Reproduce “Vanguard: Knowing When to Trust the Draft in Speculative Vision-Language Agents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hanqi Mao, Jiajun Fan, Ruihan Guo, Shuaike Shen, Ziwen Wang, Hui Ren, Xianrui Zhong, Ge Liu

Vanguard improves screenshot-driven agents by having a small model propose four actions and a separate screen-aware judge accept the highest-scoring candidate or send the step to a larger target. On 975 unseen-website Mind2Web steps, Qwen3-VL-2B drafts with a Qwen3-VL-8B target achieve 66.8% step success versus 63.3% for the target alone, with the target writing 23% of steps; this is held-out benchmark evidence, not a live-use guarantee.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.