Learning from failure to tackle hard problems

Diffusion Beats Autoregressive in Data-Constrained Settings