arXivPaperNeeds Review
ERUnderstand: Evaluating Vision-Language Models on Structured ER Diagrams
ERUnderstand shows vision-language models can recover common ERD elements but often miss rarer schema constructs, so image-to-schema agent workflows still need structural validation.
arXiv · Jul 27, 2026
Source Summary
ERUnderstand pairs **2,960 diagrams** with standardized machine-readable schemas. Tested vision-language models recovered common elements above **0.74 F1**, but reached only 0.14 on multivalued attributes and 0.07 on N-ary relationships.
Practical Implication
Builders using agents to turn ERD screenshots into migrations or models should validate uncommon constructs explicitly. Reasoning-augmented models improved overall performance by **15–25%**, but did not remove the structural weak spots.
Agent-Ready Context
ERUnderstand pairs **2,960 diagrams** with standardized machine-readable schemas. Tested vision-language models recovered common elements above **0.74 F1**, but reached only 0.14 on multivalued attributes and 0.07 on N-ary relationships. Builders using agents to turn ERD screenshots into migrations or models should validate uncommon constructs explicitly. Reasoning-augmented models improved overall performance by **15–25%**, but did not remove the structural weak spots. The dataset mixes educational, real-world, and synthetic diagrams, so aggregate scores may not match a specific team's notation. Models also remained sensitive to linguistic priors and increasing complexity; the material does not report end-to-end schema-generation accuracy.
Context Map
benchmarkimagedata#agent-evals#agent-reliabilityUncertainty
The dataset mixes educational, real-world, and synthetic diagrams, so aggregate scores may not match a specific team's notation. Models also remained sensitive to linguistic priors and increasing complexity; the material does not report end-to-end schema-generation accuracy.