Absorbing state discrete diffusion models that enable remasking benefit from inference-time scaling, improving sample quality, controlled generation, and performance on downstream tasks.