“The adaptive order policy for masked diffusion was evaluated in two settings: training only the policy with a frozen pre-trained denoiser, and training both the policy and denoiser jointly.”