BabyVision: 388 Toddler-Level Visual Reasoning Tasks for MLLMs

BabyVision provides 388 visual reasoning tasks across categories such as spatial reasoning and visual tracking to evaluate Multimodal Large Language Models. Created by UnipatAI and released in January 2026, the benchmark targets fundamental visual capabilities typically mastered by three-year-old children.

Use Cases

• Benchmarking MLLMs on 'visual tracking' to assess the model's ability to follow objects over time
• Testing 'spatial reasoning' to evaluate if models can identify the relative positions of objects in an image
• Measuring 'visual pattern recognition' performance to see if models can identify repeating visual structures

Strengths

• 388 curated visual reasoning tasks
• Categorized into 4 distinct cognitive domains: discrimination, spatial, pattern, and tracking
• MIT licensed for open academic and commercial research

Limitations

• Small sample size of 388 tasks limits its utility for model training
• Narrow focus on toddler-level reasoning may not generalize to high-complexity real-world visual scenes

Provenance

Source

UnipatAI (Arxiv: 2601.06521)

Freshness

Last updated January 2026.

Geography

United States