Uncertainty-Aware Art-Historical Dating with Vision-Language Models

Stefanie Schneider; Peter Bell

cs.CVAug 19, 2026

Uncertainty-Aware Art-Historical Dating with Vision-Language Models

Authors: Stefanie Schneider, Peter Bell

Organizations: Marburg University, Biegenstraße 11, 35037 Marburg, Germany

Abstract

Museum and archival datasets do not mirror historical artistic production, but materialize the contingent histories of collecting, preservation, cataloging, and digitization. This has direct consequences for interpreting pretrained image representations: they may appear to encode historical time while actually encoding the institutional conditions under which objects become visible as data. We describe this phenomenon as temporal entanglement and investigate it by formulating artwork dating as an uncertainty-aware regression task over frozen image embeddings. We evaluate several pretrained vision models on a temporally controlled Wikidata corpus of artworks. Our results show that these models contain usable temporal information, with Vision-Language Models (VLMs) outperforming purely visual self-supervised baselines. However, a qualitative analysis indicates that this temporal knowledge is shaped by various biases.

Figures & tables

Panel of Figure 1 — Figure 1: Pipeline for uncertainty-aware temporal estimation of artwork creation dates. From an artwork record, we extract visual features from each image and convert the temporal metadata into scalar midpoint targets for training. Quantile-regression models then estimate the conditional quantiles of the creation year, outputting a central prediction with a calibrated predictive interval.

Backbone	Estimator	Cal.	Point Estimate		Prediction Interval
Backbone			MAE	RMSE	Cov.	Width	$\overline{W}_{\alpha}$	$\cap$ $D_{A}$	$\supseteq$ $D_{A}$
–	Prior		143.6	165.0	82.5	448.0	513.9	85.2	79.2
–	Prior	✓	143.5	165.0	84.3	469.3	517.4	86.7	81.7
CLIP	QR		66.1	89.3	63.7	178.2	329.5	69.0	59.5
	QR	✓	65.7	88.5	80.5	234.5	308.9	82.5	75.7
	kNN		70.4	100.3	85.0	255.8	313.4	87.7	79.8

Table 1: Test-set performance of quantile-regression estimators and

k

-nearest-neighbor baselines. Point-estimation metrics are evaluated against the midpoint year

y_{A}

and reported as MAE and RMSE in years. Interval performance is reported for the

80\text{\,}\mathrm{\%}

prediction interval

\hat{D}_{A}^{(0.10,0.90)}

. Metrics are defined in Sec. 5.1 . Rows marked with ✓ apply split-conformal calibration. Best values are shown in bold.

Century	Point Estimate			Prediction Interval
Century	MAE	RMSE	Bias ( $\bar{e}$ )	Cov.	Width	$\overline{W}_{\alpha}$	$\cap$ $D_{A}$	$\supseteq$ $D_{A}$
1401–1500	90.7	115.3	+87.4	60.0	304.3	483.4	64.3	50.0
1501–1600	76.2	101.2	+66.5	78.0	262.6	351.7	80.0	71.0
1601–1700	56.1	74.1	+32.2	90.8	249.3	272.9	92.7	90.8
1701–1800	55.0	69.1	-32.5	95.4	230.4	241.3	96.3	91.7
1801–1900	58.2	74.3	-43.7	90.2	236.4	249.0	90.2	87.3

Table 2: Test-set performance by century of the SigLIP2-based quantile-regression estimator with split-conformal calibration. Point-estimation metrics are computed against the midpoint year

y_{A}

and reported as MAE , RMSE , and signed bias

\bar{e}

in years. Interval performance is reported for the

80\text{\,}\mathrm{\%}

prediction interval

\hat{D}_{A}^{(0.10,0.90)}

. Metrics are defined in Sec. 5.1 .

