When does weak-to-strong diffusion actually help?
It all starts with the correction we cannot observe I started reading Weak-to-Strong Diffusion with Reflection (W2SD) because its central proposal is simple enough to state in one line: use the difference between a weak model and a strong model to improve the strong model again. That sounds plausible. It also hides a serious assumption. At diffusion time $t$, let a model estimate a score field $$ s_\theta(x_t,t)=\nabla_{x_t}\log p_t^\theta(x_t). $$Roughly, the score points toward states that the model considers more likely. Suppose the strong model has score $s_s$, while an ideal model has score $s_*$. The correction we want is ...