CA-SQL: Complexity-Aware Text-to-SQL with Dynamic Compute Budget
A new Text-to-SQL pipeline called CA-SQL has been developed by researchers, which adjusts its exploration breadth according to the complexity of the task. This system estimates difficulty to determine compute budgets, employs custom prompt seeding influenced by evolutionary search, and utilizes an innovative voting technique to identify the optimal candidate query. On the Bird-Bench (BIRD) benchmark, CA-SQL has reached a leading score of 51.72% on its most difficult subset, effectively tackling the limitations of solution space exploration seen in existing inference-time learning approaches.
Key facts
- CA-SQL is a Text-to-SQL pipeline.
- It uses estimated task difficulty to scale exploration breadth.
- Custom prompt seeding is based on evolutionary search principles.
- A novel voting method selects the best candidate solution.
- Achieves 51.72% on Bird-Bench challenging subset.
- Addresses inadequate solution space exploration.
- Published on arXiv as 2605.08057.
- Announcement type: cross.
Entities
Institutions
- arXiv