LLMs Optimize Database Query Execution Plans

Researchers from Stanford University, the University of Wisconsin–Madison, and Bauplan have demonstrated that large language models (LLMs) can optimize database query execution plans, improving performance without modifying the underlying database engine. Their system, DBPlanBench, utilizes a compact serialization of physical operator graphs to enable LLMs to identify and correct logical flaws in query execution plans. In tests, this method achieved significant speedups, with one query seeing a 4.78x improvement in execution time. The findings suggest that LLMs can serve as effective semantic cardinality estimators, addressing inefficiencies in traditional query optimization methods.