Reproduce “Multimodal Language Models Cannot Spot Spatial Inconsistencies” 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
Om Khangaonkar, Hadi J. Rad, Hamed Pirsiavash
The authors introduce a benchmark testing whether an object remains compatible across two views of a static scene. They create pairs from Hypersim scenes by replacing one object using another view, retaining 615 manually reviewed pairs. Humans reach 84.8% accuracy versus 34.2% for GPT-5 low reasoning, with a random-choice baseline of 7.9%. The finding concerns tested models on the forced-choice task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.