SPOC-SQL: Stage-wise Preference Optimization for Controllable Text-to-SQL

Yingnan Chen; Chun Ding; Tianshi Xu; Xu Yang; Si Wu

cs.CLAug 24, 2026

SPOC-SQL: Stage-wise Preference Optimization for Controllable Text-to-SQL

Authors: Yingnan Chen, Chun Ding, Tianshi Xu, Xu Yang, Si Wu

Organizations: South China University of Technology, Guangzhou, Guangdong 510006, P.R. China

Abstract

Text-to-SQL aims to translate natural language questions into executable SQL queries over relational databases, requiring multi-stage structured reasoning over database schemas and query constraints. However, existing methods treat this task as single-step generation, where models optimize entire SQL sequences without targeted feedback at key decision points and lack support for interacting with and controlling the intermediate generation process. To address this issue, we propose SPOC-SQL, which decomposes Text-to-SQL into four sequential subtasks following standard SQL execution logic and designs stage-specific optimization strategies for the model to learn key decisions. Specifically, we propose the implementation of fine-grained preference optimisation at key decision points across SQL stages, with the objective of enhancing structured decision-making during query construction. Furthermore, a structured decomposition strategy is designed, facilitating stage-wise intervention and correction through explicit intermediate representations. This results in more controllable and reliable SQL generation. Experiments demonstrate that incorporating stage-wise human knowledge consistently improves performance, validating the effectiveness of stage perception controllable generation.

Figures & tables

Figure 1: The workflow of SPOC-SQL. (1) The task parsing function $\mathcal{F}$ decomposes single-turn queries into multi-turn sequences with stage-specific sub-requirements. (2) LoRA fine-tuning combined with preference-based optimization to enhance decision capability through fine-grained policy optimization. (3) Evaluation of multi-turn QA capabilities. (4) At inference, $\mathcal{F}$ decomposes user requirements into sub-tasks for iterative confirmation, generating the final SQL.

Method	Model	Venue	SF	WH	GH	OL	Dev	Real.
ACT-SQL [ 20 ]	ChatGPT	EMNLP’23	–	–	–	–	80.4	75.8
C3 [ 6 ]	ChatGPT	arXiv’23	–	–	–	–	81.8	75.4
DIN-SQL [ 13 ]	GPT-4	NeurIPS’23	–	–	–	–	82.8	78.1
ACT-SQL [ 20 ]	GPT-4	EMNLP’23	–	–	–	–	82.9	–
DAIL-SQL [ 7 ]	DS-V3	VLDB’24	–	–	–	–	83.2	77.2
DAIL-SQL [ 7 ]	GPT-4	VLDB’24	–	–	–	–	84.4	75.6

Table 1: Comparative performance on Spider benchmarks.

Method	Data	Model	SF	WH	GH	OL	Easy	Med.	Hard	X-Hard	All
MCS-SQL [ 9 ]	Dev	GPT-4	–	–	–	–	94.0	93.5	88.5	72.9	89.5
DAIL-SQL [ 7 ]	Dev	DS-V3	–	–	–	–	93.5	83.0	83.9	67.5	83.2
[1pt/1pt] SPOC-SQL	Dev	DS-V3	–	–	–	–	95.2	84.3	85.6	74.1	85.5
			$\checkmark$	–	–	–	96.0	90.1	89.1	79.5	89.7
			$\checkmark$	$\checkmark$	–	–	97.2	91.9	97.1	83.7	92.7
			$\checkmark$	$\checkmark$	$\checkmark$	–	97.6	96.6	96.6	87.5	95.4

Table 2: Performance by difficulty on Spider-Dev and Spider-Realistic.

Figure 2: Performance of SPOC-SQL across different datasets and models: Evaluations on DeepSeekV3(DSV3) and Qwen3-72B(QW3) across Spider-Dev(Dev) and Spider-Realistic(Real) Datasets.

Figure 3: Comparison of models on multi-turn QA, single-turn QA, and error-correction QA tasks.

