Reward-SQL: Boosting Text-to-SQL via Stepwise Execution-Aware Reasoning and Process-Supervised Rewards

Researchers have proposed Reward-SQL, a framework that enhances Text-to-SQL performance by incorporating stepwise execution-aware reasoning and process-supervised rewards, addressing limitations in existing reinforcement learning-based approaches^[1].

Text-to-SQL aims to translate natural language questions into executable SQL queries over structured databases. Recent advances in large language models have shown promise in this task, but existing approaches often struggle to balance strong reasoning capabilities and robust generalization^[2]. Reward-SQL is a unified approach with three stages: model initialization, process reward design, and process-supervised RL and inference. The framework introduces a process reward model that combines execution-aware trajectory scoring with entropy-based step weighting, providing dense and interpretable supervision across reasoning steps^[1]. Experiments show that Reward-SQL significantly outperforms baselines with comparable model sizes and exhibits strong cross-domain generalization. Meanwhile, CoTE-SQL, another framework, achieved state-of-the-art performance on the Bird benchmark with 53.39% EX and 59.02 VES, and strong results on the Spider benchmark with 79.60% EX and 77.19 VES^[2].