Reproduce “Disagree to Prune: Reallocating Visual Tokens Where a Small VLM Dissents” 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
Yanchao Li, Jiaqing Xie, Ben Gao, Yusu Zhang, Beilun Wang, Yanbo Wang, Xiaohua Xu, Zhongyu Huang and 2 more
PRUDENT allocates visual tokens adaptively in vision-language answering, using disagreement between a full-image checker and a low-token first answer to select questions for a higher-token re-answer while matching the one-pass average budget. On InternVL3-38B at 2.5% of visual tokens, it reaches 88.4% of full-token scores versus 83.4% for IF-Prune across benchmarks; this is not evidence of lower latency or computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.