Arctic-Text2SQL-R1: Simple Rewards, Strong Reasoning in Text-to-SQL

2025-05-22

AI Summary

Text-to-SQL systems still struggle to generate correct and executable SQL queries for complex natural language questions. To address this, we propose a lightweight reinforcement learning framework that eliminates fragile intermediate supervision and intricate reward engineering, introducing instead a minimalist binary reward signal solely based on SQL execution outcomes. Our method integrates Proximal Policy Optimization (PPO), execution-feedback-driven supervised fine-tuning, curriculum learning, and rigorous data cleaning. Additionally, we incorporate value-retrieval augmentation and majority-voting-based inference ensembling. Evaluated on six mainstream Text2SQL benchmarks, our approach achieves state-of-the-art execution accuracy, ranking first on the BIRD leaderboard. Notably, our 7B-parameter model significantly outperforms prior systems with 70B parameters, while offering superior training stability and deployment efficiency.

Technology Category

• Natural Language Processing: Question Answering
• Search and Optimization: Learning to Search
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Translating natural language into SQL (Test2SQL) is a longstanding challenge at the intersection of natural language understanding and structured data access. While large language models (LLMs) have significantly improved fluency in SQL generation, producing correct and executable SQL--particularly for complex queries--remains a bottleneck. We present Arctic-Text2SQL-R1, a reinforcement learning (RL) framework and model family designed to generate accurate, executable SQL using a lightweight reward signal based solely on execution correctness. Our approach avoids brittle intermediate supervision and complex reward shaping, promoting stable training and alignment with the end task. Combined with carefully curated data, strong supervised initialization, and effective training practices, Arctic-Text2SQL-R1 achieves state-of-the-art execution accuracy across six diverse Test2SQL benchmarks, including the top position on the BIRD leaderboard. Notably, our 7B model outperforms prior 70B-class systems, highlighting the framework's scalability and efficiency. We further demonstrate inference-time robustness through simple extensions like value retrieval and majority voting. Extensive experiments and ablation studies offer both positive and negative insights, providing practical guidance for future Test2SQL research.

Problem

• Improving accuracy of natural language to SQL translation
• Generating executable SQL for complex queries efficiently
• Avoiding brittle supervision with lightweight reinforcement learning

Innovation

• Uses reinforcement learning for SQL generation
• Lightweight reward based on execution correctness
• Strong supervised initialization and training practices