Object type	Point Estimate			Prediction Interval
Object type	MAE	RMSE	Bias ( $\bar{e}$ )	Cov.	Width	$\overline{W}_{\alpha}$	$\cap$ $D_{A}$	$\supseteq$ $D_{A}$
Painting	55.9	73.1	-9.8	82.7	234.2	291.0	84.0	78.7
Sculpture	91.6	115.9	+18.8	73.3	290.7	405.3	74.7	70.7
Print	59.6	80.3	-5.0	84.7	229.0	288.8	86.7	81.3
Drawing	63.9	82.0	-2.2	87.3	248.0	293.6	88.0	80.0

Table 3: Test-set performance by object type of the SigLIP2-based quantile-regression estimator with split-conformal calibration. Point-estimation metrics are computed against the midpoint year

y_{A}

and reported as MAE , RMSE , and signed bias

\bar{e}

in years. Interval performance is reported for the

80\text{\,}\mathrm{\%}

prediction interval

\hat{D}_{A}^{(0.10,0.90)}

. Metrics are defined in Sec. 5.1 .

Panel of Figure 2 — Figure 2: Qualitative examples of SigLIP2 date predictions with split-conformal calibration. Filled markers indicate cataloged production dates and hollow markers model predictions. Green denotes low-error cases, with absolute errors of up to $25$ years; red denotes larger errors, with absolute errors of more than $50$ years.

Explore similar work

On the Cultural Anachronism and Temporal Reasoning in Vision Language Models
May 14, 2026Mukul Ranjan, Prince Jha, Khushboo Kumari +1VLM Evaluation VLM Reasoning
MBZUAI, UAE · Inception, UAE
Chronological Blindness: Benchmarking Temporal Reasoning in Vision-Language Models with CHRONOSIGHT
Jun 15, 2026Parthaw Goswami, Jaynto Goswami DeepVLM Evaluation Temporal Reasoning
Department of Computer Science, University of Missouri, Columbia, 65201, MO, USA · SAP, Prague, Czech Republic
MMArt: A Multi-Perspective Multimodal Dataset for Visual Art Understanding
Aug 11, 2026Shuai Wang, Wangyuan Ding, Yixian Shen +5Image Captioning Vision-Language Models
University of Amsterdam, Amsterdam, The Netherlands · Amazon AGI, United States of America · College of Business and Economics, University of Johannesburg, South Africa

May 14, 2026, cs.CVJ/K move · Enter open · S save

On the Cultural Anachronism and Temporal Reasoning in Vision Language Models

Mukul Ranjan, Prince Jha, Khushboo Kumari, Zhiqiang Shen

MBZUAI, UAE · Inception, UAE

VLM Evaluation VLM Reasoning Cultural Heritage

Open paper AskPDF arXiv

Vision-Language Models (VLMs) are increasingly applied to cultural heritage materials, from digital archives to educational platforms. This work identifies a fundamental issue in how these models interpret historical artifacts. We define this phenomenon as cultural anachronism, the tendency to misinterpret historical objects using temporally inappropriate concepts, materials, or cultural frameworks. To quantify this phenomenon, we introduce the Temporal Anachronism Benchmark for Vision-Language Models (TAB-VLM), a dataset of 600 questions across six categories, designed to evaluate temporal reasoning on 1,600 Indian cultural artifacts spanning prehistoric to modern periods. Systematic evaluations of ten state-of-the-art models reveal significant deficiencies on our benchmark, and even the best model (GPT-5.2) achieves only 58.7% overall accuracy. The performance gap persists across varying architectures and scales, suggesting that cultural anachronism represents a significant limitation in visual AI systems, regardless of model size. These findings highlight the disparity between current VLM capabilities and the requirements for accurately interpreting cultural heritage materials, particularly for non-Western visual cultures underrepresented in training data. Our benchmark provides a foundation for enhancing temporal cognition in multimodal AI systems that interact with historical artifacts. The dataset and code are available in our project page.

Abstract

Explore similar work

On the Cultural Anachronism and Temporal Reasoning in Vision Language Models

Chronological Blindness: Benchmarking Temporal Reasoning in Vision-Language Models with CHRONOSIGHT

MMArt: A Multi-Perspective Multimodal Dataset for Visual Art Understanding