ExSPIN: Explicit Feedback-Based Self-Play Fine-Tuning for Text-to-SQL Parsing.

Abstract

Recently, self-play fine-tuning (SPIN) has garnered widespread attention as it enables large language models (LLMs) to iteratively enhance their capabilities through simulated interactions with themselves, transforming a weak LLM into a strong one. However, applying SPIN to fine-tune text-to-SQL models presents substantial challenges. Notably, existing frameworks lack clear signal feedback during the training process and fail to adequately capture the implicit schema-linking characteristics between natural language questions and databases. To address these issues, we propose a novel self-play fine-tuning method for text-to-SQL models, termed ExSPIN, which incorporates explicit feedback. Specifically, during fine-tuning, the SQL query execution results predicted by the LLM are fed back into the model's parameter update process. This feedback allows both the main player and the opponent to more accurately distinguish between negative and positive samples, thereby improving the fine-tuning outcomes. Additionally, we employ in-context learning techniques to provide explicit schema hints, enabling the LLM to better understand the schema-linking between the database and natural language queries during the self-play process. Evaluations on two real-world datasets show that our method significantly outperforms the state-of-the-art approaches.