VisRO: Visual Routines

       

This is the VisRO dataset from the paper AdaLoop: Thinking with the Vision Encoder. VisRO is a synthetic visual reasoning dataset of 24 procedurally generated tasks in four categories. Every example is one image, one question and one short reference answer. Each task has four difficulty levels for training and in-distribution evaluation (easy, medium, hard, xhard) and a harder fifth level (xxhard) that is held out for out-of-distribution evaluation. The generator code and the reference scorer are released with the paper's code, and the AdaLoop checkpoints trained on VisRO are on Hugging Face.

VisRO-Core6, the six tasks of the paper's experiments: (a) in-distribution examples at different difficulty levels and (b) out-of-distribution examples.

Split	Per task	All 24 tasks	Difficulty levels
train	100,000	2,400,000	easy / medium / hard / xhard, 25,000 each
eval	600	14,400	easy / medium / hard / xhard, 150 each
ood	600	14,400	xxhard
Total	101,200	2,428,800

Tasks

Category	Tasks
discrimination_grouping	count_clusters, count_points_on_lines, find_the_different, find_the_same, find_the_shadow, overlap_match, overlapping_glyphs
paths_connectivity	flow_free, maze, metro_map, path_tracing
planar_transformations	jigsaw_completion, logic_matrix, mirroring_patterns, reconstruction, rotation_patterns, tangram_reconstruction
spatial_transformations	block_delta, count_3d_blocks, cube_unfold, paper_folding, three_d_views, view_reconstruction, viewpoint_rotation

Per-task, per-split and per-difficulty counts are in counts.json.

Layout

README.md
schema.json
counts.json
SHA256SUMS
assets/visro.png
tasks/<category>/<task>/train/00000.tar ... 00019.tar
tasks/<category>/<task>/eval/00000.tar
tasks/<category>/<task>/ood/00000.tar

There are 528 uncompressed tar shards in WebDataset format. A training shard holds 5,000 examples; an eval or ood shard holds 600. Each example is two consecutive members, <key>.json (its record) and <key>.png (its image), where <key> is the eight-digit index of its id; a shard holds nothing else. The default configuration covers all 24 tasks with the splits train, eval and ood; each task also has a configuration of its own, named after the task.

Records

Every <key>.json member holds one record that follows schema.json:

Field	Meaning
id	visro24/<task>/<split>/<eight-digit index>
task, category	one of the 24 tasks and its category
split	train, eval or ood
difficulty	easy, medium, hard, xhard (train / eval) or xxhard (ood)
question	the question; it states the required answer format
answer	the reference answer, always a string
answer_type	choice (answer is one of options) or free_form
options	the answer options of a choice question; can also list suggested values; empty otherwise
image	the name of the image member in the shard, <key>.png
image_sha256, image_size	the sha256 of the PNG bytes and [width, height]

Examples are shuffled within a shard, so every shard mixes the difficulty levels.

Read a shard

With datasets (the png member becomes an image, the json member the record):

from datasets import load_dataset

maze_eval = load_dataset("Lyy0725/VisRO", "maze", split="eval")
row = maze_eval[0]        # row["png"] (PIL image), row["json"] (the record), row["__key__"]

Or with the standard library:

import json
import tarfile

with tarfile.open("00000.tar", "r:") as shard:
    members = shard.getmembers()
    for meta, image in zip(members[0::2], members[1::2]):
        row = json.loads(shard.extractfile(meta).read())
        png_bytes = shard.extractfile(image).read()
        # row["question"], row["answer"], png_bytes

Generating More Data

VisRO is a fixed release: its items were generated once, then given public ids and shuffled. The generator code regenerates items of the same tasks, difficulty levels and record schema, byte for byte for a given task, level, split and seed, but it does not re-create the published items one for one. The paper's experiments use the six core tasks (count_clusters, flow_free, maze, metro_map, overlapping_glyphs, path_tracing).

License and attribution

The dataset is released under CC BY-SA 4.0. The images contain icons from third-party collections, which keep their own terms:

• game-icons.net icons, CC BY 3.0, by game-icons.net and each icon's individual author;
• OpenMoji graphics, CC BY-SA 4.0 (OpenMoji, the open-source emoji and icon project, HfG Schwäbisch Gmünd and contributors); this share-alike term is why the dataset is CC BY-SA 4.0;
• Microsoft Fluent UI Emoji, MIT, Microsoft Corporation;
• Material Design Icons by Pictogrammers, Apache 2.0.

Text in the images is rendered with the DejaVu Sans fonts. The code repository's NOTICE lists every bundled asset with its license and attribution.

Citation

If you find this dataset useful in your research, please consider citing our paper :)

@article{adaloop2026,
  title={AdaLoop: Thinking with the Vision Encoder},
  author={Han, Xinran Nicole and Lu, Yiyang and Gan, Katelyn and Zickler, Todd},
  journal={arXiv preprint arXiv:2610.21323},
  year={2026}
}

Checksums

SHA256SUMS lists the sha256 of every file in this repository except itself.