Reproduce “Memo-SQL: Structured Decomposition and Experience-Driven Self-Correction for Training-Free NL2SQL” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zerui Yang, Weichuan Wang, Yanwei Xu, Linqi Song, Yudai Matsuda, Wei Han, Bo Bai
Memo-SQL introduces a training-free framework for Natural Language to SQL (NL2SQL) that integrates structured decomposition and experience-driven self-correction. The system achieves 68.5% execution accuracy on the BIRD dev-new set, establishing a new state of the art for open, zero-fine-tuning methods, and significantly reduces computational costs compared to leading approaches.
Key insights
- Dynamically leveraging historical error-fix pairs through memory-augmented in-context learning enables adaptive self-correction, improving robustness against diverse error types.
- Employing multiple structured decomposition strategies (entity-wise, hierarchical, atomic operational) promotes diversity in reasoning paths, yielding less correlated SQL candidates for more effective ensemble selection.
- An open, training-free framework can achieve state-of-the-art NL2SQL accuracy and superior efficiency by strategically combining experience-driven refinement and diverse query decomposition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.