activevision/hpXgvFBl7ZxO

ActiveVision

An exam for active observers — a benchmark for diagnosing whether multimodal large language models can iteratively look at an image during reasoning, instead of compressing it into a fixed embedding once.

What's in this archive

Path	Contents
data/images/	85 photorealistic PNGs, the released benchmark images
data/manifest.json	Canonical index — one record per instance with id, task, category, image, image_sha256, image_source_filename, question, answer
data/annotations/<task>.jsonl	Per-task verification metadata (5 records per file × 17 tasks = 85). Includes the structural ground truth used to compute each answer (region adjacency, arrow chains, traversal paths, Hausdorff distances, etc.)
code/<category>/<task>/creation.py	Seedable, deterministic generator. The released images at v0.4 are produced at --difficulty 4.
code/<category>/<task>/creation.md	Per-task design and anti-shortcut spec (where present).
code/<category>/<task>/data.json	Per-task definition: shared question text, answer format.
code/gpt_image_prompts.json	One gpt-image-2 image-edit prompt per task, used to re-render the matplotlib structural draft as a photorealistic variant while preserving the discriminative structure.
code/scope.md	Project specification: the three task families and the six shortcut classes the design defeats.
croissant.json	Croissant 1.0 + Croissant-RAI 1.0 metadata for this dataset.
LICENSE	CC BY 4.0.

Statistics

• 85 instances, 17 tasks, 3 task families.
• Distributed Scanning (25 instances, 5 tasks): attribute_group_counting, bounded_faces_counting, counting_connected_components, counting_regions, tangled_loops.
• Sequential Traversal (25 instances, 5 tasks): arrow_chain, color_zone_sequence, line_intersections, maze, traverse_ordering.
• Visual Attribute Transfer (35 instances, 7 tasks): constellation_match_count, contour_silhouette_count, spot_the_contour_diff, spot_the_field_diff, spot_the_signal_diff, spot_the_stroke_diff, stroke_gesture_count.

Loading

import json, pathlib
root = pathlib.Path("data_neurips2026")
manifest = json.loads((root / "data" / "manifest.json").read_text())
for item in manifest:
    image_path = root / item["image"]
    question = item["question"]
    gold     = item["answer"]

Generation pipeline

The pipeline is the artifact. Every benchmark image is produced in two deterministic stages:

1. Geometric draft (matplotlib). creation.py --seed S --difficulty 4 lays out a structural specification (region partition, arrow positions, maze graph, brush-stroke field, etc.) and computes the answer in closed form. Output: a plain matplotlib PNG.
2. Photorealistic re-render (gpt-image-2). The matplotlib draft is sent to OpenAI gpt-image-2 via the image-edit endpoint, with a per-task prompt from code/gpt_image_prompts.json that preserves silhouettes, positions, counts, and labels but replaces the surface material with a photorealistic style (stones on sand, hedge maze from above, starfield, etc.).

The released benchmark contains only the Stage-2 images. Held-out splits with unpublished seeds and additional difficulties can be regenerated from the included generators.

Responsible AI

See croissant.json for the full RAI block. Headlines:

• Synthetic only: 100% synthetic. No human subjects, no PII, no real-world events.
• Use cases: testing and validation. Not for training.
• Limitations: small evaluation set; adversarial-by-design (not predictive of general vision-language ability); photorealistic re-renders depend on a closed-source service.
• License: CC BY 4.0.

Validating the Croissant file

Before submission, validate croissant.json with the official Croissant validator at:

https://huggingface.co/spaces/JoaquinVanschoren/croissant-checker

(Run well in advance of any submission deadline — the doc warns of heavy load near deadlines.)

Citation

Anonymous. ActiveVision: An Exam for Active Observers.
NeurIPS 2026 Datasets and Benchmarks Track (under review).