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

arXiv:2610.22009 · cs.CV · Submitted 2026-10-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Visual Sanad: Auditing Vision-Language Models for False Acceptance of Qur'anic Citations".

Jane: The gist The central finding is that many open-weight models show weak discrimination and default to VALID on most images,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’ve seen the setup, and now we’re getting into what the paper actually says about these results. They found a sharp capability split between closed models and open ones in this Visual Sanad study.

Jane: It seems like proprietary systems have a much better handle on citation verification than many of the open-weight models they are currently releasing. They show near-perfect results in some cases, but it’s not universal across the board.

Lu: The paper highlights that these open-weight VLMs often struggle with weak discrimination, and they tend to default to VALID on most images unless they have really strong citation verification ability.

Meng: That sounds like a deployment risk; if you’re relying on an open model for anything sensitive, you can’t just trust it to know when it doesn't know.

Lalam: The key finding they emphasize is this "culturally grounded multimodal factuality failure," where fluent VLMs might treat displayed religious citations as self-authenticating unless they have strong citation-verification ability.

The paper's summary: Tom: That concept of a cultural ground failure is pretty telling. They’re saying it’s not just a simple visual question answering task; it’s about respecting the specific context of religious attribution.

Jane: They break down how they measured this by looking at four closed VLMs and eight open ones under both natural and de-biased prompts, measuring both visual framing effects and prompt-debiasing effects on the same invalid items.

Lu: The paper shows that while visual authority cues aren't overwhelming every model, those cues can shift errors in systematic ways. They distinguish between a general factuality test and a probe of multimodal presentation bias.

Meng: So the finding isn’t just "model X is bad"; it’s more nuanced—it’s about how the presentation of the text affects the model's decision-making process.

Lalam: They also looked at how de-biasing prompts affect things, and they found that for some models, like Qwen3-VL-32B, de-biasing reduced false acceptance but actually made discrimination worse because false rejection increased <ref:2610.22009#pg3>.

The paper's improvements: Tom: Looking at the suggestions for improvement in Visual Sanad, the paper points out that citation verification shouldn't be treated like a generic visual question answering capability. It needs to be handled differently.

Jane: They suggest that systems should either retrieve the referenced verse from a canonical source before judging the image, or they need to abstain entirely when they can’t verify the attribution.

Lu: That makes sense from a representation learning standpoint; if you can’t ground it in external knowledge, you shouldn't guess. Visual Sanad is positioned as a lightweight regression test for retrieval-augmented VLMs and moderation tools.

Meng: For deployment, that means building in a mandatory step where the AI has to check an external database before it outputs a VALID label if the confidence is low.

Lalam: The paper suggests that because religious citations often circulate as images on messaging apps or social media, testing against Qur’anic references helps us see if multimodal systems respect those specific community norms around exact source attribution.

Conclusion: Tom: To wrap up, the main lesson from Visual Sanad is that citation verification in Islamic contexts remains nontrivial for many widely used VLMs. The finding is that strong models can solve this, but many open-weight VLMs still show weak discrimination and a high VALID bias under certain conditions.

Jane: So for anyone building something that answers religious citation questions from screenshots, the paper advises you not to treat it as generic visual question answering. You should either retrieve the reference first or abstain when you can’t verify the attribution.

Lu: The study confirms that while a few models perform very well, the general trend among open models is weak discrimination and sensitivity to visual cues, which we need to address in training or prompting strategies.

Meng: From an engineering standpoint, this gives us a clear test case for where our current production systems are falling short—specifically around how they handle visually framed text.

Lalam: Visual Sanad serves as a valuable regression test for retrieval-augmented VLMs and fact-checking systems because it focuses on the failure mode where these citations circulate.

Tom: That’s the gist of Visual Sanad: it shows that handling specific cultural and textual verification needs specialized multimodal approaches, not just better general vision skills.

Niaz Mohaiman Abtahi, Reyana Shahid, Evan Maksud

Advanced AI Lab Limited · United International University

cs.CV

Submitted: 2026-10-15

Updated: 2026-10-15

Comments: 9 pages, including references and appendix. Accepted for poster presentation at the Muslims in Machine Learning (MusIML) Workshop at NeurIPS 2026. Code: https://github.com/NiazTahi/Visual-Sanad

Code: https://github.com/NiazTahi/Visual-Sanad

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 85/100

The gist: The gist The central finding is that many open-weight models show weak discrimination and default to VALID on most images, while proprietary API-served VLMs range from near-perfect citation

Key concepts

Visual Sanad Benchmark: A dataset of 1,400 images featuring English Qur’an verses paired with four factual states and six visual conditions. It is used to test models' ability to distinguish between valid and invalid citations based on text layout, design, and framing.
False Acceptance Rate: The main metric measuring how often a model incorrectly labels an invalid quote as 'VALID.' This rate reveals the model's tendency to accept incorrect religious citations, which is a critical measure of factual accuracy in this context.
De-biased Prompting: A specific instruction added to the model's request, telling it explicitly not to trust decorative elements like borders or authority headers as proof of correctness. This tests if models can overcome visual cues that might otherwise trick them into accepting false citations.
Culturally Grounded Multimodal Factuality Failure: The central problem where language models fail to correctly judge the authenticity of religious citations because they don't understand the specific cultural norms or verification requirements surrounding these texts. Fluent models often mistake visual presentation for factual truth.

