Datasets:
pretty_name: ActiveVision
license: cc-by-4.0
language:
- en
task_categories:
- visual-question-answering
- image-text-to-text
tags:
- benchmark
- visual-reasoning
- iterative-reasoning
- active-vision
- multimodal
- vqa
- evaluation
size_categories:
- n<1K
ActiveVision — An Exam for Active Observers
ActiveVision is a benchmark for iterative visual reasoning: 85 photorealistic items across 17 tasks that cannot be solved from a single glance — the model has to keep returning to the image to scan, trace, and compare. Every scene is generated by a deterministic program and re-rendered photorealistically while preserving the structure, so answers are exact by construction.
Frontier models reach about 10% with pure chain-of-thought (50.6% with an autonomous coding agent), while unaided humans average 96.1%.
🌐 Website · 📄 Paper · 💻 Evaluation code
Examples
 |
 |
The 17 tasks
| Family | Tasks |
|---|---|
| Distributed Scanning — find and count every signal | bounded_face_counting · connected_component_counting · region_counting · singleton_shape_counting · tangled_loop_counting |
| Sequential Traversal — follow a structure, step by step | arrow_chain_following · color_zone_sequencing · line_intersection_sequencing · maze_path_tracing · traversal_point_ordering |
| Visual Attribute Transfer — compare a property across regions | constellation_match_counting · contour_difference_spotting · field_difference_spotting · signal_difference_spotting · silhouette_match_counting · stroke_difference_spotting · stroke_match_counting |
5 items per task, 85 items total. Each item is one still image and one question; answers are integers, letters, or ordered sequences — never small multiple choice, so the guess floor is near zero.
Dataset structure
from datasets import load_dataset
ds = load_dataset("activevisionai/ActiveVision", split="train")
# columns: image, id, task, category, question, answer
| Field | Description |
|---|---|
image |
the photorealistic scene (PNG, up to 1536×1024) |
id |
item id, <TASK-PREFIX>-<index> (e.g. MPT-2) |
task |
one of the 17 task names above |
category |
task family: distributed_scanning, sequential_traversal, visual_attribute_transfer |
question |
the full question, instructing the model to answer in <answer>...</answer> tags |
answer |
ground truth (exact-match scoring, separator-insensitive) |
Evaluation
Score the contents of the model's last <answer>...</answer> block by exact
match after normalizing case, whitespace, and separators ("C, A, G" ≡ "CAG";
integer answers also match numerically, "07" ≡ 7).
Headline results (pure chain-of-thought, best per model)
| Model | Accuracy |
|---|---|
| Human (N=3, unaided) | 96.1% |
| GPT-5.5 (xhigh) | 10.6% |
| Gemini 3.5 Flash (high) | 8.2% |
| Gemini 3.1 Pro (high) | 5.9% |
| Claude Opus 4.7 (max) | 4.7% |
| Claude Fable 5 (max) | 3.5% |
More reasoning does not close the gap: scaling effort moves models along the cost axis, not toward the human band. Tool-using coding agents reach at most 50.6% (Claude Code, Fable 5) at $2.74–$7.63 and 12–15 minutes per item — a human answers unaided in 34 seconds.
Citation
@article{zhang2026exam,
title={An exam for active observers},
author={Zhang, Jiarui and Tao, Muzi and Wang, Shangshang and Liu, Ollie and Ma, Xuezhe and Neiswanger, Willie},
journal={arXiv preprint arXiv:2607.16165},
year={2026}
}