Reimagining Text-to-SQL: A Survey of Large Language Model Architectures and Strategies
Summary
- A comprehensive new survey reviews over 100 papers from 2022 to 2026 on Large Language Model architectures and strategies for Text-to-SQL translation.
- The survey introduces a unified taxonomy covering five paradigms: in-context learning, fine-tuning, agentic orchestration, retrieval-augmented generation, and reinforcement learning.
- It examines LLM-driven systems like ReFoRCE, Agentar-Scale-SQL, Snowflake's Arctic-Text2SQL-R1, and XiYan-SQL across benchmarks including Spider and BIRD.
- The paper also discusses open challenges such as domain generalization, efficiency, real-world deployment, and responsible AI practices for accessing sensitive databases.
Why it matters
This survey provides AI engineers and researchers with a structured taxonomy and comprehensive overview of state-of-the-art LLM-based Text-to-SQL methodologies, guiding better architectural choices for enterprise database access.