DNBENCH finds 3,275 schema traps

Researchers released DNBENCH, a 3,275-sample benchmark for LLM database normalization from 1NF through BCNF, scoring semantic equivalence, structural accuracy, and logical validity. The arXiv paper reports that MARS, which separates evidence extraction, violation diagnosis, decomposition planning, schema generation, and verification, improved DNB-SCORE 82.0% over a single-prompt baseline. Builders using agents to emit schemas should keep database-aware verification in the loop; benchmark-audit research shows why a passing aggregate score is not enough.