# Agentar-Scale-SQL: Advancing Text-to-SQL through Orchestrated Test-Time Scaling

**Authors:** Pengfei Wang, Baolin Sun, Xuemei Dong, Yaxun Dai, Hongwei Yuan, Mengdie Chu, Yingqi Gao, Xiang Qi, Peng Zhang, Ying Yan  
**arXiv:** [2509.24403](https://arxiv.org/abs/2509.24403)  
**Date:** 2025-09-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2509.24403v1/

## Key points

- Agentar-Scale-SQL is a cutting-edge framework developed by a team of researchers aiming to enhance Text-to-SQL methods by addressing the gap between machine performance and human expertise on challenging benchmarks like BIRD.
- The framework introduces an Orchestrated Test-Time Scaling strategy that leverages scalable computation to improve performance through Internal Scaling, Sequential Scaling, and Parallel Scaling perspectives.
- Agentar-Scale-SQL offers a comprehensive approach to enhancing Text-to-SQL methods by integrating these key perspectives into its design.
- It is designed as a general-purpose framework that can be easily adapted to new databases and more powerful language models.
- Extensive experiments have shown that Agentar-Scale-SQL achieves state-of-the-art performance on the BIRD benchmark with an execution accuracy of 81.67% on the test set, securing the top rank on the official leaderboard.

## Summary

, , , , and 

Agentar-Scale-SQL, developed by Pengfei Wang, Baolin Sun, Xuemei Dong, Yaxun Dai, Hongwei Yuan, Mengdie Chu, Yingqi Gao, Xiang Qi, Peng Zhang, and Ying Yan,is a cutting-edge framework that aims to enhance Text-to-SQL methods by addressing the existing gap between machine performance and human expertise on challenging benchmarks like BIRD. The current state-of-the-art approaches in this field still fall short of achieving human-level accuracy due to the lack of an orchestrated strategy and neglect of the model's internal reasoning process during test-time scaling. To overcome these limitations,<nl> Agentar-Scale-SQL introduces a novel framework that leverages scalable computation to significantly improve performance. This framework implements an Orchestrated Test-Time Scaling strategy that combines three key perspectives: Internal Scaling through RL-enhanced Intrinsic Reasoning,<nl> Sequential Scaling via Iterative Refinement,<nl> and Parallel Scaling using Diverse Synthesis and Tournament Selection.<nl> By synergistically integrating these perspectives into its design,<nl> Agentar-Scale-SQL offers a comprehensive approach to enhancing Text-to-SQL methods.<nl> Furthermore,<nl> Agentar-Scale-SQL is designed as a general-purpose framework that can be easily adapted to new databases and more powerful language models.<nl> Extensive experiments have demonstrated the effectiveness of Agentar-Scale-SQL in achieving state-of-the-art performance on the BIRD benchmark.<nl> With an impressive execution accuracy of 81.67% on the test set and securing the top rank on the official leaderboard,<nl> Agentar-Scale-SQL showcases a promising path towards bridging the gap between machine performance and human-level expertise in Text-to-SQL tasks.