Study finds vision-language models collapse to chance on simple geometric spatial test
A new arXiv paper introduces Sophon, a procedural diagnostic showing that open-weight vision-language models scoring 85-94% on spatial benchmarks like BLINK Spatial and What's Up CLEVR fall to chance on simple geometric tasks.
Read the full story at arXiv cs.CV ↗
Timeline · 1 report
- 2026-11-13 05:00 · arXiv cs.CV
Sophon: A Procedural Diagnostic for Spatial Reasoning in Vision-Language Models