SciForma: Structure-Faithful Generation of Scientific Diagrams

arXiv 2026 31.9 method

TLDR

SciForma introduces a framework for structure-faithful generation of scientific diagrams using multi-dimensional preference optimization.

Reasoning

The paper's strength lies in its novel decomposition of diagram quality into three structural axes and the development of a tailored optimization method (M-DPO) along with a curated dataset and benchmark. However, its focus is narrow, limited to diagram generation rather than broader scientific discovery, and the abstract is truncated, leaving some details unclear.

Read-first score

Read-first score 31.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 6.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,github

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,evaluation

Topical relevance 42%

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 14.

Keyword Scores

AI for scientific research

research automation

AI scientist

automated scientific discovery

autonomous research agent

automated research

literature review agent

survey generation

automated experimentation

experiment design agent

paper writing agent

scientific discovery agent

Resources

Paper