New CoTE-SQL Method Boosts Text-to-SQL Performance with Self-Enhanced Fine-Tuning.

A new research paper introduces CoTE-SQL, an innovative approach designed to enhance the accuracy and generalization of text-to-SQL systems. This method addresses the common challenge of balancing strong reasoning with robust generalization in large language models when converting natural language questions into executable SQL queries. CoTE-SQL incorporates three key innovations. First, it uses self-enhanced reasoning traces, which are distilled from LLMs without requiring human annotation. Second, it employs structured chain-of-thought (CoT) prompting, featuring modular decomposition and example retrieval to guide the model's reasoning process. Finally, it includes an error-aware revision mechanism that leverages SQL execution feedback to refine queries. Extensive experiments on the Spider and Bird benchmarks demonstrate that CoTE-SQL achieves new state-of-the-art performance among open-source LLM-based methods, particularly showing significant gains on complex queries. This highlights the effectiveness of combining self-enhancement, structured reasoning, and execution-time feedback for improved text-to-SQL generation.