DAGverse: Building Document-Grounded Semantic DAGs from Scientific Papers

Shu Wan; Saketh Vishnubhatla; Iskander Kushbay; Tom Heffernan; Aaron Belikoff; Raha Moraffah; Huan Liu

cs.AIMar 26, 2026

DAGverse: Building Document-Grounded Semantic DAGs from Scientific Papers

Authors: Shu Wan, Saketh Vishnubhatla, Iskander Kushbay, Tom Heffernan, Aaron Belikoff, Raha Moraffah, Huan Liu

Organizations: Arizona State University, Tempe AZ 85281, USA · Worcester Polytechnic Institute, Worcester MA 01609, USA

Abstract

Directed Acyclic Graphs (DAGs) are widely used to represent structured knowledge in scientific and technical domains. However, datasets for real-world DAGs remain scarce because constructing them typically requires expert interpretation of domain documents. We study Doc2SemDAG construction: recovering a preferred semantic DAG from a document together with the cited evidence and context that explain it. This problem is challenging because a document may admit multiple plausible abstractions, the intended structure is often implicit, and the supporting evidence is scattered across prose, equations, captions, and figures. To address these challenges, we leverage scientific papers containing explicit DAG figures as a natural source of supervision. In this setting, the DAG figure provides the DAG structure, while the accompanying text provides context and explanation. We introduce DAGverse, a framework for constructing document-grounded semantic DAGs from online scientific papers. Its core component, DAGverse-Pipeline, is a semi-automatic system designed to produce high-precision semantic DAG examples through figure classification, graph reconstruction, semantic grounding, and validation. As a case study, we test the framework for causal DAGs and release DAGverse-1, a dataset of 108 expert-validated semantic DAGs with graph-level, node-level, and edge-level evidence. Experiments show that DAGverse-Pipeline outperforms existing Vision-Language Models on DAG classification and annotation. DAGverse provides a foundation for document-grounded DAG benchmarks and opens new directions for studying structured reasoning grounded in real-world evidence.

Figures & tables

Figure 1: Constructing a semantic DAG (Appendix 0.A ) from a scientific paper [ 25 ] .

Figure 2: Overview of DAGverse -Pipeline. Paper collection can be updated routinely.

Panel of Figure 3 — Figure 3: Validation interface artifacts used by experts. Left: reconstructed DAG view for structure checking. Right: evidence panel for node and edge grounding checks.

Model	Pred.	True DAG	DC	SD	EA	Kept	E2E
Qwen3-VL-7B	12	6	50%	52%	96%	3	25%
InternVL3.5-8B	29	9	31%	43%	93%	7	24%
Qwen3-VL-30B	21	12	57%	25%	97%	9	43%
InternVL3.5-30B	37	16	43%	63%	98%	12	32%
GPT-5.1	59	46	78%	95%	98%	41	69%
Gemini2.5-Pro	54	39	72%	95%	100%	39	72%

Table 1: We compare DAGverse -Pipeline with popular open source and propretary VLMs. “Pred.” is the number of figures labeld as DAGs, “True DAG” are actual DAGs. “DC”, “SD”, “EA”, and “E2E” are calculated with respect to each model’s correct examples. Among all models, DAGverse -Pipeline has the highest retention rate, without highest number of examples kept.

Model	arXiv	bioRxiv
Qwen3-VL-7B	5%	7%
InternVL3.5-8B	8%	14%
Qwen3-VL-30B	14%	43%
InternVL3.5-30B	21%	43%
GPT-5.1	67%	71%
Gemini2.5-Pro	73%	79%

Table 2: End-to-end performance of DAGverse -Pipeline and baselines across different paper sources.

Figure 4: DAGverse -1 data card. The left panel summarizes key statistics, including node and edge summaries and the most frequent domain tags. The right panel visualizes the node and edge distribution by source, a domain word cloud, and a top domain count plot. We omit two general tags “Causal Inference” and “Machine Learning” in domain visualizations.

