BareBones benchmark reveals Vision-Language Models suffer texture bias cliff

Researchers have introduced BareBones, a new benchmark designed to test the geometric comprehension abilities of Vision-Language Models (VLMs). The benchmark uses pixel-level silhouettes to evaluate if VLMs can understand geometric structure independently of visual textures or contextual information. Evaluations of 26 leading VLMs, including GPT-4.1 and Gemini, revealed a significant performance drop when visual textures were removed, a phenomenon termed the "Texture Bias Cliff." AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Highlights potential limitations in current VLMs' geometric reasoning, suggesting a need for models with better grounding in spatial understanding.

RANK_REASON The cluster contains a new academic paper introducing a novel benchmark for evaluating Vision-Language Models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• BareBones
• VLMs
• GPT-4.1
• Gemini
• Claude Sonnet 4.5
• LLaVA
• ImageNet-S
• DIS5K
• ThinObject5K
• PASCAL VOC
• CUB-200
• WTP-Bench
• Aaditya Baranwal

• paper
• other