Reproduce “CueClash: Frontier Vision-Language Models Judge 3-D Relations Close to the Picture Plane” 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
Soham Dan
The authors introduce CueClash, a benchmark testing whether models judge spatial relations from image cues or ground layout. It balances image-plane cues against three-dimensional truth across five relations, comparing answers across scene presentations. For medium-tier fits items, GPT-5.5 is accurate on 96.0% of cue-right and 61.0% of cue-wrong items; the pooled result does not establish reliance, and localization is strongest for fits.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.