From Hallucination to Grounding: Diagnosing Visual Spatial Intelligence via CRISP

Zhixing Li, Yinan Yu · 2026

Read From Hallucination to Grounding: Diagnosing Visual Spatial Intelligence via CRISP by Zhixing Li, Yinan Yu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-26535.txt.

Description

Zhixing Li and Yinan Yu introduce CRISP, a diagnostic test for visual spatial intelligence in vision-language models. Such models answer questions about images. Standard tests mix up two skills: seeing the scene correctly and reasoning about it. A model can also guess well from language habits alone. CRISP separates perception from reasoning using detailed three-dimensional scene graphs.

An oracle intervention protocol lets the testers swap in perfect perception. They can then see whether reasoning alone succeeds. The diagnosis found a systematic disconnect. Leading closed models reason well underneath, but estimate distances poorly. They also fail to use their own internal structural knowledge. Open models showed the opposite weakness. Their multi-step compositional reasoning, chaining several spatial steps, was the bottleneck. The authors offer CRISP as a road map beyond one-score testing.

Catalog record

CREATORZhixing Li, Yinan Yu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.26535

Browse the full library · About the Tower of Babel