Abstract Current NL2SQL systems degrade sharply when confronted with practical constraints such as limited prompt length and the inability to fine-tune large language models (LLMs). Performance drop is especially pronounced in complex databases, where inaccurate schema linking, vague value conditions, and weak self-correction dominate the error surface. We propose Dynamic-SQL, an adaptive framework that couples multi-path chain-of-thought fusion with execution-based feedback correction. A dense–sparse hybrid vector space is first constructed to dynamically retrieve relevant schema elements, and an LLM is leveraged to generate an explicit schema subgraph. Real-value and few-shot exemplars are then injected to enrich the prompt and sharpen value conditioning. Multiple candidate SQL statements are produced via diverse reasoning paths; their chains of thought are fused to cover latent semantic interpretations, and execution feedback is exploited for iterative self-correction until convergence. On the BIRD benchmark, Dynamic-SQL, powered by the open source qwen2.5-coder-32b-instruct, reduces the average prompt length by 50.83% , raises strict schema-linking recall from 72.63% to 90.66% , and achieves 63.23% execution accuracy. By systematically addressing schema linking, exemplar augmentation, multi-path fusion reasoning, and self-correction, the framework offers a transferable paradigm for deploying LLMs in complex database querying scenarios.
Hao et al. (2026) studied this question.