Cord optimizes decoding order for masked diffusion language models

Cord is a lightweight, prompt-conditioned policy that uses geometric residuals to adapt decoding trajectories in masked diffusion language models. Instead of sampling at every step, it learns a single low-dimensional coefficient matrix per rollout to guide the reveal process based on context and proximity.