Abstract: In this talk, I will talk about the reward gap problem, which sees a tradeoff between RL training and diffusion inference. This provides some insights in choosing the level of stochasticity in diffusion generation.