Reproduce “Seeing What Matters, Verifying What Pays: Load-Adaptive Speculative Decoding for Vision-Language 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
Hong Nie, Zhangchi Gao, Yuqian Liang, Jun Yu, Bo Wang, Jianqing Sun, Shengping Liu
Uni-LaSpec accelerates vision-language model serving with image-aware block drafting and verification adapted to GPU load. It selects relevant image patches for the drafter and chooses between tree candidates and truncated linear drafts according to verification cost. It matched or exceeded the best fixed policy in 123 of 144 settings; on one bursty Qwen3.5-4B trace, latency fell 47.6% versus linear verification.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.