Policy Gradient

MDP & Timestep RL for Discrete Diffusion LLMs

A primitive formulation of dLLM RL with denoising timesteps as actions.