SJD-VP: Speculative Jacobi Decoding with Verification Prediction for Autoregressive Image Generation

Bingqi Shan, Baoquan Zhang*, Xiaochen Qi, Xutao Li, Yunming Ye, Liqiang Nie · 2026

Read SJD-VP: Speculative Jacobi Decoding with Verification Prediction for Autoregressive Image Generation by Bingqi Shan, Baoquan Zhang*, Xiaochen Qi, Xutao Li, Yunming Ye, Liqiang Nie (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-2603-27115.txt.

Description

Shan and co-authors speed up autoregressive image generation with verification-aware speculative decoding. Speculative Jacobi decoding drafts many tokens at once, but low acceptance wastes the effort. The authors notice that tokens whose chances rise across rounds tend to pass verification. Their SJD-VP method biases sampling toward those rising-chance tokens. The trick is plug-and-play with speculative Jacobi schemes. Benchmarks show faster decoding with better image quality. The gain comes from using the decoding habits, not looser checks.

Catalog record

CREATORBingqi Shan, Baoquan Zhang*, Xiaochen Qi, Xutao Li, Yunming Ye, Liqiang Nie
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45905
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.27115

Browse the full library · About the Tower of Babel