Feynman: Knowledge-Infused Diagramming Agent for Scalable Visual Designs

Researchers have developed Feynman, an AI agent that generates high-quality diagram-caption pairs at scale for training vision-language models. The system created a dataset of 100k+ well-aligned diagrams and introduced Diagramma, a benchmark for evaluating visual reasoning capabilities.

• →Feynman agent automates the creation of knowledge-rich diagram-caption pairs to address data scarcity in vision-language training.
• →The system uses domain-specific knowledge enumeration and code planning to generate diagrams through declarative programming.
• →A dataset of over 100,000 well-aligned diagram-caption pairs was synthesized using this approach.
• →Diagramma benchmark was introduced to evaluate visual reasoning capabilities of vision-language models.
• →The entire agent pipeline, dataset, and benchmark will be released as open-source.