Jiajun Fan

Vanguard: Knowing When to Trust the Draft in Speculative Vision-Language Agents

In brief

What the paper does.

Let a small model go first, and it can be right more often than the large model: Vanguard asks a general decision model whether each drafted step of a vision-language agent is right, with the screenshot in view, keeps the best draft when it is confident and calls the large model only when needed. Training-free, and the large model can come from another model family.

Published in
arXiv preprint
Paper
arxiv.org/abs/2611.23917
Project page
jiajunfan.com/projects/vanguard/

Cite

Cite the paper.

@article{mao2026vanguard,
  title={Vanguard: Knowing When to Trust the Draft in Speculative Vision-Language Agents},
  author={Mao, Hanqi and Fan, Jiajun and Guo, Ruihan and Shen, Shuaike and Wang, Ziwen and Ren, Hui and Zhong, Xianrui and Liu, Ge},
  journal={arXiv preprint arXiv:2611.23917},
  year={2026},
  url={https://arxiv.org/abs/2611.23917}
}