Multi-Agent System Generates Editable Scientific Figures from Diverse Inputs
A new multi-agent harness called Crafter, introduced in arXiv:2605.30611, enables automated generation of publication-quality scientific figures from diverse input types, including text, images, and data. Unlike existing systems that target single figure types under text-only input, Crafter generalizes across figure types and input conditions without architectural changes. A companion system, CraftEditor, converts raster outputs into editable SVGs, addressing the need for local revisions. The work also introduces CraftBench, a benchmark for evaluating figure generation. The systems are designed to handle the structured composition of discrete semantic components in scientific figures, where localized errors from generators require a harness rather than a stronger backbone. The paper is authored by researchers and is available on arXiv.
Key facts
- Crafter is a multi-agent harness for scientific figure generation.
- It generalizes across figure types and input conditions without architectural changes.
- CraftEditor converts raster outputs into editable SVGs.
- CraftBench is introduced as a benchmark for evaluation.
- Existing systems target single figure types under text-only input.
- Scientific figures are structured compositions of discrete semantic components.
- Localized errors require a harness, not a stronger backbone.
- The paper is available on arXiv with ID 2605.30611.
Entities
Institutions
- arXiv