Major multimodal AI models still struggle with basic visual tasks that toddlers solve easily

State-of-the-art multimodal AI systems lean too heavily on language understanding to compensate for weak visual perception, according to new research from computer vision specialists. A team has created BabyVision, a benchmark designed to measure the core visual reasoning abilities that humans develop long before they learn to speak.

The Problem with Today's Vision Models#

Current multimodal large language models (MLLMs) excel at knowledge-based tasks that combine images and text, but they consistently fail at elementary visual challenges that young children handle without difficulty. These failures suggest that while these systems appear sophisticated, they lack the fundamental visual building blocks that enable human perception. The gap highlights a critical weakness: these AI systems are essentially borrowing linguistic knowledge to mask deficiencies in how they actually see and understand images.

Introducing BabyVision#

Researchers designed the BabyVision benchmark to assess visual capabilities independent of language proficiency. The benchmark contains 388 test items organized into 22 subcategories spanning four major visual skill domains. This structure allows systematic evaluation of different aspects of visual reasoning—from basic object recognition to spatial relationships and visual logic—without relying on linguistic priors to bridge the gaps.

The Performance Gap#

The results are striking. Google's Gemini 3 Pro Preview achieved a score of 49.7 percent on BabyVision, falling below the typical performance of a 6-year-old child. Meanwhile, the average adult score reached 94.1 percent. This disparity reveals that despite their impressive performance on benchmark tasks laden with factual knowledge, leading MLLMs remain far removed from human-level visual understanding. The shortfall isn't a minor edge case—it represents a systemic limitation in how these models process and reason about visual information.

Extending to Generative Models#

Beyond benchmarking discrimination tasks, the researchers explored whether generative vision models could tackle visual reasoning challenges, proposing BabyVision-Gen alongside an automatic evaluation toolkit. This expansion broadens the investigation to understand how different model architectures approach visual problem-solving.

The team has made their code and benchmark data publicly available to enable further research and progress toward genuinely human-like visual perception in AI systems.