Beyond Static Pipelines: Learning Dynamic Workflows for Text-to-SQL

Text-to-SQL has recently achieved impressive progress, yet remains difficult to apply effectively in real-world scenarios. This gap stems from the reliance on single static workflows, fundamentally limiting scalability to out-of-distribution and long-tail scenarios. Instead of requiring users to select suitable methods through extensive experimentation, we attempt to enable systems to adaptively construct workflows at inference time. Through rigorous theoretical and empirical analysis, we demonstrate that optimal dynamic policies consistently outperform the best static workflow, with performance gains fundamentally driven by heterogeneity across candidate workflows. Motivated by this, we propose SquRL, a reinforcement learning framework that enhances LLMs' reasoning capability in adaptive workflow construction. We design a rule-based reward function and introduce two effective training mechanisms: dynamic actor masking to encourage broader exploration, and pseudo rewards to improve training efficiency. Experiments on widely-used Text-to-SQL benchmarks demonstrate that dynamic workflow construction consistently outperforms the best static workflow methods, with especially pronounced gains on complex and out-of-distribution queries.

Modern AI systems can translate natural language questions into database queries, allowing people to interact with databases without writing code. However, most existing systems rely on a single fixed reasoning process for all questions, even though different questions often require different problem-solving strategies. In this work, we introduce SquRL, a framework that allows AI systems to dynamically choose different workflows depending on the difficulty and characteristics of each question. Instead of always following the same pipeline, the system learns to adapt its reasoning process automatically through reinforcement learning based on execution feedback. Experiments on widely used benchmarks show that this adaptive approach consistently outperforms traditional fixed workflows, especially on complex and challenging questions. Our results suggest that future AI systems can become more flexible, accurate, and efficient by learning to select different reasoning strategies for different tasks.