Training a competitive Text-to-SQL agent normally relies on human‑annotated NL/SQL pairs, which are costly and domain‑specific. This work shows that a strong solver can be learned without any labeled pairs.
We introduce SQL‑Zero, a proposer‑solver self‑play framework. A challenger and a solver start from the same base LLM, and the only ground truth comes from executing queries against the target database. The challenger generates SQL calibrated to the solver’s current difficulty (aiming for “hard but solvable”), and both roles are updated alternately with GRPO while a template‑level repetition penalty on the challenger prevents diversity collapse.
Training on BIRD databases without labels, the 3B model improves over the zero‑shot base by 6.6 points on BIRD dev, and the 7B model by 7.3 points. Compared with a matched control trained on human BIRD gold using the same recipe, SQL‑Zero also achieves higher scores, although a paired test does not resolve the margin statistically.
Scale matters for transfer: at 3B every iteration outperforms the base on unseen Spider databases and under lexical perturbation (Spider‑Syn), and degrades less than the BIRD‑gold control. At 7B only the first iteration retains transfer ability.
By leveraging self‑play and execution feedback, SQL‑Zero enables zero‑annotation Text‑to‑SQL learning, opening a new path for cross‑database generalization.
Review