DDRL reportedly yields comparable rewards with or without separate supervised fine-tuning in one test

A user reviewing the work says removing a separate 20,000-step supervised fine-tuning (SFT) stage and running DDRL from pretraining on the same dataset produced comparable rewards and diffusion loss to SFT→DDRL in this setup. They caution that expensive reinforcement-learning rollouts still dominate compute and advise against skipping SFT because RL needs good rollouts. They also note that the human evaluation used only 15 raters, roughly half of them authors.