CiteVQA exposes source attribution failures
researchers introduced citevqa, revealing that while models like gpt-5.4 answer correctly 87.1% of the time, their accuracy plummets to 59% when required to cite precise sources.
citevqa evaluated 20 models on 1,897 questions across 711 pdfs, requiring exact evidence markers rather than just page numbers. gemini-3.1-pro-preview scored 76, while gpt-5.4 fell to 59 due to citation errors. open-source models scored below 23. the study concludes that locating the correct document passage is the primary bottleneck for visual document understanding.