作者位次 #3/7ACL 2026Findings· Findings of ACL
PURE : Post-hoc Unlocking and RE finement for Discrete Diffusion Decoding
Yangryeol Park · Kunhui Lee · Hanback Choi · Cheoneum Park · Donghyeon Jeon · Inho Kang · Seung-Hoon Na
摘要
Masked diffusion language models (MDLMs) enable efficient parallel decoding but are limited by a monotonic unmasking policy, where committed tokens cannot be revised. While remasking-based methods mitigate early errors, they mainly intervene during generation. In this work, we study post-hoc refi...