MEM-SQL improves how Text-to-SQL systems handle complex database questions by reusing historical SQL templates, SQL blocks, and schema mappings. The paper also introduces the HARE benchmark, which contains 1,600 requests. Across six language models, average execution accuracy increased from 37.1% to 50.0%, average input tokens decreased by 76.6%, and average latency fell by 68.8%. When a request requires combining multiple validated queries or placing reusable components into a new query structure, accuracy drops from 56.7% to 29.7%, indicating that complex historical reuse remains a bottleneck.
Source:Hugging Face 社区论文精选 · Read Original · 前沿热点
