TLDR: HES-SQL is a hybrid training framework for Text-to-SQL generation that combines supervised fine-tuning with reinforcement learning. It introduces skeleton-completeness scoring for structural alignment, a query-latency-aware reward system for computational efficiency, and self-distillation to preserve reasoning capabilities. This approach enables models to generate highly accurate and efficient SQL queries, achieving competitive performance on benchmarks like BIRD and KaggleDBQA with significant efficiency gains, and has been deployed in real-world applications.
Interacting with databases often requires knowledge of complex SQL syntax, which can be a barrier for many users. The field of Text-to-SQL aims to bridge this gap by allowing users to query databases using natural language. While significant progress has been made, especially with the advent of large language models (LLMs), developing systems that are both highly accurate and computationally efficient remains a considerable challenge.
A new research paper introduces HES-SQL, a novel hybrid training framework designed to address these dual challenges. Developed by Suming Qiu, Jing Li, Zhicheng Zhou, Junjie Huang, Linyuan Qiu, and Zhijie Sun, HES-SQL integrates advanced supervised fine-tuning with a sophisticated reinforcement learning approach to generate SQL queries that are not only semantically correct but also execute efficiently.
Understanding HES-SQL’s Core Innovations
HES-SQL stands out with three key innovations that contribute to its superior performance:
First, a Skeleton-Completeness Scoring mechanism evaluates how well a generated SQL query’s structure aligns with the ideal SQL pattern. It acts as a filter, ensuring that only syntactically and logically sound queries are considered, thereby improving the structural integrity of the generated SQL.
Second, a Query-Latency-Aware Reward System recognizes that a correct query is not enough if it runs slowly. HES-SQL incorporates a reward system that explicitly encourages the model to produce computationally efficient SQL queries. This means the system learns to generate queries that execute faster on database management systems like MySQL and SQLite.
Third, Self-Distillation for Thinking-Mode Completion is used to prevent the model’s reasoning capabilities from degrading during the training process. This helps the model maintain its ability to switch between “fast” (non-reasoning) and “slow” (reasoning with chain-of-thought) thinking modes, ensuring robust and adaptable query generation.
How HES-SQL Works
The framework combines supervised fine-tuning (SFT) with Group Relative Policy Optimization (GRPO), a reinforcement learning strategy. The SFT phase initializes the model with high-quality reasoning examples, including step-by-step thought processes. This “cold-start” SFT provides a strong foundation. Subsequently, GRPO fine-tunes the model using a hierarchical reward system. This system evaluates queries based on several factors: format validity, successful execution, correct schema linking (matching natural language conditions to database elements), and execution efficiency (how fast the query runs compared to a gold standard).
By using this multi-stage reward system, HES-SQL guides the model to learn not just what the correct SQL is, but also how to generate it efficiently and robustly, even for complex queries involving multiple joins or nested subqueries.
Also Read:
- CLARITY: Enhancing LLM Reasoning Quality Through Consistency-Aware Reinforcement Learning
- Deepening Knowledge Integration: How Semantic-Condition Tuning Enhances LLMs for Knowledge Graph Completion
Impressive Performance and Real-World Impact
Experimental evaluations on standard Text-to-SQL benchmarks—BIRD, Spider, and KaggleDBQA—demonstrate HES-SQL’s competitive performance. It achieved execution accuracies of 79.14% on BIRD and 54.9% on KaggleDBQA, consistently outperforming strong supervised baselines. Crucially, it also delivered efficiency gains ranging from 11% to 20% (reduced latency) compared to these baselines, meaning faster query execution times.
These results establish a new standard for Text-to-SQL systems, effectively balancing semantic accuracy with computational efficiency. The methodology has significant implications for developing robust natural language interfaces to databases, making data access more accessible and efficient for non-technical users. HES-SQL has already been deployed as part of the CEM Copilot application at Saudi STA sites, showcasing its practical utility in an industrial setting. For more technical details, you can refer to the full research paper: HES-SQL: Hybrid Reasoning for Efficient Text-to-SQL with Structural Skeleton Guidance.
