Reproduce “VELD-QA: A Large-Scale Dataset for Visual Evidence Grounding and Question Answering in Long Documents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhivar Sourati, Marianne Menglin Liu, Rongguang Wang, Zheng Wang, Yazhe Hu, Lan Guo, Yuying Wang, Mengqing Guo and 7 more

VELD-QA evaluates long-document question answering by linking human-written questions to supporting pages and visual evidence, including descriptions and transcribed table worksheets. It compares answers with gold pages and top-three retrieval. On the 1,452-question test set, Claude Opus 5.5 scores 81.7% and 74.9%, respectively. The results distinguish retrieval from interpretation; annotation gains rely on question-specific gold evidence, not automation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.