Stability-Weighted Decoding for Diffusion Language Models

Yue Wu, Jian Huang · 2026

Read Stability-Weighted Decoding for Diffusion Language Models by Yue Wu, Jian Huang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-17068.txt.

Description

Wu and Huang improve diffusion language models, which generate text in parallel by unmasking a fully masked sequence. Existing decoders rely on static confidence scores from a single step, ignoring history and sometimes freeing unstable tokens too early. Their Stability-Weighted Decoding weights token scores by temporal stability, working as a plug-and-play add-on for any score-based policy, and lifts accuracy on code and math benchmarks while staying strong across acceleration ratios.

Catalog record

CREATORYue Wu, Jian Huang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53146
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.17068

Browse the full library · About the Tower of Babel