VLMs Exhibit Cultural Anachronism in Interpreting Indian Artifacts, New Benchmark Shows
A recent study published on arXiv highlights a critical issue in Vision-Language Models (VLMs) concerning cultural heritage: cultural anachronism, which involves the incorrect interpretation of historical artifacts through concepts that do not align with their time periods. To assess this problem, researchers developed the Temporal Anachronism Benchmark for VLMs (TAB-VLM), consisting of 600 questions across six categories that test temporal reasoning on 1,600 Indian cultural artifacts spanning from prehistoric to modern times. Evaluations of ten advanced models showed notable shortcomings, with the top-performing model (GPT-5.2) reaching merely 58.7% accuracy overall. This performance gap is consistent across different architectures and scales, indicating that cultural anachronism is a pervasive challenge in existing VLMs.
Key facts
- The study defines cultural anachronism as the tendency of VLMs to misinterpret historical objects using temporally inappropriate concepts.
- TAB-VLM is a dataset of 600 questions across six categories.
- The dataset covers 1,600 Indian cultural artifacts spanning prehistoric to modern periods.
- Ten state-of-the-art models were evaluated.
- The best model, GPT-5.2, achieved only 58.7% overall accuracy.
- Performance gaps persist across varying architectures and scales.
- The study suggests cultural anachronism is a systemic issue.
- VLMs are increasingly applied to cultural heritage materials like digital archives and educational platforms.
Entities
Institutions
- arXiv
Locations
- India