EvoSQL: Memory-Augmented Critic-Generator Co-Evolution for Text-to-SQL

Recent advances in NLP, especially large language models like GPT-4 and Codex, have propelled semantic parsing and Text-to-SQL tasks. Early methods relied on rule-based systems, but deep learning approaches, particularly Transformer-based models, have dominated recent progress. Schema-aware models (Li et al., 2023a) incorporate database structure understanding, while content-aware retrieval enhances semantic alignment. Despite these developments, complex queries involving multi-hop reasoning, nested subqueries, and error correction remain challenging. Existing solutions primarily focus on single-pass decoding, which limits their ability to handle multi-step reasoning, leading to high error rates and poor generalization. Multi-turn approaches, including iterative refinement and multi-agent systems, have been explored but often lack structured memory and effective error diagnosis, constraining their effectiveness in real-world scenarios. Therefore, there is a pressing need for frameworks that can dynamically explore multiple hypotheses, diagnose errors, and adaptively refine SQL queries across multiple reasoning steps.

Current Text-to-SQL systems face significant bottlenecks in reasoning over complex schemas and nested queries. Single-pass models often produce incomplete or incorrect SQL, especially when multi-step reasoning or error diagnosis is required. Existing methods lack effective mechanisms for maintaining candidate diversity, structured error diagnosis, and iterative refinement, resulting in limited robustness and generalization. The core challenge is to develop a system that can perform multi-round candidate generation, verification, and correction, leveraging execution feedback and structured critique to improve accuracy in complex scenarios. Addressing this gap is crucial for deploying reliable natural language interfaces to databases in real-world applications, where queries are often intricate and error-prone.

This paper introduces several key innovations: 1) Memory-augmented co-evolution framework that couples a generator and critic for multi-round candidate refinement; 2) Dual-stage verification combining deterministic execution signals with LLM-based critique, enabling precise error diagnosis; 3) Utility-guided candidate selection and aggregation leveraging episodic memory, balancing exploration and exploitation; 4) Integration of Self-Distillation Policy Optimization (SDPO) to enhance execution-aware supervision during offline fine-tuning. These innovations collectively enable the system to perform iterative, structured reasoning, improving robustness and accuracy over traditional single-pass models. The approach effectively addresses multi-step reasoning, error diagnosis, and correction, setting a new standard for complex Text-to-SQL tasks.

• �� Initial candidate generation: The generator produces multiple SQL hypotheses based on input question, schema, and context.
• �� Candidate verification: Each candidate undergoes execution, capturing signals like success, errors, and runtime.
• �� Critique and diagnosis: The critic analyzes the question, candidate SQL, and execution results, providing structured scores and correction suggestions.
• �� Candidate filtering: Using a utility function that considers execution confidence and consistency bonuses, high-utility candidates are selected.
• �� Memory update: Verified candidates and summaries are stored in episodic memory for future rounds.
• �� Multi-round evolution: The process repeats, with retrieval of high-utility candidates for further refinement, until early stopping criteria are met.
• �� Final selection: The candidate with the highest utility score is chosen as the output.
• �� Offline fine-tuning: SDPO leverages execution feedback to distill supervision signals into the model, improving subsequent candidate generation.

The evaluation employs Spider and BIRD benchmarks, measuring execution accuracy (EX) with a 30-second timeout. Baselines include Maj@16, with models like Coder-3B, Coder-7B, Qwen3-4B, and SQL-R1. The experiments compare single-pass, multi-round co-evolution, and SDPO-enhanced variants. Hyperparameters include 16 candidates per round, up to 3 rounds, and specific utility weighting. Ablation studies analyze the impact of candidate memory, verification stages, and critique mechanisms. Results show consistent improvements across datasets, with the largest gains on challenging queries, validating the framework’s effectiveness in complex reasoning scenarios.

EvoSQL outperforms baseline models with improvements up to 9.19% on BIRD-Dev, demonstrating superior robustness in complex queries. Multi-round evolution yields an average 1.37-point performance boost, confirming the benefit of iterative refinement. The integration of memory, verification, and critique significantly enhances accuracy, especially on difficult datasets. SDPO further boosts performance, particularly on models like Coder-3B and Qwen3-4B, indicating the value of offline execution-aware supervision. Ablation results highlight the importance of each component, confirming that structured feedback and multi-turn optimization are critical for high-quality SQL synthesis.

This framework is immediately applicable to enterprise data management, enabling automated, accurate SQL generation from natural language queries, reducing reliance on manual scripting. It can be integrated into intelligent assistants, customer support bots, and data analysis tools, improving user experience and operational efficiency. Long-term, the approach paves the way for autonomous, self-improving database interfaces capable of handling increasingly complex queries, supporting multi-modal inputs, and adapting to diverse domains. Such systems could revolutionize how non-experts interact with large-scale data repositories, democratizing data access and analysis across industries.

Despite its strengths, the approach faces challenges with extremely nested or ambiguous queries where execution signals are insufficient for precise diagnosis. The multi-round process incurs higher computational costs, limiting real-time deployment in resource-constrained environments. Dependence on annotated execution data and critic feedback increases resource requirements, and scalability to very large schemas or multi-modal inputs remains to be validated. Future work should focus on efficiency improvements, reducing inference latency, and extending the framework to broader multi-task and multi-modal scenarios.