CiteVQA: AI Excels in Answers but Fails in Accurate Citations

A New Benchmark for Accurate Attribution

Researchers from Peking University and the Shanghai Artificial Intelligence Laboratory have developed an innovative benchmark called CiteVQA. This new tool aims to evaluate the ability of artificial intelligence models to provide not only correct answers but also to accurately cite their sources. This phenomenon, referred to as attribution hallucination, highlights a crucial problem: models can give the right answer while pointing to an incorrect source.

CiteVQA stands out from other document analysis tests like DocVQA or MMLongBench-Doc by requiring that each correct answer be accompanied by an exact citation. A correct answer with an incorrect citation receives a 0 SAA score. In fields such as law, financial audits, or medicine, where traceability is essential, this precision is vital to ensure the usability of AI outputs.

Strict Requirements for Models

The CiteVQA benchmark requires models to justify each claim with a precise reference in the document, whether it be a paragraph, a table, or a figure. A simple page number is not sufficient. The dataset includes 1,897 questions drawn from 711 PDFs covering seven domains, with 451 documents in English and 260 in Chinese. The documents average 40.6 pages, which is significantly longer than most conventional benchmarks.

To construct this dataset, the team established an automated pipeline that breaks down documents into individual elements. Models like Gemini 3.0 Flash then follow the chain of evidence to determine which elements are essential. Each document is tested to see if the model can respond without it, allowing for verification of its importance. The dataset is entirely constructed automatically, and in the final step, the pipeline removes each document one by one to check which ones are truly necessary.

Performance of Current Models

The primary metric used is Strict Attributed Accuracy. A model earns points only if both the answer and the citation are correct. Among the twenty models tested, Gemini-3.1-Pro-Preview achieved the highest score with 76 out of 100. In contrast, GPT-5.4 often found the correct answer but failed to justify its reasoning, scoring 87.1 for raw answer quality but only 59 with correct citations.

Open-source models showed significantly lower results. Qwen3-VL-235B-A22B, the best free model, only reached 22.5 points, while other open-source models often scored below 10, making them "extremely risky" for regulated industries, according to the researchers.

Challenges in Source Localization

Many models struggle to identify the correct page. The Gemini 3 series succeeds in over 87% of cases, while Qwen3-VL-235B-A22B hovers just below 58%. More complex tasks, requiring information from multiple documents, reduce recall for Gemini 3.1 Pro Preview from 69% to 55%.

Academic documents with clear layouts achieve the best scores, while disorganized newspapers and magazines limit performance to around 63 points. In an ablation study, researchers deliberately reduced the search space by providing models only with relevant pages or the correct document. Scores quickly increased—by over 13 points for Qwen3-VL-8B.

Importance of Contextual Engineering

An ablation study showed that restricting the search space improves model scores. By providing only the relevant pages, scores for Qwen3-VL-8B increased by more than 13 points. This demonstrates that precise citations enhance not only transparency but also the accuracy of responses. The researchers have shared their code on GitHub, and the dataset is available on Hugging Face.

Another benchmark from the Shanghai AI Laboratory in 2024 revealed similar difficulties with long documents, underscoring the persistent challenges for language models. Google DeepMind is tackling a related issue with FACTS Grounding, which measures whether responses strictly come from the provided document or if the model incorporates external knowledge. Even Gemini 3 Pro and GPT-5.1 do not achieve reliable scores.

OpenAI recently examined why models guess instead of saying "I don't know." In an analysis, the company presented hallucinations as a problem of systemic incentives. Training and evaluation reward confident answers and punish hesitation. This same dynamic likely fuels the attribution hallucination that CiteVQA now detects in source citations.