Candidate-Aware Decoding for Faster Diffusion Language Models

This training-free framework introduces Confidence-Verified Commit (CVC) to enable early-exit decoding in diffusion language models. By verifying argmax stability, the method avoids premature termination on long chain-of-thought outputs common in existing fixed-region exit gates.