R³-SQL: A New Framework for Text-to-SQL with Ranking and Resampling
A research paper introduces R³-SQL, a Text-to-SQL framework that addresses two key limitations in existing systems: inconsistent scoring of functionally equivalent SQL queries and the inability to recover when the correct SQL is missing from the candidate pool. R³-SQL groups candidates by execution result and ranks groups for consistency, combining pairwise preference across groups with pointwise utility from the best group rank and size. It also introduces agentic resampling to judge the generated candidate pool and selectively resample when the correct SQL is likely absent. The paper is available on arXiv.
Key facts
- R³-SQL stands for Ranking Reward and Resampling for Text-to-SQL.
- The paper is published on arXiv with ID 2604.25325.
- Modern Text-to-SQL systems generate multiple candidate SQL queries and rank them.
- Existing methods score functionally equivalent SQL queries inconsistently.
- Ranking cannot recover when the correct SQL is absent from the candidate pool.
- R³-SQL groups candidates by execution result for consistent ranking.
- It combines pairwise preference and pointwise utility for scoring.
- Agentic resampling selectively resamples when correct SQL is likely absent.
Entities
Institutions
- arXiv