Across 26 VLMs and six silhouette datasets, accuracy collapses under RGB deprivation—a Texture Bias Cliff—showing models rely on texture and priors more than pure shape.
Gpt-4.1 technical report
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
BareBones: Benchmarking Zero-Shot Geometric Comprehension in VLMs
Across 26 VLMs and six silhouette datasets, accuracy collapses under RGB deprivation—a Texture Bias Cliff—showing models rely on texture and priors more than pure shape.