Visual language models show widespread visual deficits on neuropsychological tests

2026-02-06 · Source: Nature Machine Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, long

Summary

A study published in Nature Machine Intelligence on February 6, 2026, reveals that despite their high performance in complex visual reasoning, visual language models (VLMs) exhibit significant deficits in fundamental visual abilities. Researchers systematically evaluated three state-of-the-art VLMs using 51 tests from 6 clinical and experimental psychology batteries, comparing their performance to healthy adults across low, mid, and high visual domains. While VLMs excelled at straightforward object recognition, they showed widespread, clinically significant impairments in low- and mid-level visual concepts like orientation, position, continuity, and occlusion. This suggests that current artificial systems can achieve advanced object recognition without developing the foundational visual understanding that humans acquire implicitly.

Key takeaway

For research scientists developing or deploying VLMs, you should integrate neuropsychological testing into your evaluation pipelines. This will help identify critical gaps in foundational visual understanding, beyond high-level object recognition, ensuring more robust and human-aligned VLM development. Prioritize addressing deficits in elemental visual concepts like orientation and occlusion to improve model reliability.

Key insights

VLMs excel at high-level object recognition but fail at foundational low- and mid-level visual concepts.

Principles

• Human-like vision requires foundational visual concepts.
• Complex object recognition does not imply basic visual understanding.

Method

Three state-of-the-art VLMs were evaluated using 51 neuropsychological tests from 6 clinical batteries, comparing performance to normative human data across visual domains.

In practice

• Use neuropsychological tests to benchmark VLM visual capabilities.
• Focus VLM development on low- and mid-level visual processing.

Topics

• Visual Language Models
• Visual Reasoning
• Neuropsychological Assessment
• Human-AI Alignment
• Model Evaluation

Code references

• Gene-Tangt/neuropsych_vlm_bench

Best for: Research Scientist, AI Researcher, AI Scientist, Computer Vision Engineer

Related on AIssential

• Vision-Language Models vs Human: Perceptual Image Quality Assessment — VLMs can approximate human image quality judgments, but performance varies significantly by perceptual attribute.
• From Pixels to Prompts: Vision-Language Models — Understanding Vision-Language Models requires a clear mental map beyond just knowing buzzwords.
• Would you still call this Dax? Novel Visual References in VLMs and Humans — VLMs overgeneralize novel visual concepts, especially when conflicting with prior knowledge, despite human-like sensitivity to perturbations.
• Diagnosing Visual Ignorance in Vision-Language Models — VLMs' visual ignorance results from internal routing failures and benchmarks that reward language-prior reliance.
• Why Vision Language Models Ignore What They See [Munawar Hayat] - 758 — Current VLMs struggle with physics-based generation, visual grounding, and multi-person image fidelity, necessitating improved training and architectural approaches.
• What does RL improve for Visual Reasoning? A Frankenstein-Style Analysis — RL primarily refines mid-to-late VLM layers, enhancing vision-to-reasoning alignment, not uniform visual perception.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Nature Machine Intelligence.