ReToolSQL introduces a two-stage training framework that combines supervised fine-tuning on rejection-sampled traces with agentic reinforcement fine-tuning over multi-turn tool-use. This approach addresses the limitation of single-turn SQL generation by allowing models to iteratively refine queries based on execution feedback. The method enables smaller models to match or exceed larger systems by leveraging complementary learning axes.
- Multi-turn agentic RL allows iterative error recovery in SQL generation.
- Smaller models can outperform larger systems using this two-stage framework.
- Rejection-sampled reasoning traces provide a strong supervised warm-start.
- Tool-use trajectories enable robust execution feedback loops.
