CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence

2026-05-12

AI Summary

Current document visual question answering (Doc-VQA) evaluations focus solely on answer correctness while neglecting the accuracy of supporting evidence, potentially leading to untrustworthy model outputs in high-stakes scenarios. This work introduces CiteVQA, a new benchmark that proposes Strict Attribution Accuracy (SAA)—a metric requiring multimodal large language models to simultaneously generate answers and element-level bounding box citations, jointly evaluating both answer and citation region correctness. We construct the first large-scale, cross-domain, bilingual (Chinese–English) dataset with citation annotations, produced via an automated mask-ablation pipeline and refined through expert review. Evaluation of 20 state-of-the-art models reveals that even the strongest closed-source model, Gemini-3.1-Pro-Preview, achieves only 76.0% SAA, while the best open-source model reaches merely 22.5%, exposing a pervasive “attribution hallucination” problem and underscoring the urgent need for traceable document intelligence.

Abstract

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.

Problem

• Evidence Attribution
• Document Intelligence
• Multimodal Large Language Models
• Visual Question Answering
• Trustworthiness

Innovation

• Evidence Attribution
• Multimodal Large Language Models
• Document VQA
• Strict Attributed Accuracy
• Citation Grounding