A study examines the performance of Large Language Models (LLMs) in database normalization tasks, specifically focusing on functional dependencies, lossless join decompositions, and inter-table constraints. The research introduces the Database Normalization Benchmark (DNBENCH), containing 3,275 samples, designed to evaluate LLM-driven normalization from 1NF to BCNF. DNBENCH employs a three-axis protocol to assess semantic equivalence, structural accuracy, and logical validity. The benchmark reveals consistent failures in dependency inference, schema decomposition, and inter-table constraint reconstruction across Single, Complex, and Real World levels.
Source: https://arxiv.org/abs/2609.11141