Reproduce “StackTok: Accelerating VLMs Inference with Budget-Adaptive Visual Token Selection” 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
Zhenbin Wang, Lei Zhang, Lituan Wang, Wei Huang, Yan Wang, Zhenwei Zhang
StackTok selects visual tokens for vision-language models by combining query–vision relevance with vision–vision coverage. It uses a coverage reference curve and entropy-calibrated, state-dependent gate to make coverage primary when support is insufficient and relevance primary otherwise. On LLaVA-NeXT-7B, 160 tokens retain 95.26% of full-token benchmark performance versus MMTok’s 95.18%; this is not a measured speedup, and gains vary by task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.