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

Text-to-SQL systems often struggle with complex database queries that require multi-step reasoning. Traditional approaches typically use either fixed, rigid pipelines or 'plan-then-execute' strategies that commit to a full workflow before seeing any intermediate results. SQLConductor introduces a step-wise orchestration framework that treats Text-to-SQL as a dynamic decision-making process, allowing the system to adapt its strategy based on the feedback and artifacts generated at each step of the query-solving process.

SQLConductor separates the orchestration of the reasoning process from the execution of specific subtasks. It defines an action space consisting of seven specialized modules—such as question decomposition, schema pruning, and SQL generation—that are driven by frozen large language models.

The core innovation is the Search-to-Policy learning framework:

1. Workflow Exploration: It uses Monte Carlo Tree Search (MCTS) to explore various action sequences, identifying successful workflows without requiring human-annotated examples.
2. Supervision Curation: It estimates the stability of these explored workflows to filter out unreliable or accidental successes, creating high-quality training data.
3. Policy Training: The policy model is trained using stability-weighted supervised fine-tuning, followed by curriculum reinforcement learning, which prioritizes challenging queries to ensure the policy remains robust.

By enabling step-wise orchestration, SQLConductor allows for more flexible and adaptive reasoning. Instead of being locked into a predefined sequence, the system can decide to revisit earlier steps, gather more evidence, or refine its query based on intermediate outputs. This flexibility leads to higher execution accuracy and better generalization to out-of-distribution datasets, all while using a compact policy model to coordinate larger, frozen backbones, making it more efficient than training monolithic models from scratch.