acceptodds
Under review as a conference paper at ICLR 2027

Decomposer: Editable Scientific Figure Reconstruction with Evolving Knowledge

21 views

Abstract

Scientific figures are essential for explaining complex methods and communicating research ideas. Although generative models enable automated figure creation, their raster outputs remain difficult to edit precisely and reuse. Recovering editable figures requires preserving heterogeneous visual components, their spatial organization, and their semantic relationships, posing substantial challenges for both reconstruction and evaluation. We introduce Decomposer, a framework that reconstructs scientific figures as faithful, natively editable slides. Decomposer organizes reconstruction into recognition, representation, and composition stages while distinguishing six common component categories: text, equation, image, chart, arrow, and shape. It further incorporates Diagnosis-Guided Refinement, which follows a localize–repair–verify loop to trace visible rendering errors to the responsible stages and components, modify the corresponding decisions, and verify the corrections through rerendering. Beyond correcting individual figures, Evolving Reconstruction Knowledge allows successful experience to benefit subsequent reconstructions by retaining verified experiences as memory, distilling reusable strategies into skills, and consolidating repeatedly validated skills into native framework capabilities. For systematic evaluation, we introduce FigureEditBench, a benchmark comprising 436 complete scientific method figures and six manually curated component-level subsets with 2,760 annotated regions. Coupled with human validation, a fine-grained, rubric-based VLM-as-Judge evaluates both complete figures and local components across multiple dimensions. Experiments demonstrate that Decomposer achieves strong reconstruction performance and generalizes consistently across multiple backbone models. Further ablation studies validate the effectiveness of both Diagnosis-Guided Refinement and Evolving Reconstruction Knowledge.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Computer Vision, Multimodal Learning, Multimodal Agents, Self-Evolving Agents, Visual Content Editing

Related papers

Loading the map…

Discussion (0)

Sign in to comment.