ReToolSQL improves text-to-SQL accuracy by combining supervised fine-tuning on rejection-sampled reasoning traces with multi-turn agentic reinforcement fine-tuning. This two-stage approach enables models to iteratively refine queries through tool-use rather than relying on single-turn generation.
HOW THIS AFFECTS YOU
●
builderYou can improve single-pass SQL accuracy and error recovery by implementing agentic feedback loops.
●
researcherThe framework demonstrates how SFT and RFT act on complementary axes to improve pass@k and single-turn success.