Reproduce “BtsPruner: Role-Aware Visual Token Pruning for Large 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
Zihao Qi, Zhibo Zhang, Miaoxin Zhang, Jincheng Cai, Zengqiang Yan
BtsPruner reduces visual tokens in vision-language models without retraining by selecting a fixed-budget set before language-model processing. It uses attention concentration to identify a transition, selects diverse global representatives from later encoder stages, and recursively retrieves local details from earlier stages. On LLaVA-1.5-7B, it reports 91.6% retention with 32 of 576 tokens; this is a ten-benchmark aggregate, not one-task accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.