TAB-VLM: Temporal Anachronism Benchmark for Vision-Language Models

Paper: On the Cultural Anachronism and Temporal Reasoning in Vision Language Models (ACL 2026 Findings) Authors: Mukul Ranjan, Prince Jha, Khushboo Kumari, Zhiqiang Shen

TAB-VLM is a benchmark for measuring cultural anachronism in Vision-Language Models — the tendency to misinterpret historical artifacts using temporally inappropriate concepts, materials, or cultural frameworks. The benchmark consists of 600 multiple-choice questions across six categories, grounded in 1,600 Indian cultural artifacts spanning prehistoric to modern periods, sourced from the Museums of India portal.

Systematic evaluation of ten state-of-the-art VLMs on TAB-VLM reveals significant deficiencies in temporal reasoning over cultural heritage. Even the strongest model tested (GPT-5.2) reaches only 58.7% overall accuracy, and the gap persists across architectures and scales — suggesting that cultural anachronism is a structural limitation of current VLMs rather than a capacity issue, particularly for non-Western visual cultures underrepresented in training data.

---

Dataset summary

• Questions: 600 (across 6 reasoning categories)
• Artifacts: 1,600 Indian cultural artifacts (prehistoric to modern)
• Rows in CSV: 3,198 (one row per image; images per question are grouped by question_id)
• Modality: Image + text
• Format: CSV
• Task: Multi-image multiple-choice VQA
• Image source: museumsofindia.gov.in (URLs only; images are not redistributed)

Task

Each question presents a group of artifact images and asks the model to reason about their temporal, cultural, or material relationships — for example, identifying which artifact belongs to a different time period than the others. Unlike single-image VQA, the task inherently requires cross-image comparative reasoning against background knowledge about Indian cultural history.

The six question categories target different axes of temporal and cultural reasoning. One observed category is odd_one_out_period (temporal odd-one-out). Please see the paper for the full taxonomy of the six categories and their construction.

---

Dataset structure

Data fields

Each row corresponds to one image belonging to a question. A full question is reconstructed by grouping rows that share the same question_id.

Field	Type	Description
question_id	string	Unique identifier for the question (shared across all images in that question).
question_type	string	Reasoning category (one of 6).
question_text	string	Natural-language prompt shown to the model.
image_url	string	Public URL of the image on museumsofindia.gov.in.
local_path	string | null	Filesystem path used during dataset construction. May be null.
options	string	Semicolon-separated list of answer choices (typically 4). Each option is a short artifact description like "Coin (Coin)".
correct_answer	string	Zero-indexed position of the correct option in options.

Data splits

Split	Rows
train	3,198

TAB-VLM is released as a single evaluation split. It is intended as a diagnostic benchmark for evaluating existing VLMs, not as a training set.

Example row

question_id	question_type	question_text	options	correct_answer
odd_period_1	odd_one_out_period	Which artifact belongs to a different time period than the others?	Architectural Stone Slab (Sculpture); Seated Annapurna (Archaeology); Head of Divinity (Sculpture); Pot (Archaeology)	1

Reconstructing a question from rows

Because each image for a question is stored on a separate row, group by question_id to assemble the full multi-image question:

import pandas as pd

df = pd.read_csv("path/to/dataset.csv")

grouped = (
    df.groupby("question_id")
      .agg({
          "question_type": "first",
          "question_text": "first",
          "options": "first",
          "correct_answer": "first",
          "image_url": list,
      })
      .reset_index()
)

print(grouped.iloc[0])

---

Usage

Loading with datasets

from datasets import load_dataset

ds = load_dataset("mukul54/tab-vlm", split="train")
print(ds[0])

Parsing options and answers

row = ds[0]
options = [opt.strip() for opt in row["options"].split(";")]
answer_idx = int(row["correct_answer"])
answer_text = options[answer_idx]

Fetching images

Images are hosted externally on museumsofindia.gov.in and referenced via image_url. local_path reflects the authors' build environment and will not resolve on other machines.

import io
import requests
from PIL import Image

img = Image.open(io.BytesIO(requests.get(row["image_url"], timeout=30).content))

Please be respectful of the source server when downloading at scale — rate-limit requests and cache locally.

---

Benchmark results

Ten state-of-the-art VLMs were evaluated in the paper. Headline finding: the best model, GPT-5.2, reaches only 58.7% overall accuracy. The performance gap holds across architectures and scales, indicating that cultural anachronism is not resolved simply by scaling. For the full per-model and per-category breakdown, partial-credit metrics (Kendall's τ, multi-label F1, Jaccard), annotator agreement (pairwise Cohen's κ: human–human 0.92; human–auto 0.85, 0.84), and the three-way failure taxonomy (knowledge gap 29%, visual grounding 8%, relational reasoning 63%), see the paper.

---

Dataset creation

Source data

Images are collected from the Museums of India portal (museumsofindia.gov.in), which hosts digitized artifact records from Indian museums under the Ministry of Culture, Government of India. Each artifact record includes one or more photographs along with metadata such as category, material, and period. The benchmark covers 1,600 artifacts ranging from prehistoric to modern periods.

Annotations

Questions were constructed to isolate a single discriminative attribute (such as time period) while holding other attributes roughly constant, so that a correct answer requires genuine temporal/cultural reasoning rather than surface cues. Options are written as short "<n> (<category>)" phrases derived from the source records. Annotation of model outputs was performed by the authors as a binary correct/incorrect judgment; no external human subjects were involved. See the paper for the full construction protocol.

Personal and sensitive information

The dataset consists exclusively of photographs of inanimate historical artifacts (pottery, sculptures, manuscripts, textiles, coins, etc.) from public museum collections. It contains no personally identifying information or depictions of identifiable individuals.

---

Considerations for using the data

Intended use

• Benchmarking multimodal reasoning in VLMs on non-Western cultural heritage.
• Studying cross-image comparative reasoning along temporal, cultural, and material axes.
• Diagnosing cultural anachronism as a failure mode in visual AI systems.

Limitations

• Artifacts are drawn from a specific set of Indian museum collections and are not a uniform sample of Indian cultural history.
• Labels inherit any inaccuracies present in the source museum catalog metadata.
• image_url points to a live third-party server; link rot is possible over time.
• local_path values reflect the authors' build machine and are not portable.
• As the paper notes, this is a diagnostic evaluation benchmark, not a training resource; its conclusions apply to the models and tasks evaluated and should be interpreted accordingly.

---

Licensing

• Dataset annotations (questions, options, labels): CC BY 4.0, matching the paper's license.
• Images: Images are hosted by museumsofindia.gov.in and are not redistributed in this dataset — only URLs are provided. Any use of the images must follow the terms of the source portal.

Citation

If you use TAB-VLM in your research, please cite:

@inproceedings{ranjan2026tabvlm,
  title     = {On the Cultural Anachronism and Temporal Reasoning in Vision Language Models},
  author    = {Ranjan, Mukul and Jha, Prince and Kumari, Khushboo and Shen, Zhiqiang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026},
  url       = {https://openreview.net/forum?id=N8TknYQbdU}
}

Contact

For questions or issues, please open a discussion on the dataset's Community tab or contact the authors via the paper.