Reproduce “DAGverse: Building Document-Grounded Semantic DAGs from Scientific Papers” 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
Shu Wan, Saketh Vishnubhatla, Iskander Kushbay, Tom Heffernan, Aaron Belikoff, Raha Moraffah, Huan Liu
A research team presents DAGverse, a semi-automatic pipeline and dataset for constructing document-grounded semantic Directed Acyclic Graphs (DAGs) from scientific papers. The DAGverse-1 dataset, containing 108 expert-validated causal DAGs with rich textual evidence, demonstrates the pipeline's effectiveness in accurately extracting and contextualizing graph structures, outperforming monolithic Vision-Language Models.
Key insights
- A staged, high-precision pipeline, integrating specialized AI models (VLMs, LLMs) with human oversight, is more effective and robust for document-grounded DAG construction than end-to-end, monolithic models.
- Defining and grounding 'Semantic DAGs' with explicit textual evidence from the source document (node/edge descriptions, aliases, graph context) is crucial for richer, interpretable knowledge representation.
- Distinguishing genuine DAGs from visually similar 'DAG-like negatives' (e.g., flowcharts, model diagrams) is a critical challenge where high-precision classification stages in the pipeline offer a distinct advantage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.