Model	Spider-Dev	Spider-Real.	T2S-MTD
SPOC-SQL	95.6	93.1	84.6
w/o PMDO	93.2	90.1	80.7
w/o RDIM	87.5	84.6	78.4
LoRA-only	83.7	79.8	75.4
Base	80.1	74.5	66.3

Table 3: Ablation study of SPOC-SQL.

Panel of Figure 4 — Figure 4: Case Study: (a) Decomposition of complex nested queries and set operations; (b) Modular parsing of multi-component SQL queries.

Figure 5: Comparison of SQL Info versions across subjective and objective metrics.

Text-to-SQL

Text-to-SQL translates natural-language questions into executable database queries, linking user intent to schemas, values, and SQL logic.

See all 173 papers

SQLConductor: Search-to-Policy Learning for Step-wise Text-to-SQL Orchestration
Jun 22, 2026Yizhang Zhu, Zhangyang Peng, Boyan Li +1Text-to-SQL RL for Tool Use
HKUST(GZ) Guangzhou, China
EXPO-SQL: Execution-based Clause-level Policy Optimization for Text-to-SQL
Apr 29, 2026Jaehoon Lee, CheolWon Na, Suyoung Bae +4Reward Shaping Text-to-SQL
College of Computing and Informatics Sungkyunkwan University, South Korea
ACE-SQL: Adaptive Co-Optimization via Empirical Credit Assignment for Text-to-SQL
Jun 4, 2026Xiaobing Chen, Ai Jian, Eryu Guo +1Execution-Guided Code Generation Schema Linking
Harbin Engineering University, Harbin, China · Harbin Institute of Technology, Harbin, China · Beijing University of Posts and Telecommunications, Beijing, China

Jun 22, 2026J/K move · Enter open · S save

SQLConductor: Search-to-Policy Learning for Step-wise Text-to-SQL Orchestration

Yizhang Zhu, Zhangyang Peng, Boyan Li, Yuyu Luo

HKUST(GZ) Guangzhou, China

Text-to-SQL RL for Tool Use

Open paper AskPDF arXiv

Text-to-SQL enables users to access relational databases via natural language, but real-world settings remain challenging due to coordinated reasoning over complex database environments. Existing systems often use multi-stage pipelines or reasoning models specialized for individual stages. However, fixed pipelines rely on predefined stage orders, limiting their adaptivity to query demands and intermediate evidence. Recent orchestration-based methods provide flexibility by composing specialized modules for each query, but typical plan-then-execute approaches still commit to a complete workflow before execution and cannot adapt to intermediate artifacts and feedback. In this paper, we propose SQLConductor, a step-wise orchestration learning framework for Text-to-SQL. SQLConductor formulates Text-to-SQL subtasks as specialized actions for workflow composition and trains a policy model to select the next action based on intermediate artifacts and feedback. To learn this policy, SQLConductor introduces Search-to-Policy Learning, which uses Monte Carlo Tree Search to explore candidate workflows and stability estimation to identify robust supervision. The policy model is trained with Stability-weighted Supervised Fine-tuning to prioritize high-quality orchestration patterns and further enhanced through Curriculum Reinforcement Learning. This transforms offline workflow search into a deployable policy for step-wise orchestration at inference time. Experiments on BIRD-Dev and out-of-distribution datasets show that SQLConductor achieves superior execution accuracy and strong generalization, reaching 73.2% EX on BIRD-Dev with a compact orchestration policy coordinating frozen larger action models, outperforming prior methods that directly train comparable or larger Text-to-SQL backbones. Further analyses show that the learned policy adapts orchestration to diverse query demands.

Abstract

Text-to-SQL

Explore similar work

SQLConductor: Search-to-Policy Learning for Step-wise Text-to-SQL Orchestration

EXPO-SQL: Execution-based Clause-level Policy Optimization for Text-to-SQL

ACE-SQL: Adaptive Co-Optimization via Empirical Credit Assignment for Text-to-SQL