Anchored, Not Graded: Vision-Language Models Fail at Slant-from-Texture Perception

2026-06-04 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

Vision-Language Models (VLMs) demonstrate a significant failure in perceiving surface slant from texture, a task where human perception exhibits systematic, graded biases. Unlike humans and unsupervised Convolutional Neural Networks (CNNs), VLMs, across various families and scales, predict slant only at specific "anchors" such as 0°, ±25°, and ±45°. This anchoring behavior shows little dependence on crucial stimulus properties like field of view, optical slant, or surface curvature. While supervised fine-tuning can partially mitigate this issue, residual anchoring remains. Researchers interpret this as a failure in the VLM's representation-to-output language interface, suggesting an inability to express geometric encoding in a graded form, rather than a complete absence of geometric understanding. This finding indicates that strong performance on high-level vision-language benchmarks does not guarantee sensitivity to low-level geometric cues.

Key takeaway

For Machine Learning Engineers developing or deploying Vision-Language Models, recognize that current VLMs may lack fine-grained geometric perception, specifically for tasks like slant-from-texture. Your models might default to anchored, discrete predictions (e.g., 0°, ±25°, ±45°) rather than continuous, graded responses. If your application requires precise low-level spatial understanding, you should implement targeted supervised fine-tuning and rigorous evaluation to address these inherent limitations.

Key insights

VLMs struggle to express graded geometric perception, anchoring predictions to discrete values.

Principles

• VLMs anchor slant predictions to discrete values.
• High-level VLM success doesn't imply low-level geometric sensitivity.
• Graded geometric perception is a representation-to-output interface challenge.

In practice

• Supervised fine-tuning partially remediates VLM slant anchoring.
• Evaluate VLMs for low-level geometric perception tasks.

Topics

• Vision-Language Models
• Geometric Perception
• Slant from Texture
• Anchoring Bias
• Supervised Fine-tuning
• Psychophysical Biases

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Anchored, Not Graded: Vision-Language Models Fail at Slant-from-Texture Perception — VLMs fail to express continuous geometric perception due to a language readout bottleneck, despite encoding the information.
• [R] Can Vision-Language Models See Squares? Text-Recognition Mediates Spatial Reasoning Across Three Model Families — VLMs excel at text-based spatial tasks but struggle with non-textual visual patterns, indicating a strong implicit OCR pipeline.
• Grid2Matrix: Revealing Digital Agnosia in Vision-Language Models — VLMs exhibit "Digital Agnosia," failing to express fine-grained visual details despite their vision encoders retaining the information.
• Learning Geometric Representations from Videos for Spatial Intelligent Multimodal Large Language Models — GeoVR enables MLLMs to learn 3D spatial intelligence from 2D videos by distilling geometric knowledge via multi-objective learning.
• GeoWeaver: Grounding Visual Tokens with Geometric Evidence before Scene Reasoning — GeoWeaver grounds visual tokens with token-adaptive geometric evidence as a representational prerequisite for spatio-temporal reasoning.
• DiffSpot: Can VLMs Spot Fine-Grained Visual Differences in Web Interfaces? — VLMs struggle with fine-grained visual difference detection in web UIs, achieving only 40.7% accuracy on a new benchmark.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.