Reproduce “The Sharp Directions Are Against You: Curvature Analysis of Activation Steering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seyedarmin Azizi, Arya Fayyazi, Parsa Razmara, Erfan Baghaei Potraghloo, Souvik Kundu, Massoud Pedram
The paper introduces sharp removal, a method to stabilize activation steering at inference. It estimates sensitive directions using behavioral-loss curvature with respect to a steering vector, then subtracts its component in the sharpest directions. It outperformed CAA in all 25 generation conditions; on Qwen refusal, rates were 78% baseline, 36% with CAA, and 98% after removal, within tested conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.