Pothen.
Can Vision-Language Models Tell Where a Cultural Object Comes From?
750 museum objects. Five catalogue fields. One attribution.
How much of the story can models recover?
Nine models, from individual fields to a complete attribution. Choose a metric, then select two models to compare.
Complete-record accuracy
All five fields compatible with the museum record.
Loading model results…
Scores are percentages of valid outputs. Each model attempted 750 objects. Overall: 6,542 valid · 208 invalid. Invalid outputs are excluded from accuracy denominators.
Whiskers: 95% object-bootstrap intervals (10,000 shared resamples). Sorting shows point estimates; overlapping intervals limit fine-grained ranking.
Look beneath the overall score.
Choose any two models above, or change them here.
Each bar is a field-level compatibility rate. A higher field score does not ensure a coherent complete record. Field-specific confidence intervals are not reported in this release.
A world of objects.
Browse the benchmark’s cultural objects. Open an object to explore its museum record.
Loading the collection…
No objects match these filters. Try a different search or reset the collection.
The record is
open to inquiry.
Explore the images, reference alternatives, model predictions, human annotations, judge decisions, and reproducibility code in the dataset release.
Cite Pothen
@misc{jiang2026pothen,
title={Pothen: Can Vision-Language Models Tell Where a Cultural Object Comes From?},
author={Yuechen Jiang and Enze Zhang and Mohsinul Kabir and Chenying Lin and Ying Peng and Fengyuan Liu and Qianqian Xie and Stavroula Golfomitsou and Konstantinos Arvanitis and Sophia Ananiadou},
year={2026},
url={https://arxiv.org/abs/2604.07338}
}