Figure 5: A single semantic DAG in DAGverse can be used in different ways. Using the cirrhosis example [ 25 ] , we demonstrate 3 tasks: text-to-graph recovery from grounded evidence, graph-to-text generation of a coherent causal narrative, and causal question answering over graph structure and semantics.

	DAG Statistics	Data Curation
Dataset	# Domains	Automatic	Real-world	Growing
CLadder [ 10 ]	Synthetic
Corr2Cause [ 11 ]	Synthetic
CausalProbe [ 4 ]	News
CausalGraph2LLM [ 23 ]	Mostly synthetic
ReCAST [ 21 ]	Econ

Table 3: Comparison between our dataset and existing causal DAGs datasets.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Number of Instances	arXiv	bioRxiv
Metadata	2,720,631 (100%)	401,231 (100%)
Metadata (processed)	13,441 (0.49%)	1,187 (0.30%)
Papers (Downloaded)	8,410 (0.31%)	693 (0.17%)
Papers (Candidates)	260 (0.0096%)	688 (0.17%)
Figures (Pre-DAG classification)	2,233 (0.08%)	9,261 (2.31%)
Figures (Post-DAG classification)	1,552 (0.06%)	2,169 (0.54%)

Appendix

Table 4: End-to-end pipeline funnel from raw metadata to validated outputs. For each source, the percentage in parentheses is the retention rate at that step relative to the total number of papers in that source’s metadata.

KG Construction

Knowledge graph construction turns text, tables, images, and databases into structured entities and relationships that machines can query and reason over.

See all 195 papers

EviDAG: Auditable Causal DAG Authoring with Biomedical Literature
Jul 23, 2026Yi-han Sheu, Michael R. Steigman, Yu Zhou +3Data Provenance Evidence-Grounded Reasoning
Massachusetts General Hospital, USA · Harvard University, USA · Northeastern University, USA
DAGent: Evaluate-then-Grow Planning for Deep Research Agents
Sep 30, 2026Hanwen Liu, Yuanfu Sun, Qiaoyu TanMulti-Agent LLM Systems Deep Research Agents
New York University · New York University Shanghai
SciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding
Sep 4, 2026Shenxi Wu, Yuhong Liu, Haosong Zhang +8Document Understanding LLM Evaluation
The Chinese University of Hong Kong · Shanghai Artificial Intelligence Laboratory · New York University +5

Jul 23, 2026J/K move · Enter open · S save

EviDAG: Auditable Causal DAG Authoring with Biomedical Literature

Yi-han Sheu, Michael R. Steigman, Yu Zhou, Bo Wang+2

Massachusetts General Hospital, USA · Harvard University, USA · Northeastern University, USA

Data Provenance Evidence-Grounded Reasoning Causal Structure Learning Causal Inference

Open paper AskPDF arXiv

Constructing causal directed acyclic graphs (DAGs) is a core step in biomedical causal analysis, yet it remains a largely manual process. Analysts must connect study variables to prior literature, evaluate uncertain causal claims, and preserve sufficient provenance for expert review. We present EviDAG, a browser-based system for authoring causal DAGs as auditable, evidence-linked artifacts from biomedical literature. Given free-text descriptions of study concepts, EviDAG creates a reproducible literature snapshot, uses an LLM-based reasoning module to generate structured pairwise causal judgments, links literature-supported judgments to verbatim evidence excerpts, and assembles the judgments into a constraint-checked graph. Each proposed edge includes confidence estimates, provenance, and a reviewable rationale. The interface supports study specification, progress monitoring, evidence review, graph comparison, adjustment-set computation, and export. In evaluations against both compact benchmark DAGs and reference DAGs derived from published literature, EviDAG achieves high edge recall on the literature-based cohort while retaining verifiable evidence trails absent from LLM-only baselines. EviDAG thus reduces the burden of causal DAG curation while making the resulting assumptions auditable, supporting the design, analysis, and interpretation of biomedical studies.

Abstract

KG Construction

Explore similar work

EviDAG: Auditable Causal DAG Authoring with Biomedical Literature

DAGent: Evaluate-then-Grow Planning for Deep Research Agents

SciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding