Write a comment

Summary

This paper identifies a fundamental limitation of current Text-to-SQL systems: their reliance on single static workflows that cannot adapt to the heterogeneous distribution of real-world queries. The authors propose SquRL, a reinforcement learning framework that trains LLMs to dynamically construct task-specific workflows at inference time by selecting among modular actor components — schema reduction, schema linking (parse), SQL generation, and SQL refinement (optimize). They provide theoretical analysis showing that dynamic policies consistently outperform the best static workflow, with gains driven by workflow heterogeneity. Two training mechanisms — dynamic actor masking for exploration and pseudo rewards for efficiency — address the sparse reward problem inherent to execution-based feedback. Experiments on multiple benchmarks demonstrate consistent improvements over static baselines, with larger gains on complex and out-of-distribution queries.

Strengths

The framing of Text-to-SQL as a tool-based reasoning problem is the paper's most conceptually valuable contribution. Casting workflow construction as dynamic selection among modular actors — rather than as a fixed pipeline — naturally connects to the broader trend toward agentic LLM systems and opens the door to principled optimization of multi-step reasoning workflows. The Actor-Template-Workflow abstraction is clean and well-motivated: separating the structural template from the concrete actor instantiation allows the system to reason about workflow composition at the right level of abstraction, without over-committing to specific implementation choices.

The theoretical analysis supporting the claim that dynamic policies outperform the best static workflow is an important contribution that is often absent from empirical NLP papers. Grounding the empirical gains in a theoretical bound, and tying those bounds to workflow heterogeneity as the driving factor, gives the paper intellectual depth beyond benchmark comparisons. This also provides a principled criterion for practitioners: if the workflow candidate pool is heterogeneous, dynamic selection is worth the added complexity; if workflows are largely homogeneous, static pipelines may suffice.

The two training mechanisms address real engineering challenges. Dynamic actor masking is a sensible solution to the exploration problem in RL for text generation — without it, the policy would collapse to a locally optimal but globally suboptimal workflow preference. Pseudo rewards to accelerate training under delayed feedback are a practical contribution that should transfer to related agentic training settings beyond Text-to-SQL.

The experimental focus on complex and out-of-distribution queries is exactly the right evaluation axis. Enterprise Text-to-SQL systems encounter long-tail queries — multi-hop joins across dozens of tables, queries with implicit business logic, dialect-specific constructs — that benchmark-tuned static pipelines systematically fail on. The finding that SquRL's gains are most pronounced on these cases is the result most relevant to real-world deployment.

Weaknesses and Limitations

The paper evaluates on standard academic benchmarks (Spider, BIRD), but enterprise Text-to-SQL involves additional challenges that these benchmarks do not capture: databases with hundreds of tables where schema reduction is the dominant bottleneck, queries requiring external business context not present in the schema, multi-turn conversational interfaces, and strict latency requirements. The generalization of dynamic workflow construction to these settings is not discussed, and the claim that the approach addresses "real-world scenarios" warrants more careful qualification.

The reinforcement learning training setup introduces computational costs that are not fully characterized. Training a policy to select workflows requires repeated execution of complete SQL pipelines — each involving LLM calls for schema linking, generation, and refinement — to obtain execution-based rewards. The paper introduces pseudo rewards to improve efficiency, but does not provide a clear comparison of training compute versus the static baselines. For practitioners evaluating whether to adopt SquRL over a well-tuned static pipeline, this information is essential.

The actor pool in the experiments appears to be drawn from a fixed set of existing methods. It is unclear how sensitive the dynamic policy is to the composition and quality of this actor pool. If the pool contains a dominant actor that outperforms others on most query types, the dynamic policy may degenerate to near-static behavior. A systematic ablation varying pool composition and size would clarify when dynamic selection provides genuine benefit versus when it approximates static selection of the best individual actor.

The connection between workflow heterogeneity — the theoretical driver of dynamic policy gains — and measurable properties of the actor pool or query distribution is not operationalized. The paper demonstrates empirically that heterogeneity matters, but does not provide a practical way for users to assess whether their specific deployment context has sufficient heterogeneity to justify the overhead of RL training. A heterogeneity metric, even an approximate one, would make the theoretical insight actionable.

Suggestions

The paper would benefit from a qualitative case study showing the workflow selection decisions made by SquRL on representative complex queries from Spider 2.0 or BIRD, alongside the decisions a static pipeline would make. This would give readers concrete intuition for when and why dynamic selection helps, and would make the paper more accessible to practitioners building production systems.

An analysis of how the policy generalizes to new actor implementations not seen during training would be valuable. If practitioners add a new schema linker or SQL generator, can the existing SquRL policy incorporate it without retraining, or does the policy need to be retrained from scratch? This question is central to the framework's long-term maintainability.

The authors should discuss the relationship between SquRL and recent work on RL-trained reasoning agents for code generation, such as Search-R1 and similar frameworks. The technical challenges — sparse execution feedback, exploration in a combinatorial action space, reward shaping — overlap substantially, and connecting to this literature would situate the contribution more clearly.

Overall Assessment

This is a technically solid and well-motivated paper that addresses a genuine limitation of current Text-to-SQL systems. The Actor-Template-Workflow abstraction is clean, the theoretical analysis adds depth, and the empirical results on complex queries are compelling. The main weaknesses are the gap between benchmark evaluation and enterprise deployment conditions, the undercharacterized computational costs of RL training, and the lack of a practical operationalization of the workflow heterogeneity criterion. These are addressable gaps that do not undermine the core contribution. Recommended for acceptance with minor revisions.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they did not use generative AI to come up with new ideas for their review.