Sophrosyne: Curbing Over-Exploration in LLM-Powered Text2SQL Agents
A recent publication on arXiv (2605.30862v1) presents Sophrosyne, a novel data system environment aimed at enhancing the precision of Text2SQL agents utilizing large language models (LLMs). These agents convert natural language into SQL by navigating data systems via tool calls, but they encounter a dilemma between cost-effective exploration and producing accurate queries. The study classifies APIs into coarse-grained and fine-grained categories, revealing that fine-grained APIs lead to excessive exploration and the inclusion of irrelevant schema components, resulting in inaccurate outputs. Sophrosyne improves API responses with instructions that direct the agent's exploration, reducing over-exploration. Preliminary findings suggest enhanced performance. This paper was released as a cross-type publication.
Key facts
- arXiv paper 2605.30862v1 introduces Sophrosyne.
- Sophrosyne is a data system environment for Text2SQL agents.
- Text2SQL agents use LLMs to translate natural language into SQL.
- Agents explore data systems through tool calls before formulating queries.
- APIs are categorized as coarse-grained or fine-grained.
- Fine-grained APIs lead to over-exploration and inaccurate results.
- Sophrosyne augments API responses with directives to guide exploration.
- Initial results show improved accuracy.
Entities
Institutions
- arXiv