ResearchNovember 17, 2026
-18
āļø
Frontier VLMs flunk basic depth
CueClash examines how vision-language models judge 3-D spatial relations near the picture plane.
#VLM#Computer Vision#Benchmark#3D Reasoning#Evaluation

š„ What happened
Soham Dan's CueClash benchmark exposes Claude Opus 5.5, GPT-5.5 and Gemini 3.6 Flash: they lean on 2-D picture cues instead of true 3-D layout. When cues mislead, accuracy drops 18 to 37 points.
š” Why it matters
Given ground coordinates, they hit 99ā100%. The failure is the picture-to-layout step ā GPT-5.5's maps nearly copy the image, Opus corrects depth about as much as humans do.
ā” Our take
This isn't a reasoning gap, it's a perception shortcut. Anyone shipping VLMs for robotics should bolt on pixel measurements with camera depth ā otherwise perspective wins.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions ā when in doubt, read the linked original source.