Terminology

Verification — 3× in this paper · explained in 2 episodes
Benchmark — 2× in this paper · explained in 2 episodes
Vision-language models — 2× in this paper · explained in 2 episodes
Vision-Language Models (VLMs) — 1× in this paper · explained in 7 episodes

Summary

The gist The central finding is that many open-weight models show weak discrimination and default to VALID on most images, while proprietary API-served VLMs range from near-perfect citation verification to moderate false acceptance, exposing a culturally grounded multimodal factuality failure where fluent VLMs may treat displayed religious citations as self-authenticating unless they have strong citation-verification ability<ref:2610.22009#pg6>

Dataset Construction

The Visual Sanad benchmark is constructed from 1,400 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<ref:2610.22009#pg7> The dataset contains 65% valid images and 35% invalid images, which yields 1800 invalid images for the main metric<ref:2610.22009#pg8> Each base verse is paired with four factual states: Valid, Near-neighbor, Same-surah far, and Cross-surah lexical<ref:2610.22009#pg9> Invalid foils are screened with exact, character-level, and asymmetric token-containment checks to avoid cases where the target quote is effectively embedded in the foil

Experimental Setup

The task requires models to return VALID, INVALID, or ABSTAIN for a rendered image The main metric is the false-accept rate on invalid trials, defined as the percentage of invalid quote–reference pairs labeled VALID Models are evaluated across four closed VLMs and eight open VLMs under both natural and de-biased prompts The natural prompt asks the model to decide whether the displayed quote belongs to the displayed reference The de-biased prompt adds one cautionary instruction: “Do not treat decorative design, religious-looking borders, or verification headers as evidence that the citation is correct”

Key Findings on Model Performance

Table 1 summarizes the main model comparison showing a sharp capability split where GPT-5.5 shows near-perfect performance and open models often show weak discrimination and default to VALID on most images Closed/API VLMs range from near-perfect citation verification to moderate false acceptance, whereas recent open-weight VLMs often show weak discrimination and default to VALID on most images For instance, GPT-5.5 achieved 99.83% in false accept rate on invalid trials under the natural prompt Open models like Qwen3-VL-32B showed a high false accept rate of 51.96% under the natural prompt

Visual Cue and Prompt Effects

Visual framing effects are smaller than the model-capability effect, but they are not negligible, with authority cues sometimes inducing skepticism and decorative cues often increasing false acceptance For example, InternVL3.5-38B showed a natural-prompt false acceptance of 68.67% on plain invalid images but 89.33% under both Islamic-authority and neutral-authority headers De-biased prompting reduces false acceptance for several models under paired McNemar tests, but Table 1 shows that this does not replace citation verification For Qwen3-VL-32B, the de-biased false-accept rate improved from 51.96% to 43.17%, while discrimination fell from 12.16 to 10.83 because false rejection increased

Implications and Conclusion

The study supports the claim that multimodal citation verification in Islamic contexts remains nontrivial for several widely used VLMs The main lesson is that citation verification should not be treated as a generic visual question-answering capability Systems that answer religious citation questions from screenshots should either retrieve the referenced verse from a canonical source before judging the image, or abstain when they cannot verify the attribution Visual Sanad can therefore serve as a lightweight regression test for retrieval-augmented VLMs, moderation tools, educational assistants, and community-facing fact-checking systems The failure mode matters because religious citations often circulate as images in messaging apps, lecture slides, posters, and social media By focusing on Qur’anic references, Visual Sanad tests whether multimodal systems respect community norms around exact source attribution The study concludes that strong models can solve it, while many open-weight VLMs show weak discrimination, high VALID bias, and model-dependent cue sensitivity

How it works

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 The final dataset has 1800 invalid images, so raw accuracy must be read alongside false acceptance, false rejection, and label bias The models receive one rendered image and must return VALID, INVALID, or ABSTAIN The natural prompt asks the model to decide whether the displayed quote belongs to the displayed Qur’an reference The de-biased prompt adds one cautionary instruction: “Do not treat decorative design, religious-looking borders, or verification headers as evidence that the citation is correct”

Limitations

The study has limitations including using English translations instead of Arabic Qur’anic text It also evaluates rendered static images rather than real social-media screenshots Although it includes several larger open VLMs, the set is still not exhaustive and fast-moving models may change the frontier Furthermore, the dataset is clustered by verse, so downstream statistical models should include verse-level random effects rather than treating 2400 images as independent The authors retain the locked manifest, prompts, audit notes, inference outputs, and analysis scripts used for the reported results The open-source runs use greedy deterministic decoding with do sample=False and fixed seed 31 API runs are reported from cleaned full-dataset CSVs

Community Relevance

The failure mode matters because religious citations often circulate as images in messaging apps, lecture slides, posters, and social media where a model that confidently validates an incorrect attribution may amplify rather than correct misinformation By focusing on Qur’anic references, Visual Sanad tests whether multimodal systems respect community norms around exact source attribution The paper was presented at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026) and Workshop: MusIML@NeurIPS 2026 The arXiv identifier for this work is arXiv:2610.22009v1 [cs.

Improvements for AI systems

Improve citation verification robustness by implementing a retrieval-augmented mechanism for religious citations, as Systems that answer religious citation questions from screenshots should either retrieve the referenced verse from a canonical source before judging the image, or abstain when they cannot verify the attribution. This directly addresses the vulnerability identified in Visual Sanad, which exposes a culturally grounded multimodal factuality failure.
Develop model-specific prompt engineering strategies based on visual cue analysis, utilizing insights such as how authority wording can either induce skepticism or increase trust and how de-biasing shifts the decision threshold. This allows for dynamic adjustment of the model's sensitivity to visual framing, mitigating the model-family dependent effects observed in Table 5.
Implement a fine-grained verification layer specifically targeting high false acceptance rates under authority framing, such as those seen in InternVL3.5-38B, where authority wording is especially consequential. This system would use the visual cues (e.g., Verified Qur’anic Citation) to trigger a mandatory secondary check against a verified source database before outputting a 'VALID' label.