Residual Skill Optimization for Text-to-SQL Ensembles

DivSkill-SQL builds complementary Text-to-SQL agent ensembles by optimizing each new skill on examples that earlier skills fail to solve. This targets the additional coverage each skill contributes to Pass@K without fine-tuning the underlying language model. The paper evaluates the approach on Spider2-Lite and studies transfer across SQL dialects and to BIRD-Critic, showing improvements in selected-answer accuracy and fewer schema and function hallucinations.