CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence

Hugging Face Daily Papers Papers

Summary

CiteVQA is a benchmark for document vision-language models that evaluates both answer correctness and citation of supporting evidence, revealing widespread attribution hallucinations where models provide correct answers but cite wrong regions.

Multimodal Large Language Models (MLLMs) have significantly advanced document understanding, yet current Doc-VQA evaluations score only the final answer and leave the supporting evidence unchecked. This answer-only approach masks a critical failure mode: a model can land on the correct answer while grounding it in the wrong passage -- a critical risk in high-stakes domains like law, finance, and medicine, where every conclusion must be traceable to a specific source region. To address this, we introduce CiteVQA, a benchmark that requires models to return element-level bounding-box citations alongside each answer, evaluating both jointly. CiteVQA comprises 1,897 questions across 711 PDFs spanning seven domains and two languages, averaging 40.6 pages per document. To ensure fidelity and scalability, the ground-truth citations are generated by an automated pipeline-which identifies crucial evidence via masking ablation-and are subsequently validated through expert review. At the core of our evaluation is Strict Attributed Accuracy (SAA), which credits a prediction only when the answer and the cited region are both correct. Auditing 20 MLLMs reveals a pervasive Attribution Hallucination: models frequently produce the right answer while citing the wrong region. The strongest system (Gemini-3.1-Pro-Preview) achieves an SAA of only 76.0, and the strongest open-source MLLM reaches just 22.5. Ultimately, towards trustworthy document intelligence, CiteVQA exposes a reliability gap that answer-only evaluations overlook, providing the instrumentation needed to close it. Our repository is available at https://github.com/opendatalab/CiteVQA.
Original Article
View Cached Full Text

Cached at: 05/18/26, 10:25 AM

Paper page - CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence

Source: https://huggingface.co/papers/2605.12882 Published on May 13

·

Submitted byhttps://huggingface.co/zr-wang

Wangon May 18

Authors:

,

,

,

,

,

,

,

,

,

Abstract

CiteVQA introduces a benchmark for document vision-language models that evaluates both answer accuracy and correct citation of supporting evidence, revealing significant attribution hallucinations in current models.

Multimodal Large Language Models(MLLMs) have significantly advanceddocument understanding, yet currentDoc-VQAevaluations score only the final answer and leave the supporting evidence unchecked. This answer-only approach masks a critical failure mode: a model can land on the correct answer while grounding it in the wrong passage -- a critical risk in high-stakes domains like law, finance, and medicine, where every conclusion must be traceable to a specific source region. To address this, we introduce CiteVQA, a benchmark that requires models to return element-levelbounding-box citationsalongside each answer, evaluating both jointly. CiteVQA comprises 1,897 questions across 711 PDFs spanning seven domains and two languages, averaging 40.6 pages per document. To ensure fidelity and scalability, the ground-truth citations are generated by an automated pipeline-which identifies crucial evidence viamasking ablation-and are subsequently validated throughexpert review. At the core of our evaluation isStrict Attributed Accuracy(SAA), which credits a prediction only when the answer and the cited region are both correct. Auditing 20 MLLMs reveals a pervasiveAttribution Hallucination: models frequently produce the right answer while citing the wrong region. The strongest system (Gemini-3.1-Pro-Preview) achieves an SAA of only 76.0, and the strongest open-source MLLM reaches just 22.5. Ultimately, towards trustworthy document intelligence, CiteVQA exposes a reliability gap that answer-only evaluations overlook, providing the instrumentation needed to close it. Our repository is available at https://github.com/opendatalab/CiteVQA.

View arXiv pageView PDFProject pageGitHub53Add to collection

Get this paper in your agent:

hf papers read 2605\.12882

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.12882 in a model README.md to link it from this page.

Datasets citing this paper1

#### opendatalab/CiteVQA Viewer• Updated4 days ago • 1.9k • 238 • 5

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.12882 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

ProvenAI: Provenance-Native Traces of Evidence in Generated Answers

arXiv cs.CL

ProvenAI introduces a framework for decomposing transparency in multi-hop question answering into three independently measurable layers: answer correctness, citation fidelity, and per-document influence, revealing a citation-influence gap where cited sources may have weak influence while uncited sources significantly shape the output.

Evidence-Backed Video Question Answering

Hugging Face Daily Papers

This paper introduces Evidence-Backed Video Question Answering (E-VQA), a new task requiring models to output both semantic answers and precise spatio-temporal evidence like tracked object segmentation masklets. The authors create a human-verified benchmark and a scalable training dataset, showing significant improvements over baselines.