SciFig: Towards Automating Editable Figure Generation for Scientific Papers

A new multi-agent framework called SciFig can generate editable methodology figures from scientific text in about 10 minutes, according to a paper posted on arXiv. The system produces XML-based diagrams that can be revised in standard tools, addressing a long-standing trade-off between visual richness and editability in automated figure generation. The framework, detailed in a preprint by Shraman Pramanick, decomposes figure generation into four stages: planning, layout synthesis, component rendering, and iterative refinement [1]. The output is an XML figure that can be edited in conventional diagramming software and further refined through human or vision-language model feedback [1]. The work arrives as arXiv, the open-access repository that hosts the paper, processes roughly 24,000 submissions per month, underscoring the scale of scientific communication that such tools aim to serve [6]. Existing automation methods have typically forced a choice between structured, editable outputs — such as those produced by TikZ or SVG pipelines — and the polished appearance of raster images generated by image models, which are difficult to revise [1]. SciFig targets both qualities simultaneously. The authors also introduce SciFig-Bench, a human-verified benchmark comprising 435 author-drawn methodology figures drawn from 37 arXiv domains and 15 top-tier AI and machine learning venues [1]. Evaluation was conducted using SciFig-Eval, a four-axis protocol for measuring figure quality. Across seven single-agent and agentic baselines, SciFig recorded the highest scores on all four axes [1]. Qualitative results indicate the system can generalize beyond methodology figures to teaser diagrams and statistical plots [1]. The initial submission, dated January 7, 2026, weighed 2,436 KB; a revised version filed on June 25, 2026, expanded to 20,831 KB [1]. The paper appears on arXiv’s abstract page alongside experimental community tools developed under the arXivLabs framework, which allows third-party collaborators to build features such as bibliographic explorers and code finders directly into the site [4][5]. arXivLabs was formalized in 2020 to enable community innovation while requiring partners to adhere to values of openness, community, excellence, and user data privacy [5]. The SciFig dataset and code are publicly available through the project’s website [1].