DecoSearch Text-to-SQL (DeepSeek): 88.31% Spider, 70.53% BIRD

DecoSearch, a training-free Text-to-SQL framework by Esteban Schafir, Xu Zheng, Hojat Allah Salehi, Zhuomin Chen, Mo Sha, Wei Cheng and Dongsheng Luo (submitted 16 Jun 2026 as arXiv:2606.17821), routes each question by reasoning complexity and repairs flawed plans at the decomposition level. With a DeepSeek backbone it achieves 88.31% execution accuracy on Spider and 70.53% on BIRD while consuming an order of magnitude fewer tokens than competing methods.

How does DecoSearch route and decompose queries?

DecoSearch routes each query to an appropriate reasoning tier using a lightweight Schema Selector and an LLM Judger: straightforward questions go to direct SQL generation and complex questions are escalated to a Directed Acyclic Graph of atomic sub-questions. The system then solves each atomic sub-question with a targeted SQL generation step and grounds decomposition via a retrieval-augmented generation component.

The pipeline begins with the Schema Selector pruning the full database schema to relevant tables and columns. The LLM Judger evaluates whether the question requires decomposition. If judged complex, DecoSearch constructs a DAG of atomic sub-questions; a RAG component supplies semantically similar training examples to ground each decomposed node. Each node maps to a dedicated SQL generation step and subsequent execution.

What components handle errors and repair plans?

DecoSearch uses a Topology Refiner to restructure reasoning plans when execution failures indicate a flawed decomposition rather than a fixable SQL error, enabling plan-level repair rather than repeated SQL edits. This means the system can change the decomposition topology when execution signals a deeper planning problem.

In practice the Topology Refiner inspects execution failures to decide whether the issue stems from an incorrect decomposition. When it detects a plan-level fault, it restructures the DAG and reissues the atomic generation steps. The paper presents this flow as an explicit part of the training-free wrapper so that the core LLM backbones remain unchanged.

How does DecoSearch perform on benchmarks?

DecoSearch achieves 70.53% execution accuracy on BIRD and 88.31% on Spider with a DeepSeek backbone, outperforming all training-free baselines cited in the paper while using an order of magnitude fewer tokens than competing methods. The authors also report that DecoSearch operates as a model-agnostic wrapper that consistently improves fine-tuned SQL generation backbones without modifying the underlying pipelines.

Those two benchmark numbers are the primary performance claims in the submission. The paper frames the gains as coming from routing queries by complexity, targeted atomic SQL steps, RAG grounding, and plan-level repair via the Topology Refiner.

Why it matters

DecoSearch addresses where many LLM-based Text-to-SQL approaches struggle: multi-step, data-aware reasoning. Routing by complexity and repairing the decomposition plan shifts effort away from repeated SQL regeneration toward correcting the reasoning plan itself. That reduces token consumption and, in the authors' results, raises execution accuracy on standard benchmarks while leaving existing model backbones untouched.

What to watch

Look for code and demos linked from the arXiv entry or the authors' repositories that show the Topology Refiner in action and reproduce the 88.31% Spider and 70.53% BIRD figures. Also watch whether other backbones beyond DeepSeek show similar token-efficiency gains when wrapped by DecoSearch.

Reference: DecoSearch: Complexity-Aware Routing and Plan-Level Repair for Text-to-SQL, Esteban Schafir et al., arXiv:2606.17821, submitted 16 Jun 2026. DOI: https://doi.org/10.48550/arXiv.2606.17821