Reproduce “VSpark: Speculative Decoding for Autoregressive Vision-Language-Action Models” 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

Suho Yoo, Youngjoon Jang, Hyebin Cho, Taeyeop Lee, Jaewook Jeong, Joon Son Chung

VSpark accelerates autoregressive vision-language-action decoding by reusing target predictions from rejected draft branches as candidates under corrected prefixes. It trains the draft with perturbed robot states and target-feature alignment, while keeping exact target verification unchanged. On π₀-FAST across LIBERO, VSpark reaches mean committed length 3.94 versus DSpark’s 3.16 and decoding speedup 1.93× versus 1.60×; these are simulated decoding-time results, not demonstrated task-success or real-world control-frequency gains.

Summary by alphaXiv.

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