CA-SQL: Complexity-Aware Inference-Time Reasoning Allocates Compute Budget for Text-to-SQL

arXiv·medium signal

Introduces a system that dynamically allocates inference-time compute based on SQL query complexity — simple queries get fast single-pass generation while hard queries trigger exploration and verification. Addresses the key bottleneck in production text-to-SQL: uniform compute wastes resources on easy queries and underspends on complex joins and subqueries. Achieves new SOTA on Spider and BIRD benchmarks.