Crafter: Multi-Agent System Redefines Scientific Figure Generation Beyond Single-Model Bottlenecks

Article Summary

The article criticizes the prevailing assumption that a single, larger model can solve the diverse and complex problem of scientific figure generation. Instead, it proposes Crafter, a multi-agent harness that treats figures not as monolithic predictions, but as structured compositions of discrete, locally erroneous components. Crafter uses four specialized roles—including an Intent Reasoner and a Plan Generator—that iterate to produce a figure. This approach allows it to generalize across vastly different figure types (e.g., bar charts from captions, phylogenetic trees from sketches, molecule diagrams from images) and input modalities without requiring fundamental architectural changes, unlike existing systems that are narrowly focused.

Key Points

• Crafter's core innovation is shifting figure generation from a single model prediction task to a coordinated, multi-agent problem-solving process.
• The system uses an Intent Reasoner to create a common semantic language from diverse user inputs (captions, sketches, images), decoupling intent from rendering.
• By generating multiple candidate plans upfront, Crafter explores a broader solution space and mitigates the risks associated with committing to a single, potentially flawed generation path.