SQL-Zero: Self-Evolving Text-to-SQL

The researchers proposed SQL-Zero, a method for training competitive Text-to-SQL agents without the need for manual annotation of data. This method uses a proposer-solver game mechanism, with the challenger and solver starting from the same base large language model, with database execution results as the only true value. The challenger generates calibrated SQL pairs for the current difficulty level, and both sides alternately use GRPO updates. A template-level repetition penalty is applied to the challenger to prevent diversity collapse. After unlabeled training on the BIRD database, the 3B and 7B models improved by 6.6 points and 7.3 points respectively on the development set compared to zero-sample baselines, and their performance was better than that of control models trained with manually annotated data under the same conditions. Transferability depends on scale: the 3B model outperformed the baseline in each iteration on the unseen Spider database and vocabulary perturbation task (Spider-Syn) and degraded less compared to the control models; while the 7B model only maintained transfer effects in the first iteration.