Large language models struggle to normalize databases accurately
Can LLMs Normalize Databases? A Benchmark and Multi-Agent Framework for Schema Normalization
Computation and Language
Summary
Making sure databases are organized properly is hard because it needs checking complex rules and relationships between data. The authors created a big test set to see how well large language models do this kind of organization, called normalization. They found these models often make mistakes understanding the necessary rules and breaking the databases into correct parts. To fix this, the authors designed a new system where multiple special agents work together to better understand and organize the database schemas, improving accuracy a lot.
What this means in practice
- •For database engineers: Use the multi-agent framework to improve automated schema normalization and reduce errors in database design workflows.
- •For data integration teams: Enhance schema reconciliation processes by leveraging multi-agent reasoning to detect and fix data dependencies more reliably.
Authors
Dong-Jae Koh, Huisu Kim, SeongHwan Yoon, Lasse M. Jantsch, Chun-Hee Lee, Seonghyeon Lee, Young-Kyoon Suh
Abstract
Large Language Models (LLMs) are increasingly used to generate structured outputs, but their reliability remains unclear when those outputs must satisfy database-level constraints. We study this issue through database normalization, involving reasoning about functional dependencies, lossless join decompositions, and inter-table constraints. We introduce a Database Normalization Benchmark (DNBENCH), comprising 3,275 samples for evaluating LLM-driven database normalization from 1NF to BCNF. DNBENCH uses a three-axis protocol to measure semantic equivalence, structural accuracy, and logical validity. Across Single, Complex, and Real World levels, DNBENCH uncovers recurring failures in dependency inference, schema decomposition, and inter-table constraint reconstruction. We further propose Multi-Agent Reasoning for Schemas (MARS), which separates evidence extraction, violation diagnosis, and decomposition planning from schema generation and verification. MARS improves the DNB-SCORE by 82.0% over the single-prompt baseline. All artifacts will be released upon acceptance.