CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence
CiteVQA evaluates evidence attribution in MLLMs using Strict Attributed Accuracy, revealing reliability gaps in answer-only evaluations.
Key Findings
Methodology
CiteVQA employs an automated annotation pipeline to generate a high-fidelity dataset, using MLLMs for evidence chain extraction, ensuring each answer is visually verifiable.
Key Results
- Gemini-3.1-Pro-Preview achieves only 76.0 in Strict Attributed Accuracy, while the strongest open-source model reaches just 22.5, indicating significant attribution hallucination.
- Among 20 mainstream MLLMs, a pervasive attribution hallucination is observed, where models provide correct answers but cite incorrect regions.
- Experiments reveal a significant gap between answer accuracy and evidence attribution, especially in multi-document scenarios.
Significance
CiteVQA provides a new evaluation standard for document intelligence, addressing reliability gaps in answer-only evaluations, crucial for high-stakes domains like law, finance, and medicine.
Technical Contribution
CiteVQA introduces Strict Attributed Accuracy (SAA) as an evaluation standard, combined with an automated annotation pipeline, addressing the cost and consistency bottlenecks of granular visual annotation.
Novelty
CiteVQA is the first to introduce element-level bounding-box citations in document VQA, ensuring joint evaluation of answers and evidence, revealing blind spots in existing evaluation methods.
Limitations
- Current models show weak evidence attribution in multi-document scenarios, especially cross-page evidence linking.
- Automated annotations may have biases, despite expert validation.
Future Work
Future research could explore more efficient evidence chain extraction methods and apply CiteVQA in more domains to enhance multimodal model reliability.
AI Executive Summary
CiteVQA exposes blind spots in existing document VQA evaluation methods, particularly in high-stakes domains where answer accuracy must be combined with evidence traceability. CiteVQA uses an automated annotation pipeline to generate a high-fidelity dataset, employing MLLMs for evidence chain extraction, ensuring each answer is visually verifiable. Experiments show that while some models excel in answer accuracy, they exhibit significant gaps in evidence attribution, especially in multi-document scenarios. This finding provides a new evaluation standard for document intelligence, addressing reliability gaps in answer-only evaluations, and advancing the application of multimodal models in law, finance, and medicine. Future research could explore more efficient evidence chain extraction methods and apply CiteVQA in more domains to enhance multimodal model reliability.
Deep Analysis
Background
In recent years, multimodal large language models (MLLMs) have made breakthrough progress in document understanding, particularly in complex visual layout analysis and cross-modal reasoning. However, existing document VQA evaluation frameworks focus almost exclusively on final answer accuracy, neglecting the logical path through which the model derives that answer—namely, the precise extraction of evidence. In high-stakes domains such as legal consultation, financial auditing, and evidence-based medicine, 'evidence' is the cornerstone of decision-making.
Core Problem
Existing document VQA evaluation methods focus solely on answer accuracy, neglecting evidence traceability. This approach masks the risk of models relying on incorrect evidence to arrive at correct answers, especially in high-stakes domains where answer accuracy must be combined with evidence traceability.
Innovation
CiteVQA introduces Strict Attributed Accuracy (SAA) as an evaluation standard, ensuring joint evaluation of answers and evidence. It uses an automated annotation pipeline to generate a high-fidelity dataset, employing MLLMs for evidence chain extraction, ensuring each answer is visually verifiable.
Methodology
- �� Automated annotation pipeline generates high-fidelity dataset, ensuring data precision and consistency. • Uses MLLMs for evidence chain extraction, ensuring each answer is visually verifiable. • Introduces Strict Attributed Accuracy (SAA) as an evaluation standard, ensuring joint evaluation of answers and evidence.
Experiments
The experimental design includes evaluating 20 mainstream multimodal large language models using the CiteVQA dataset. Evaluation metrics include answer accuracy, evidence attribution's Strict Attributed Accuracy (SAA), recall, and relevance. Experiments reveal a significant gap between answer accuracy and evidence attribution, especially in multi-document scenarios.
Results
Experiments show that while some models excel in answer accuracy, they exhibit significant gaps in evidence attribution, especially in multi-document scenarios. Gemini-3.1-Pro-Preview achieves only 76.0 in Strict Attributed Accuracy, while the strongest open-source model reaches just 22.5, indicating significant attribution hallucination.
Applications
CiteVQA can be used in high-stakes domains like law, finance, and medicine, ensuring each conclusion is traceable to a specific source. Its evaluation standard can enhance the reliability of multimodal models, advancing their application in more domains.
Limitations & Outlook
Current models show weak evidence attribution in multi-document scenarios, especially cross-page evidence linking. Automated annotations may have biases, despite expert validation. Future research could explore more efficient evidence chain extraction methods and apply CiteVQA in more domains to enhance multimodal model reliability.
Plain Language Accessible to non-experts
Imagine you're in a library trying to find specific information in a book. You not only need to find the right book but also the specific paragraph in the book. CiteVQA acts like a smart assistant that not only tells you the answer but also points to the exact location in the book. This is crucial for fields that require precise information, like law and medicine. CiteVQA helps ensure that every answer has a reliable source, not just a guess.
ELI14 Explained like you're 14
Imagine you're playing a treasure hunt game, and you need to find the treasure on the map. CiteVQA is like your compass, not only telling you where the treasure is but also guiding you to find the right path. This is crucial for fields that need precise information, like law and medicine. CiteVQA helps ensure that every answer has a reliable source, not just a guess.
Glossary
Multimodal Large Language Models (MLLMs)
Language models that combine multiple modalities (e.g., text and image), capable of understanding and generating complex multimodal data.
Used for document understanding and evidence chain extraction.
Strict Attributed Accuracy (SAA)
An evaluation standard requiring both the answer and the cited evidence region to be correct.
Used to evaluate models' evidence attribution capability.
Evidence Chain
A collection of evidence supporting an answer, typically involving multiple sources and elements.
Generated through an automated annotation pipeline.
Attribution Hallucination
A phenomenon where models provide correct answers but cite incorrect regions.
Pervasive in experiments.
Automated Annotation Pipeline
An automated process for generating high-fidelity datasets, ensuring data precision and consistency.
Used for CiteVQA dataset construction.
Open Questions Unanswered questions from this research
- 1 How to improve evidence attribution in multi-document scenarios, especially cross-page evidence linking.
- 2 How to reduce biases in automated annotations and improve dataset precision.
Applications
Immediate Applications
Legal Consultation
Helps legal professionals quickly find relevant legal clauses and precedents, improving efficiency.
Long-term Vision
Medical Diagnosis
Assists doctors in finding reliable diagnostic evidence in vast medical literature, improving diagnostic accuracy.
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.