HeiSD : Hybrid Speculative Decoding for Embodied Vision-Language-Action Models with Kinematic Awareness

Zihao Zheng, Zhihao Mao, Sicheng Tian, Jiayu Chen, Maoliang Li, Xinhao Sun, Zhaobo Zhang, Xuanzhe Liu, Donggang Cao, Hong Mei, † Xiang Chen · 2026

Read HeiSD : Hybrid Speculative Decoding for Embodied Vision-Language-Action Models with Kinematic Awareness by Zihao Zheng, Zhihao Mao, Sicheng Tian, Jiayu Chen, Maoliang Li, Xinhao Sun, Zhaobo Zhang, Xuanzhe Liu, Donggang Cao, Hong Mei, † Xiang Chen (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-2603-17573.txt.

Description

Zheng and co-authors present HeiSD, a faster decoding scheme for robot vision-language-action models. Speculative decoding drafts ahead, but drafter-based and retrieval-based forms each have weak spots. HeiSD mixes both and adds a verify-skip step plus looser sequence-level acceptance. A kinematic score sets the switch point between the two modes on its own. Speedups reached 2.45 times in simulation, they report, with task success intact.

Catalog record

CREATORZihao Zheng, Zhihao Mao, Sicheng Tian, Jiayu Chen, Maoliang Li, Xinhao Sun, Zhaobo Zhang, Xuanzhe Liu, Donggang Cao, Hong Mei, † Xiang Chen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63460
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.17573

Browse the full library · About the Tower of Babel