Reproduce “SciFigBench: A Benchmark for Scientific Method Figure Understanding” 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

Lijia Wu, Zheng Li, Yuxuan Liu, Jingwen Zhang, Penghai Zhao, Xiang Li

SciFigBench tests AI method figures with questions about labels, connections, and information flow. It contains 4,352 figures and 40,670 question-answer pairs from AI conference papers published in 2024–2025. SciFigCue adds parsed visual cues for arrow locations and directions. On Qwen3.5-VL-9B, it raises accuracy from 50.8% to 56.6% on edge verification; results cover AI method figures, not diagrams across all scientific fields.

Summary by alphaXiv.

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