ConceptContrast Activation Addition / 对比激活加法
CAA
Contrastive Activation Addition, an activation steering method that extracts direction vectors by computing the difference between activations of positive and negative samples, then applies these vectors to model activation layers to guide output behavior.
Timeline (last 90 days)
Sep 18
Mainstream activation steering methods in academia include CAA (Contrast Activation Addition) based on contrastive activation differences, linear probe-based steering approaches, and various variants under the Representation Engineering framework.
Unverified50%