Visual Sanad: Auditing Vision-Language Models for False Acceptance of Qur'anic Citations

Key points

• Vision-language models (VLMs) are increasingly used to interpret religious, educational, and culturally sensitive visual content, yet their reliability on citation verification remains underexplored.1
• The benchmark contains 2,400 rendered images built from 100 English Qur'an verses, four factual states per verse, and six visual conditions ranging from plain text to decorative and authority-framed layouts.1
• De-biased prompting can reduce false acceptance, but often by shifting the decision threshold toward INVALID rather than by producing strong citation discrimination.1
• Visual Sanad therefore exposes a culturally grounded multimodal factuality failure: fluent VLMs may treat displayed religious citations as self-authenticating unless they have strong citation-verification ability.1

Sources (1)

• [1]Visual Sanad: Auditing Vision-Language Models for False Acceptance of Qur'anic Citations
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 02:49 PM

  We introduce Visual Sanad, a controlled benchmark for testing whether VLMs falsely accept incorrect Qur'anic verse attributions when the same quote-reference pair is rendered under different visual framings.

  Vision-language models (VLMs) are increasingly used to interpret religious, educational, and culturally sensitive visual content, yet their reliability on citation verification remains underexplored.

Extractive summary: sentences quoted from the sources.