README.md

metadata

license: cc-by-4.0
language:
  - en
  - zh
task_categories:
  - visual-question-answering
pretty_name: KidVis
size_categories:
  - n<1K
tags:
  - multimodal
  - visual-question-answering
  - benchmark
  - visual-perception
  - cognitive-development
  - multiple-choice

KidVis: Benchmarking Fundamental Visual Primitives in Multimodal Large Language Models

KidVis is a diagnostic benchmark for evaluating foundational visual perception in Multimodal Large Language Models (MLLMs). Unlike semantic-heavy multimodal benchmarks, KidVis focuses on controlled, low-semantic, and motor-free visual tasks inspired by cognitive developmental psychology.

The benchmark decomposes visual perception into six basic visual primitives:

• Visual Concentration
• Visual Tracking
• Visual Discrimination
• Visual Memory
• Visual Spatial
• Visual Closure

Dataset Description

KidVis contains 500 multiple-choice visual questions across 10 task categories. Each task contains 50 questions and is designed to probe one or more foundational visual capabilities.

The 10 task categories are:

1. Body Part Counting
2. Clock Reading
3. Complex Scene Counting
4. Hidden Figures
5. Schulte Grid
6. Spatial Orientation
7. Visual Completion
8. Visual Reasoning
9. Jigsaw Assembly
10. Path Tracing

Each example consists of an image, a bilingual question, and a single correct answer option from A, B, C, and D.

Dataset Structure

Each data instance contains the following fields:

Field	Description
image	The visual question image
subset	The task category identifier
question_id	The question index
question_zh	The Chinese question prompt
question_en	The English question prompt
answer	The correct answer option, one of A, B, C, or D