README.md

metadata

license: apache-2.0
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - cognitive-science
  - multimodal
  - vision
  - reasoning
  - webdataset
  - benchmark
  - core-knowledge
  - developmental-psychology
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: id
      dtype: string
    - name: concept
      dtype: string
    - name: stage
      dtype: string
    - name: type
      dtype: string
    - name: question
      dtype: string
    - name: images
      dtype: string
    - name: videos
      dtype: string
    - name: answer
      dtype: string
    - name: choices
      dtype: string
    - name: image_paths
      sequence: image
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
  - config_name: complete
    data_files:
      - split: train
        path: CoreCognition_20250622.zip

CoreCognition: A Core Knowledge Benchmark for Multi-modal Large Language Models

Dataset Description

CoreCognition is a large-scale benchmark encompassing 12 core knowledge grounded in developmental cognitive science, designed to evaluate the fundamental core abilities of Multi-modal Large Language Models (MLLMs).

While MLLMs demonstrate impressive abilities over high-level perception and reasoning, their robustness in the wild remains limited, often falling short on tasks that are intuitive and effortless for humans. We examine the hypothesis that these deficiencies stem from the absence of core knowledge—rudimentary core abilities innate to humans.

This dataset contains 1,423 multimodal samples with images/videos and questions, covering fundamental concepts like object permanence, spatial reasoning, counting, and other core abilities that emerge in human development.

(Additional 80 Concept Hacking questions in our paper will be released separately)

• 🔗 Website: https://williamium3000.github.io/core-knowledge/
• 🔗 Paper: https://arxiv.org/abs/2410.10855
• 🔗 Github: https://github.com/williamium3000/core-knowledge

Formats

1. HuggingFace Preview - For browsing and exploration (visible in HuggingFace viewer, contains embedded 448*448-pixel image preview but no videos)

⚠️ Warning: this format is primarily for HuggingFace viewer; it DOES NOT contain full data.

2. Complete Dataset ZIP (Recommended) - Full data with all images and videos before resizing, 6.41GB

CoreCognition_20250622.zip
├── CoreCognition.csv          # Complete metadata CSV
└── media/                     # All images and videos
    ├── imagename1.png
    ├── imagename2.png
    ├── videoname1.mp4
    └── ...

Quick Start

1. Browse metadata and image preview in this huggingface repo
2. Download the complete dataset (6.41GB) by

from datasets import load_dataset

# this will download huggingface.co/datasets/williamium/CoreCognition/blob/main/CoreCognition_20250622.zip
dataset = load_dataset("williamium/CoreCognition", "complete")

Dataset Fields

Metadata Fields (visible in viewer)

• id: Unique sample identifier
• concept: Core knowledge concept detailed below
• type: Question type ("MC" for multiple choice, "TF" for True/False)
• question: The question text with interleaved <image-placeholder: ...> and/or <video-placeholder: ...>
• images: Semicolon-separated image filenames, can be found in ZIP data
• videos: Semicolon-separated video filenames, can be found in ZIP data
• answer: Correct answer choice
• choices: Choice options as JSON string
• image_paths: Embedded image column for HuggingFace viewer only

Core Knowledge Concepts (12 Categories)

The benchmark covers these fundamental cognitive concepts grounded in developmental science:

• Boundary: The transition from one object to another
• Continuity: Objects persist as unified, cohesive entities across space and time
• Permanence: Objects do not cease to exist when they are no longer perceived
• Spatiality: The a priori understanding of the Euclidean properties of the world
• Perceptual Constancy: Changes in appearances don't mean changes in physical properties
• Intuitive Physics: Intuitions about the laws of how things interact in the physical world
• Perspective: To see what others see
• Hierarchy: Understanding of inclusion and exclusion of objects and categories
• Conservation: Invariances of properties despite transformations
• Tool Use: The capacity to manipulate specific objects to achieve goals
• Intentionality: To see what others want
• Mechanical Reasoning: Inferring actions from system states and vice versa

Paper Citation

If you use CoreCognition in your research, please cite our paper:

@inproceedings{
    li2025core,
    title={Core Knowledge Deficits in Multi-Modal Language Models},
    author={Yijiang Li and Qingying Gao and Tianwei Zhao and Bingyang Wang and Haoran Sun and Haiyun Lyu and Robert D. Hawkins and Nuno Vasconcelos and Tal Golan and Dezhi Luo and Hokin Deng},
    booktitle={Forty-second International Conference on Machine Learning},
    year={2025},
    url={https://openreview.net/forum?id=EIK6xxIoCB}
}

License

Apache 2.0 License