Cost‑effective integration of LLMs and PLMs for NL2SQL

Determine effective strategies to integrate and balance Large Language Models (LLMs) and Pre‑trained Language Models (PLMs) within modular or multi‑agent NL2SQL systems so as to optimize SQL generation performance while minimizing token consumption, inference time, and overall resource cost.

Background

The survey highlights that LLM-based NL2SQL methods are promising but often incur high token usage and inference time, while PLM-based methods tend to better handle complex SQL and schema understanding. Combining these complementary strengths is proposed as a promising direction via modular or multi-agent designs.

However, the paper explicitly notes that achieving an effective integration and balance between LLMs and PLMs to simultaneously optimize performance and resource efficiency remains unresolved, motivating a concrete open problem in system design and optimization.