SQLConductor: Search-to-Policy Learning for Step-wise Text-to-SQL Orchestration

SQLConductor introduces a novel step-wise orchestration learning framework for Text-to-SQL, utilizing a policy model to dynamically select actions based on intermediate artifacts and feedback. It employs Search-to-Policy Learning via Monte Carlo Tree Search and Stability-weighted Supervised Fine-tuning, achieving a 73.2% execution accuracy on the BIRD-Dev dataset while coordinating larger frozen action models. This approach enhances adaptability in real-world database interactions, offering significant improvements over traditional fixed pipelines and prior methods, making it valuable for practitioners developing flexible AI-driven database query systems.