Tag
Researchers propose a training-free method called Suffix-Anchored Confidence Modulation to improve confidence-based decoding in diffusion language models by addressing issues with EOT tokens and premature decoding.