Mixture-of-Agents based Text-to-SQL Study on TURSpider Dataset

This study focuses on experiments conducted on the TURSpider dataset, developed for the Turkish Text-to-SQL task. TURSpider is a large-scale Turkish dataset containing SQL queries of varying difficulty levels and serves as a valuable resource for research in this field. The study investigates the effectiveness of the feedback-driven Mixture-of-Agents (MoAF) approach on this task. In the MoAF structure, multiple large language models (LLMs) collaborate to improve SQL generation performance. In this setup, agent collaboration enables models to learn from each other and correct errors through feedback mechanisms. According to the experimental results, the MoAF approach achieved an execution accuracy of 60.63%, achieving the highest performance reported on the TURSpider dataset.