Datasets:

regap-bench
/

ReGap-Bench

pretty_name: ReGap-Bench
language:
  - en
task_categories:
  - visual-question-answering
tags:
  - multimodal
  - visual-reasoning
  - reference-gap
  - benchmark
size_categories:
  - n<1K
configs:
  - config_name: shape_counting
    data_files:
      - split: test
        path: shape_counting/test-*.parquet
  - config_name: angle_ranking
    data_files:
      - split: test
        path: angle_ranking/test-*.parquet
  - config_name: size_ranking
    data_files:
      - split: test
        path: size_ranking/test-*.parquet
  - config_name: clock_ranking
    data_files:
      - split: test
        path: clock_ranking/test-*.parquet
  - config_name: transformation_prediction
    data_files:
      - split: test
        path: transformation_prediction/test-*.parquet

ReGap-Bench

ReGap-Bench is the benchmark from Diagnosing the Reference Gap: What Visual Scaffolds Can and Cannot Fix in Multimodal Reasoning.

It is a paired reference-intervention benchmark of five visual reasoning tasks. Each scene is rendered under a plain condition and one or more reference conditions (numbering, shading, grid lines), while the scene content, question, and answer stay fixed. Comparing a model's accuracy across conditions isolates the effect of the reference itself.

Subsets

Each subset has 200 scenes in a single test split.

Subset	Question	Answer	Conditions
`shape_counting`	Count shapes of a given class	integer	plain, numbering, shading, grid
`angle_ranking`	Rank angles by size	list of labels	plain, shading, grid
`size_ranking`	Rank shapes by size	list of labels	plain, shading, grid
`clock_ranking`	Rank clocks by time	list of labels	plain, numbering, shading, grid
`transformation_prediction`	Pick the result of a transformation chain	option letter (A–D)	plain, shading, grid

Fields

Field	Description
`image_plain`	The scene without added references.
`image_numbering`	The scene with numbering. Only in `shape_counting` and `clock_ranking`.
`image_shading`	The scene with shaded interiors.
`image_grid`	The scene with grid lines.
`question`	The question asked under every condition.
`answer`	The gold answer, identical across conditions.
`info`	Task-specific metadata (below).

info by subset:

shape_counting: grid_dimensions, total_shape_count, target_shape (labels of the shapes that match the question).
angle_ranking: grid_dimensions, angles_present (angle sizes in degrees, in label order).
size_ranking: grid_dimensions, scales (scale factor of each shape, in label order).
clock_ranking: grid_dimensions, times_present (time shown on each clock, in label order).
transformation_prediction: transform_chain (the transformation steps), options_details (how each option A–D was constructed).

Usage

from datasets import load_dataset

ds = load_dataset("regap-bench/ReGap-Bench", "shape_counting", split="test")
row = ds[0]
row["image_plain"]      # PIL image
row["question"], row["answer"]

Citation

@misc{chen2026diagnosingreferencegapvisual,
      title={Diagnosing the Reference Gap: What Visual Scaffolds Can and Cannot Fix in Multimodal Reasoning}, 
      author={Guizhen Chen and Weiwen Xu and Wotao Yin and Anh Tuan Luu and Hou Pong Chan},
      year={2026},
      eprint={2610.39419},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.39419}, 
}