Long-document question answering benchmarks have typically focused on text-only reasoning, leaving a gap when answers depend on figures, tables, or other visual elements. According to a new arXiv paper, few existing datasets combine large-scale human-authored questions with detailed visual evidence annotations, which makes it hard to tell whether a model fails because it cannot find the right evidence or because it misreads what that evidence shows.
The paper introduces VELD-QA, a large-scale dataset built specifically for visual evidence grounding and question answering in long documents. By pairing questions with explicit visual evidence annotations, the dataset is meant to support more precise diagnosis of where systems go wrong—whether in localizing relevant figures or tables or in interpreting their content.
The authors position VELD-QA as a resource for probing failures that earlier benchmarks overlook. Because the source is only an abstract, the paper's exact scale, annotation process, and evaluation results are not detailed here, but the stated aim is clear: to make long-document QA evaluation more sensitive to the role of visual evidence.