Adaptive Visual Autoregressive Acceleration via Dual-Linkage Entropy Analysis

Yu Zhang, Jingyi Liu, Feng Liu, Duoqian Miao, Qi Zhang, Kexue Fu, Changwei Wang, Longbing Cao · 2026

Read Adaptive Visual Autoregressive Acceleration via Dual-Linkage Entropy Analysis by Yu Zhang, Jingyi Liu, Feng Liu, Duoqian Miao, Qi Zhang, Kexue Fu, Changwei Wang, Longbing Cao (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-2602-01345.txt.

Description

Yu Zhang and co-authors speed visual autoregressive nets with NOVA, a no-train token trim guided by doubt. Such nets burn compute on huge token counts, while past trim plans use fixed stage cuts, stiff schedules, and narrow scope, leaving speed untapped. Since doubt shifts track guess risk, NOVA spots the bend point of scale doubt growth live and fires trimming there. Scale- and layer-linked ratios then trim low-doubt tokens per scale and layer, reusing prior-scale caches to hold quality while cutting run time.

Catalog record

CREATORYu Zhang, Jingyi Liu, Feng Liu, Duoqian Miao, Qi Zhang, Kexue Fu, Changwei Wang, Longbing Cao
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32516
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.01345

Browse the full library · About the Tower of Babel