Tag
This paper introduces WINO and WINO+, methods that enable revokable parallel decoding in diffusion LLMs and distill efficient denoising trajectories, significantly improving the quality-speed trade-off.