianxiang Cao, Tristan Tsoi, Zhili Tan, Simon Lui, Xunying Liu

What Does Debiasing Really Remove? A Geometric Study of PCA-Based Gender Debiasing in Word Embeddings5 days agoAlexey Kresin, Tchifou M. Dieffi, Tomer Caspi

Rotate2Think: Geometric Priming via Orthogonal Rotation to Improve Language Model Reasoning4 days agoAditya Sharma, Christopher J. Pal, Amal Zouaq

Mechanistic Analysis of Alignment Algorithms in Language Models4 days agoAarush Sinha, Ishan Garg, Veeraraju Elluru, Arth Singh, Kushal Garg

MechLens: Late Crystallization of Factual Knowledge Explains Intervention Effectiveness in Language Models5 days agoXueping Gao

Null-Space Constrained Low-Rank Adaptation for Response-Specified Large Language Model Unlearning4 days agoBocheng Ju, Jianhua Wang, Chengliang Liu, Xiaolin Chang

Measuring, Localizing, and Ablating Alignment Signatures in LLMs13 days agoAniket Anand, Janvijay Singh, Zhewei Sun, Dilek Hakkani-T\"ur, Nick Feamster

Arxiv CS.AI (top papers)·Georgii AparinTatiana Gaintseva

6 days ago