Tag
This paper introduces DARE, a method for improving the inference efficiency of Diffusion Large Language Models by reusing cached key-value and output activations to reduce computational redundancy with negligible quality loss.