# The Death of Schema Linking? Text-to-SQL in the Age of Well-Reasoned Language Models

**Authors:** Karime Maamari, Fadhil Abubaker, Daniel Jaroslawicz, Amine Mhedhbi  
**arXiv:** [2408.07702](https://arxiv.org/abs/2408.07702)  
**Date:** 2024-08-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.07702v1/

## Key points

- Large language models (LLMs) have significantly advanced the field of Text-to-SQL by generating SQL queries from natural language inquiries.
- LLM-based approaches in Text-to-SQL typically involve a multi-stage pipeline, starting with retrieval and ending with correction.
- Schema linking is a critical aspect in accurate query generation, providing context by selecting relevant elements of the database schema.
- Recent advancements in LLM reasoning have questioned the necessity of traditional schema linking, as newer models can identify relevant schema elements without explicit linking, reducing noise while preserving signal.
- As model reasoning improves, the benefits of noise reduction from traditional schema linking become less significant, challenging conventional wisdom.

## Summary

, , , , 

In the field of Text-to-SQL, the generation of SQL queries from natural language inquiries has been significantly advanced by large language models (LLMs). These LLM-based approaches typically follow a multi-stage pipeline, starting with retrieval and ending with correction. One critical aspect is schema linking, which selects relevant elements of the database schema to provide context for accurate query generation. However, recent advancements in LLM reasoning have led to reevaluating traditional schema linking's necessity. Empirical findings suggest that newer models can identify relevant schema elements without explicit linking, reducing noise while preserving signal. As model reasoning improves, the benefits of noise reduction become less significant, challenging conventional wisdom around schema linking. To address this shift, we propose alternative methods that improve Text-to-SQL accuracy without compromising essential schema information. Our approach leverages empirical insights and currently ranks first in execution accuracy at 71.83% on the BIRD benchmark. In summary, as LLMs continue to evolve and improve their reasoning abilities, there may be opportunities to streamline Text-to-SQL pipelines by bypassing traditional schema linking methods in favor of more efficient and accurate approaches.