Reproduce “VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs” 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
Yuan Feng, Qize Yang, Ruizhe Chen, Sibo Song, Haolin He, Muzhi Zhu, Zihan Liu, Yunfei Chu and 4 more
VisionWeave reduces visual-token sequences by choosing locally between four native fine-grained tokens and one learned coarse-grained token for each image or video-frame block. A router makes these choices from visual features, retaining more detail where needed. Across eight benchmarks, Qwen3.8-27B averages 43.0% fewer tokens while retaining 98.9% of native-model performance; serving gains were measured only on a specified long-video workload.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.