DivSkill-SQL: Residual Skill Optimization for Text-to-SQL Ensembles
DivSkill-SQL is a groundbreaking framework aimed at improving Text-to-SQL performance by leveraging residual skills that existing ensembles may lack. It achieves noteworthy enhancements in Pass@K metrics without the necessity for fine-tuning, recording upticks of 11.1 points on Snowflake and 8.3 points on BigQuery using the Spider2-Lite dataset, surpassing the top baseline results. Additionally, DivSkill-SQL demonstrates consistent advancements with both Opus-4.6 and GPT-5.4 models. A significant advantage of this framework is its capability to transfer skills across different SQL dialects effortlessly, eliminating the need for additional retraining.
Key facts
- DivSkill-SQL is a residual skill optimization framework.
- It builds complementary agentic Text-to-SQL ensembles without model fine-tuning.
- Each new skill is optimized on examples the current ensemble fails on.
- It targets marginal contribution to Pass@K.
- On Spider2-Lite, it improves selected accuracy by up to +11.1 points on Snowflake.
- On Spider2-Lite, it improves selected accuracy by up to +8.3 points on BigQuery.
- Consistent gains across Opus-4.6 and GPT-5.4 base models.
- Skills optimized on a single dialect transfer without retraining.
Entities
—