#artificial intelligence
Apr 2026
What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal
It is shown that steering vectors can be sparsified by up to 85-96% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions, and that different steering methodologies agree on a subset of important dimensions.
Stephen Cheng, Sarah Wiegreffe, Dinesh Manocha
· arXiv.org · 6 citations