Hyperphantasia / README.md
shahab7899's picture
Update README.md
2400948 verified
metadata
license: mit
pretty_name: Hyperphantasia
viewer: true
tags:
  - mental_visualization
  - text
  - image
  - benchmark
  - puzzles
task_categories:
  - multiple-choice
  - question-answering
  - visual-question-answering
language:
  - en
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: easy
        path: data/easy-*
      - split: hard
        path: data/hard-*
      - split: medium
        path: data/medium-*
dataset_info:
  features:
    - name: Question
      dtype: string
    - name: ID
      dtype: string
    - name: Image_file
      dtype: image
    - name: Pseudo_solution_file
      dtype: image
    - name: Answer
      dtype: string
    - name: Category
      dtype: string
    - name: Difficulty
      dtype: string
  splits:
    - name: easy
      num_bytes: 6652257
      num_examples: 400
    - name: hard
      num_bytes: 5911842
      num_examples: 400
    - name: medium
      num_bytes: 6091700
      num_examples: 400
  download_size: 10553705
  dataset_size: 18655799

drawing

A Benchmark for Evaluating the Mental Visualization Capabilities of Multimodal LLMs

Mohammad Shahab Sepehri Berk Tinaz Zalan Fabian Mahdi Soltanolkotabi

Github Repository

License

Hyperphantasia is a synthetic Visual Question Answering (VQA) benchmark dataset that probes the mental visualization capabilities of Multimodal Large Language Models (MLLMs) from a vision perspective. We reveal that state-of-the-art models struggle with simple tasks that require visual simulation and imagination. Our dataset consists of 1200 samples with four different puzzles in two categories of Interpolation and Extrapolation.

drawing

Hyperphantasia has three levels of difficulty to evaluate the extent and generalizability of mental visualization capabilities of MLLMs.

drawing

Usage

You can find our evaluation code in our Github repository.

Acknowledgement

We would like to thank Microsoft for an Accelerating Foundation Models Research grant that provided the OpenAI credits enabling this work. This research was also in part supported by AWS credits through an Amazon Faculty Research Award and a NAIRR Pilot Award. M. Soltanolkotabi and MS. Sepehri were supported by the USC–Capital One Center for Responsible AI and Decision Making in Finance (CREDIF) Fellowship. M. Soltanolkotabi is also supported by the Packard Fellowship in Science and Engineering, a Sloan Research Fellowship in Mathematics, an NSF-CAREER under award #1846369, DARPA FastNICS program, and NSF-CIF awards #1813877 and #2008443. and NIH DP2LM014564-01.