Reproduce “Where Do Vision Language Models Fail at 3D Reasoning?” 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

Khai Le-Duc, Nguyen Kim Hai Bui, An Ngo, Nam D. Hoang, Yueqi Song, Jiarui Liu, Thai Quoc Hoang, Fang Wu and 6 more

BedBench tests whether vision-language models can answer linked questions about generated three-dimensional objects. It compares judgments with exact descriptions and consistency across related answers. On color distribution, Gemma4-31B in Thinking mode scores 7.2% exact match, versus 93.3% for 30 human participants, each answering one randomly drawn instance per task. The findings concern synthetic geometry, not natural or embodied scenes.

Summary by alphaXiv.

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