Reproduce “When Seeing Is Not Enough: Benchmarking Interactive Visual Grounding in LVLMs” 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
Zhengxiang Wang, Owen Rambow
The authors introduce an interactive visual-grounding benchmark that varies how much target description a Matcher receives and whether it can question a Director. Across four controlled visual contexts, eight LVLMs, and 38,848 target selections, NONE was worse than STATIC despite more questions. Gemini-3.1-Pro scored 56.7% versus 87.5%; causes remain unresolved.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.