How Does Entropy Influence Modern Text-to-SQL Systems?

In the field of text-to-SQL candidate generation, a critical challenge remains in quantifying and assessing the confidence in the generated SQL queries. Existing approaches often rely on large language models (LLMs) that function as opaque processing units, producing outputs for every input without a mechanism to measure their confidence. Current uncertainty quantification techniques for LLMs do not incorporate domain-specific information. In this study, we introduce the concept of query entropy for Text-to-SQL candidate confidence estimation and integrate it into existing popular self-correction pipelines to guide generations and prevent resource overuse by including a novel clustering technique for generated SQL candidates based on entropy. We further study the treatment of different candidate generation techniques under this paradigm.