Tag
The paper proposes a training-free decoding method called Refusal-Aware Early Commitment (RAEC) to improve safety alignment in diffusion language models by leveraging refusal signals from early denoising steps.