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

SQLConductor is a new AI framework that improves Text-to-SQL systems—tools that convert natural language queries into database commands—by using adaptive, step-wise orchestration rather than fixed pipelines. The system achieves 73.2% execution accuracy on complex database queries while using smaller, frozen models, suggesting significant efficiency gains for database accessibility applications.

SQLConductor addresses a fundamental limitation in current Text-to-SQL systems: their inability to adapt to real-world database complexity. Traditional approaches rely on rigid, predefined execution stages that cannot adjust based on intermediate results or unexpected query structures. This paper presents a learning-based orchestration framework that treats SQL generation as a dynamic decision-making problem, selecting the next action based on current context rather than committing to a complete workflow upfront.

The technical innovation centers on Search-to-Policy Learning, which combines Monte Carlo Tree Search exploration with stability-weighted training. This approach generates high-quality training examples by exploring diverse orchestration patterns and identifying robust strategies that generalize well. The resulting policy model operates as a lightweight conductor that coordinates specialized frozen modules, reducing computational overhead while improving accuracy.

For the database and enterprise software markets, this represents meaningful progress toward democratizing data access. Organizations increasingly expect natural language interfaces to their data warehouses, yet current systems struggle with schema complexity and semantic ambiguity. SQLConductor's 73.2% execution accuracy on BIRD-Dev, combined with strong out-of-distribution performance, suggests practical viability for real deployments where query diversity is inherently high.

The efficiency gains—achieving superior results with smaller coordinated models rather than single large models—have implications for AI deployment economics. This orchestration pattern could influence how enterprises architect their AI infrastructure, prioritizing flexible composition over monolithic scaling. Future development should focus on handling increasingly complex database environments and multi-step reasoning scenarios.

• →SQLConductor uses adaptive step-wise orchestration to improve Text-to-SQL accuracy, reaching 73.2% execution accuracy through flexible action selection rather than fixed pipelines.
• →Search-to-Policy Learning combines Monte Carlo Tree Search with stability estimation to train efficient policy models that coordinate frozen specialized modules.
• →The framework demonstrates strong generalization on out-of-distribution datasets, addressing a critical real-world challenge for natural language database access.
• →Orchestration-based approaches achieve superior performance with smaller composite models compared to larger monolithic Text-to-SQL backbones.
• →Adaptive workflow selection enables the system to adjust to diverse query demands and intermediate evidence, improving robustness across different database scenarios.