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

2026-06-22

AI Summary

Existing Text-to-SQL approaches struggle to dynamically adapt to intermediate results and feedback in complex databases, and their fixed pipeline strategies exhibit limited generalization. This work proposes SQLConductor, which models subtasks as optional actions and employs a lightweight policy model to dynamically orchestrate query construction based on intermediate outputs. The approach introduces a novel Search-to-Policy Learning framework that integrates Monte Carlo Tree Search, stability-weighted supervised fine-tuning, and curriculum reinforcement learning to generate high-quality training signals and enhance policy generalization. Evaluated on BIRD-Dev, SQLConductor achieves a 73.2% execution accuracy, substantially outperforming baselines that rely solely on scaling model size, and demonstrates strong out-of-distribution generalization capabilities.

Abstract

Text-to-SQL enables users to access relational databases via natural language, but real-world settings remain challenging due to coordinated reasoning over complex database environments. Existing systems often use multi-stage pipelines or reasoning models specialized for individual stages. However, fixed pipelines rely on predefined stage orders, limiting their adaptivity to query demands and intermediate evidence. Recent orchestration-based methods provide flexibility by composing specialized modules for each query, but typical plan-then-execute approaches still commit to a complete workflow before execution and cannot adapt to intermediate artifacts and feedback. In this paper, we propose SQLConductor, a step-wise orchestration learning framework for Text-to-SQL. SQLConductor formulates Text-to-SQL subtasks as specialized actions for workflow composition and trains a policy model to select the next action based on intermediate artifacts and feedback. To learn this policy, SQLConductor introduces Search-to-Policy Learning, which uses Monte Carlo Tree Search to explore candidate workflows and stability estimation to identify robust supervision. The policy model is trained with Stability-weighted Supervised Fine-tuning to prioritize high-quality orchestration patterns and further enhanced through Curriculum Reinforcement Learning. This transforms offline workflow search into a deployable policy for step-wise orchestration at inference time. Experiments on BIRD-Dev and out-of-distribution datasets show that SQLConductor achieves superior execution accuracy and strong generalization, reaching 73.2% EX on BIRD-Dev with a compact orchestration policy coordinating frozen larger action models, outperforming prior methods that directly train comparable or larger Text-to-SQL backbones. Further analyses show that the learned policy adapts orchestration to diverse query demands.

Problem

• Text-to-SQL
• workflow orchestration
• adaptive reasoning
• complex database environments
• intermediate feedback

Innovation

• step-wise orchestration
• Search-to-Policy Learning
• Monte Carlo Tree Search
• Stability-weighted Supervised Fine-tuning
• Curriculum Reinforcement Learning