LLMs for Query Optimization

Large language models (LLMs) have fantastic reasoning capabilities in some domains. This project explores how that reasoning can be applied to database query optimization. Core challenges include (1) ensuring that any generated plan is correct, and (2) overcoming high LLM inference overhead. Our work has considered a number of different directions, including embeddings, RAG, and fine-tuning.

Papers

• SEFRQO: A Self-Evolving Fine-Tuned RAG-Based Query Optimizer
  1. Hanwen Liu
  2. Qihan Zhang
  3. Ryan Marcus
  4. Ibrahim Sabek
  SIGMOD '26 (pdf) (doi)
• SERAG: Self-Evolving RAG System for Query Optimization
  1. Hanwen Liu
  2. Qihan Zhang
  3. Ryan Marcus
  4. Ibrahim Sabek
  aiDM@ SIGMOD '25 (pdf) (doi)
• The Unreasonable Effectiveness of LLMs for Query Optimization
  1. Peter Akioyamen
  2. Zixuan Yi
  3. Ryan Marcus
  MLForSystems @ NeurIPS '24 (pdf) (doi)

People