Research map / Large language models / Language, safety and interpretability

Structure-Guided Large Language Models for Text-to-SQL Generation

ICML 2025 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SGU-SQL, a text-to-SQL framework that builds query and schema graphs, links them with an RGAT-based structure learner, and decomposes SQL generation via syntax-tree prompting for GPT-4.

Problem

LLM-based text-to-SQL struggles with ambiguous user intent, complex database schemas, and the syntax structure of SQL; existing decomposition methods are ineffective for SQL generation.

Approach

Combine dual query-schema graph construction and RGAT-based structure linking with syntax-tree-based decomposition of SQL generation into subtasks that an LLM (GPT-4) solves incrementally.

Key results

• SGU-SQL with GPT-4 reaches 87.95% execution accuracy on Spider-dev, above the strongest listed baseline SuperSQL (86.82%).
• On BIRD dev, SGU-SQL with GPT-4 reaches 61.80% execution accuracy versus 58.60% for SuperSQL.
• Ablation shows removing syntax-based decomposition causes the largest drop (-5.60 on Spider-dev, -8.02 on BIRD-dev).

Measured effects reported in the paper

Surprising findings

• Chain-of-thought prompting reduced PaLM-2 accuracy relative to few-shot prompting, according to the paper.
• Table results show inconsistency: the main Spider table reports SGU-SQL EM 78.26 whereas the Spider-dev ablation reports 87.95 execution accuracy.

Limitations

• Relies on closed-source GPT-4, making reproducibility and cost difficult.
• Reported results contain inconsistencies across tables and text (e.g., differing numbers for the same configuration), and the paper provides limited methodological detail on the graph learner and decomposition.

Experimental setup

Datasets: Spider BIRD

Benchmarks: Spider BIRD Spider 2.0-Snow Spider 2.0-Lite

Metrics: execution accuracy exact match accuracy valid efficiency score

Baselines: RAT-SQL LGESQL RESDSQL Graphix-T5 DIN-SQL DAIL-SQL SuperSQL MAC-SQL C3-GPT PURPLE CHESS CHASE-SQL

Builds on: RAT-SQL DIN-SQL DAIL-SQL MAC-SQL RGAT

Closest papers by meaning

• ROUTE: Robust Multitask Tuning and Collaboration for Text-to-SQL (ICLR 2025 · similarity 0.91)
• CogSQL: A Cognitive Framework for Enhancing Large Language Models in Text-to-SQL Translation (AAAI 2025 · similarity 0.90)
• SQLens: An End-to-End Framework for Error Detection and Correction in Text-to-SQL (NeurIPS 2025 · similarity 0.89)
• Alpha-SQL: Zero-Shot Text-to-SQL using Monte Carlo Tree Search (ICML 2025 · similarity 0.89)
• Enhancing SQL Query Generation with Neurosymbolic Reasoning (AAAI 2025 · similarity 0.88)
• PARROT: A Benchmark for Evaluating LLMs in Cross-System SQL Translation (NeurIPS 2025 · similarity 0.87)