Reproduce “LoopZip: Visual Token Compression for Efficient Looped 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

Weizhong Huang, Zhicheng Liu, Yan Zhang, Fei Chao, Xiawu Zheng, Liujuan Cao

LoopZip compresses visual tokens for looped vision-language models by forming candidate summaries, selecting tokens with attention from two calls of a shared layer, and adjusting retained values to preserve both calls’ outputs. On LoopVL-1B, it scores 51.09 versus VisionZip’s 48.49 across eleven benchmarks at 87.5% average pruning; this supports the method in tested settings, not a universal solution.

Summary by alphaXiv.

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