SF-SQL: A Stage-Free Text-to-SQL Framework Enhanced by Reinforcement Learning

Abstract

Text-to-SQL, which translates natural language into SQL queries, has attracted significant attention in the database community. Recent advances leverage large language models (LLMs) to generate syntactically and semantically correct SQL. However, these approaches typically rely on prompt engineering and rigid multi-stage pipelines, making them sensitive to prompt variations and limiting their generalization across SQL dialects. To address these limitations, we propose SF-SQL, a Stage-Free Text-to-SQL framework enhanced by reinforcement learning. Instead of following a fixed pipeline, SF-SQL formulates SQL generation as a dynamic decision-making process, enabling the LLM to reason about the next action based on the context and execution feedback. To facilitate effective training, we develop a trajectory synthesis strategy that efficiently generates diverse and high-quality reasoning-action sequences. Moreover, SF-SQL proposes a reward scheme that integrates both final and intermediate query-aware signals without incurring additional inference overhead. Our experiments show that SF-SQL achieves 88.73% execution accuracy on Spider and 69.23% on BIRD, demonstrating strong generalization across five types of database dialects.

Keywords

Text-to-SQL / Large Language Model / Reinforcement Learning / Trajectory Synthesis

RIGHTS & PERMISSIONS

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.