Keywords
constraint-aware reranking
schema drift
paired simulation
reproducibility
Abstract
We evaluated constraint-aware reranking for text-to-SQL execution under 34% schema drift. A deterministic paired simulation generated 56 cases and preserved a boundary-condition stratum. Mean execution accuracy changed from 0.534 to 0.571; the paired difference was +0.036 (95% interval +0.035 to +0.038). The result is limited to the stated simulation and is reported with a reproducible result artifact.
References
Zhang, B., Xie, H., Du, P., Chen, J., Cao, P., Chen, Y., Liu, S., Liu, K., & Zhao, J. (2023). ZhuJiu: A Multi-dimensional, Multi-faceted Chinese Benchmark for Large Language Models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 479-494. https://doi.org/10.18653/v1/2023.emnlp-demo.44
Liang, Z., liu, L., Quan, R., zou, M., li, D., Tang, Y., & qin, H. (2026). Hierarchical Adaptive Reward-based Reinforcement Learning Model for High-Precision Text-to-SQL Generation. https://doi.org/10.2139/ssrn.6767042
Wang, Y., Lv, H., & Qian, Y. (2026). CIR-SQL: A Dual-Model Intent Recognition Framework for Chinese Text-to-SQL. AI, 7(3), 91. https://doi.org/10.3390/ai7030091
Nascimento, E. R. S., & Casanova, M. A. (2024). Querying Databases with Natural Language: The use of Large Language Models for Text-to-SQL tasks. Anais Estendidos do XXXIX Simpósio Brasileiro de Banco de Dados (SBBD Estendido 2024), 196-201. https://doi.org/10.5753/sbbd_estendido.2024.240552
Mellah, Y., Kocaman, V., Ul Haq, H., & Talby, D. (2024). Efficient schema-less text-to-SQL conversion using large language models. Artificial Intelligence in Health, 1(2), 96. https://doi.org/10.36922/aih.2661
Souto Prego, B., Vilares, D., & Cabado Lousa, B. (2024). Large Language Model Based Chatbot for Database Interaction through Natural Language. VII Congreso XoveTIC: impulsando el talento científico, 335-342. https://doi.org/10.17979/spudc.9788497498913.47
Jiang, G., Li, W., Yu, C., Zhu, Z., & LI, W. (2025). FGCSQL: A Three-Stage Pipeline for Large Language Model-driven Chinese Text-to-SQL. https://doi.org/10.20944/preprints202502.1059.v1
Hairan, B., & Şahman, M. A. (2026). A comparative evaluation of large language models for detecting SQL injection vulnerabilities in web applications. PeerJ Computer Science, 12, e4015. https://doi.org/10.7717/peerj-cs.4015
Muppala, M. (2025). ETL pipelines and SQL database management. SQL Database Mastery: Relational Architectures, Optimization Techniques,and Cloud-Based Applications, 84-101. https://doi.org/10.70593/978-93-7185-191-6_5
Rehman, A. U. (2026). FinSight AI: Coupling a Large Language Model with a Live NoSQL Database for Conversational Financial Analytics and Rule-Assisted Fraud Screening. https://doi.org/10.2139/ssrn.7093099