Diffusion Steering Lens (DSL)

Linearly interpolates a corrupted vision-transformer representation toward a clean reference (x_new = x* + alpha(x - x*)) and patches the result into subsequent submodule outputs before decoding through the model's own generative (diffusion) decoder — a causally-validated alternative to a vanilla Diffusion Lens/Logit Lens for visualizing what a given attention head or MLP directly contributes.