What's this about?
Causal-interpretability work showing that boilerplate template tokens in text-to-image diffusion transformers aren't inert — they act as implicit registers where the model stashes semantic information during denoising.
What exactly is a "register token"?
A token the model repurposes as scratch memory rather than for its literal meaning — like a sticky note margin. Known in vision transformers; this shows the same trick emerges in image generation.
Why is it worth noting?
66↑ — interpretability results on generation models are rarer than on classifiers, and this one suggests levers for controlling image generation.