EvoSQL Improves Text-to-SQL with Co-Evolution and Memory

While large language models have greatly improved Text-to-SQL capabilities, generating complex database queries still demands advanced reasoning, including multi-step decomposition and execution-based error diagnosis. To address these challenges, EvoSQL, a novel co-evolution framework, has been developed. EvoSQL conceptualizes SQL synthesis as an iterative dialogue between a generator, which proposes SQL candidates, and a critic, which evaluates them.The framework maintains a dynamic, contextualized memory of candidate SQL queries. These candidates are rigorously verified using both actual execution signals and LLM-based critique. The memory is then updated through a utility-guided aggregation process, ensuring that the system continuously learns and refines its approach. To further enhance the underlying generator-critic pair, EvoSQL incorporates a Self-Distillation Policy Optimization (SDPO) fine-tuning stage. This stage injects execution-aware supervision into modern coding LLM backbones, making them more adept at generating correct SQL.Extensive experiments on challenging benchmarks like Spider and BIRD demonstrate that EvoSQL consistently outperforms existing open-source models. The gains are particularly notable on BIRD-Dev, with improvements ranging from +1.37% to +9.19% for various models. The SDPO initialization further boosts performance. These results underscore the effectiveness of memory-grounded co-evolution as a robust strategy for developing more reliable and generalizable Text-to-SQL systems.