CORA-Diff: Confidence-Oriented Residual Acceptance for Efficient Diffusion Language Model Inference | AIChainDay