MEM-SQL: Memory-Empowered Text-to-SQL
Reusing verified SQL lifted mean execution accuracy from 37.1% to 50.0%, while cutting input tokens by 76.6% and latency by 68.8%.
The results span six language models on HARE, a new benchmark with 1,600 requests built from 823 verified queries across four datasets. MEM-SQL stores whole-query templates, reusable SQL blocks and domain-to-schema mappings, consulting the schema when retrieved evidence falls short. Combining several past queries remains a weakness: accuracy fell from 56.7% to 29.7% on requests requiring that reuse, especially when components needed a new query structure or shared join.
HF Daily Papers' note
The results span six language models on HARE, a new benchmark with 1,600 requests built from 823 verified queries across four datasets. MEM-SQL stores whole-query templates, reusable SQL blocks and domain-to-schema mappings, consulting the schema when retrieved evidence falls short. Combining several past queries remains a weakness: accuracy fell from 56.7% to 29.7% on requests requiring that reuse, especially when components needed a new query structure or shared join.
HF Daily Papers' note
score 4