Reproduce “Glance, then Select: Evidence-Tempered Visual Token Selection for Extremely Compressed Multimodal LLMs” 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

Yuxuan Wang, Guangxiang Zhao, Yuxuan Tian, Yahui Han, Xinyu Tan, Feiyu Wang, Zongwei Lv, Xiangzheng Zhang and 2 more

Glance-ETD uses a question-guided coarse image pass to allocate high-resolution encoding, then selects visual tokens by balancing answer evidence and scene diversity. At a strict 5% token budget, it achieves 91.6 ± 0.3% normalized retention across fourteen benchmarks, versus 85.6 ± 0.2% with DivPrune. On HR-Bench 4K, TTFT falls from 4.03 to 1.29 seconds, a first-token result rather than a decoding-speed claim.

Summary by alphaXiv.

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