Residual Skill Optimization for Text-to-SQL Ensembles

In three linesDivSkill-SQL optimizes Text-to-SQL ensembles through residual skill learning on failed examples, without fine-tuning. On Spider2-Lite, gains of +11.1 pts (Snowflake) and +8.3 pts (BigQuery) vs baseline. Skills transfer across SQL dialects without retraining.

Summary generated by Claude — human-verified