Component and Dimension Sparsity in Transformer Refusal Mechanisms

Activation steering manipulates large language model behavior by intervening on internal activations, but the mechanistic basis of these interventions remains poorly understood.