Dataset Card for Idis

• Dataset Description
• Paper Information
• Dataset Structure
• Dataset Usage
• License
• Citation

Dataset Description

Idis (Images with distractors) is a VQA benchmark suite for studying how distractors affect the test-time scaling of reasoning vision-language models. Starting from two base datasets, we add distractors while keeping the target and the answer unchanged, and vary them along three axes: modality (visual and linguistic), number (1 to 4), and semantic relationship to the target (aligned / conflicting / irrelevant). To ensure dataset quality, we perform iterative human verification until all samples in the benchmark satisfy subtask-specific acceptance criteria.

• Idis-perception is built on ImageNet-9 (4,050 images, 9 classes). Objects are inserted with an image editor or a text segment is rendered directly into the image.
• Idis-math is built on MathVerse testmini (3,152 problems). Shapes, handwritten expressions or tables are placed next to the diagram, or distractor sentences are added to the question.

Paper Information

• Paper: https://arxiv.org/abs/2511.21397
• Project page: https://idis-rvlms.github.io/
• Code: https://github.com/effl-lab/Idis

Dataset Structure

Idis-perception/
  visual_distractor/<class>/<n>/<aligned|conflicting|irrelevant>/<stem>.png
  textual_distractor/<class>/4/conflicting/<stem>.png
Idis-math/
  visual_distractor/<aligned|conflicting|irrelevant>/{meta.jsonl, n1..n4/}
  typographic/<handwritten|mathwriting>/{meta.jsonl, n1..n4/}
  textual_distractor/{meta.jsonl, n1..n4/meta.jsonl}

Split	Distractor
Idis-perception	visual (aligned / conflicting / irrelevant)
Idis-perception	typographic (class name rendered into images)
Idis-math	visual (aligned / conflicting / irrelevant)
Idis-math	typographic (handwritten notes / MathWriting expressions)
Idis-math	textual (distractor sentences added to the question)

<class> is the ImageNet-9 class directory (00_dog ... 08_fish), <n> the number of distractors and <stem> the ImageNet file name, so every image pairs with its original. Each meta.jsonl maps a MathVerse sample to its augmented images or question; the MathVerse images themselves are available from AI4Math/MathVerse.

Dataset Usage

from huggingface_hub import snapshot_download

root = snapshot_download("Vail-2000/Idis", repo_type="dataset")                 # everything
root = snapshot_download("Vail-2000/Idis", repo_type="dataset",
                         allow_patterns=["Idis-perception/visual_distractor/*"])   # one benchmark

Evaluation scripts that read these folders are in https://github.com/effl-lab/Idis.

License

This project is released under the MIT license.

Citation

@inproceedings{bae2026idis,
  title     = {Understanding the Effects of Distractors on Reasoning Vision-Language Models},
  author    = {Bae, Jiyun and Ok, Hyunjong and Mo, Sangwoo and Lee, Jaeho},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2026}
}