residual-acceptance

CORA-Diff is a training-free method that accelerates diffusion language model inference by using native confidence and persistence signals to accept residual positions early, skipping redundant dense denoising passes while preserving task quality.