Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AGRO-SQL: Agentic Group-Relative Optimization with High-Fidelity Data Synthesis

Cehua Yang, Dongyu Xiao, Junming Lin, Yuyang Song, Hanxu Yan, Shawn Guo, Wei Zhang, Jian Yang, Mingjie Tang, Bryan Dai

TL;DR

AGRO-SQL tackles data scarcity and reasoning limitations in Text-to-SQL by integrating a high-fidelity, RL-ready data factory with a model-centric agentic RL framework. It introduces Generation-as-Verification and a two-stage agentic RL pipeline featuring Diversity-Aware Cold Start and Group Relative Policy Optimization (GRPO) to stabilize learning and improve semantic-logic accuracy. Inference is enhanced via a three-stage context augmentation and Multi-turn Refinement with execution feedback. On the BIRD and Spider benchmarks, AGRO-SQL achieves state-of-the-art performance among single-model approaches, exemplified by an EX of 70.66% (72.10% with self-consistency) on BIRD, demonstrating strong practical impact for open-source Text-to-SQL.

Abstract

The advancement of Text-to-SQL systems is currently hindered by the scarcity of high-quality training data and the limited reasoning capabilities of models in complex scenarios. In this paper, we propose a holistic framework that addresses these issues through a dual-centric approach. From a Data-Centric perspective, we construct an iterative data factory that synthesizes RL-ready data characterized by high correctness and precise semantic-logic alignment, ensured by strict verification. From a Model-Centric perspective, we introduce a novel Agentic Reinforcement Learning framework. This framework employs a Diversity-Aware Cold Start stage to initialize a robust policy, followed by Group Relative Policy Optimization (GRPO) to refine the agent's reasoning via environmental feedback. Extensive experiments on BIRD and Spider benchmarks demonstrate that our synergistic approach achieves state-of-the-art performance among single-model methods.

AGRO-SQL: Agentic Group-Relative Optimization with High-Fidelity Data Synthesis

TL;DR

AGRO-SQL tackles data scarcity and reasoning limitations in Text-to-SQL by integrating a high-fidelity, RL-ready data factory with a model-centric agentic RL framework. It introduces Generation-as-Verification and a two-stage agentic RL pipeline featuring Diversity-Aware Cold Start and Group Relative Policy Optimization (GRPO) to stabilize learning and improve semantic-logic accuracy. Inference is enhanced via a three-stage context augmentation and Multi-turn Refinement with execution feedback. On the BIRD and Spider benchmarks, AGRO-SQL achieves state-of-the-art performance among single-model approaches, exemplified by an EX of 70.66% (72.10% with self-consistency) on BIRD, demonstrating strong practical impact for open-source Text-to-SQL.

Abstract

The advancement of Text-to-SQL systems is currently hindered by the scarcity of high-quality training data and the limited reasoning capabilities of models in complex scenarios. In this paper, we propose a holistic framework that addresses these issues through a dual-centric approach. From a Data-Centric perspective, we construct an iterative data factory that synthesizes RL-ready data characterized by high correctness and precise semantic-logic alignment, ensured by strict verification. From a Model-Centric perspective, we introduce a novel Agentic Reinforcement Learning framework. This framework employs a Diversity-Aware Cold Start stage to initialize a robust policy, followed by Group Relative Policy Optimization (GRPO) to refine the agent's reasoning via environmental feedback. Extensive experiments on BIRD and Spider benchmarks demonstrate that our synergistic approach achieves state-of-the-art performance among single-model methods.
Paper Structure (18 sections, 1 figure, 1 table)

This paper contains 18 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Data Synthesis in Text-to-SQL
  4. Reinforcement Learning in Text-to-SQL
  5. Inference-Time Strategies for Text-to-SQL
  6. Methodology
  7. RL-Ready Iterative Data Pipeline
  8. Agentic RL Training Framework
  9. Inference
  10. Experiments
  11. Datasets
  12. Metrics
  13. Baselines
  14. Implementation Details
  15. Main Results
  16. ...and 3 more sections

Figures (1)

  • Figure 1: The overall pipeline of our framework. Given a natural language question and database schema, the policy model generates SQL candidates. Our core Advantage Shaping Module then computes reshaped token-level advantages, which are used to update the policy model via the GRPO algorithm.