Visual Sparse Steering trains sparse autoencoders on frozen CLIP activations to build label-free steering vectors that improve zero-shot classification by up to 4.12 percent via centroid-deviation steering with reconstruction-error gating.
Delta-Adapter extracts a semantic delta from single image pairs to train exemplar-based editors without paired examples, improving accuracy and generalization.
Kernelized Activation Steering lifts activation steering into a reproducing kernel Hilbert space to induce locally adaptive, geometry-aware steering via implicit kernel evaluations, recovering Difference-in-Means as a linear special case and outperforming standard methods on LLM and image control ta