Discrete diffusion language models (dLLMs) iteratively denoise masked sequences, updating multiple tokens in parallel per step.
Progressive verification speeds up diffusion language model inference
PV-Decoding uses confidence thresholds and progressive verification to reduce model forward passes by 8–37% while keeping accuracy drop under 1.3%.
Chinese Tech
Renxi Liu · Yiming Li · Shengyin Sun · Chen Chen · Xian Wang · Shuai Yang · +5 more
Huawei Technologies · City University of Hong Kong
Research Digest··2 min read
The authors propose PV-Decoding, a confidence-driven speculative decoding framework for discrete diffusion language models.
§