Research map / Large language models / Agents, code and math
LearNAT: Learning NL2SQL with AST-guided Task Decomposition for Large Language Models
Weibin Liao, Xin Gao, Tianyu Jia, Rihong Qiu, Yifan Zhu, Yang Lin, Xinyu Ma, Junfeng Zhao, Yasha Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| provide manually written subtask decompositions to the LLM (ablation) | Qwen2.5-Coder-32B · 500 BIRD-train examples | performance: 30.4% up | Improved |
| let the LLM decompose queries by prompting (ablation) | Qwen2.5-Coder-32B · 500 BIRD-train examples | performance: 3.4% up | No clear effect |
| use language models to verify subtask correctness by semantic similarity (ablation) | Sentence-Transformer and GLM-4 · 500 subtasks | verification accuracy (%): 46.8 and 36.0 total accuracy | Hurt |
| add AST-guided search and pruning to MCTS decomposition (ablation) | Decomposition Synthesis Procedure · BIRD-train decomposition | decomposition success rate and token cost: 80.00% success; +8.45% over naive MCTS; 56.38% fewer tokens | Improved |
| add self-improving adaptive demonstrations over rounds (ablation) | Decomposition Synthesis Procedure · BIRD-train decomposition | decomposition success rate: +1.32% first round, then +0.4% and +0.27% | Mixed |
| add LearNAT training (decomposition data, SFT and margin DPO) (ablation) | Qwen2.5-Coder-7B · BIRD-dev total | execution accuracy (%): 47.5 -> 58.1 | Improved |
| add AST guidance to decomposition synthesis (ablation) | LearNAT · BIRD-dev total (7B) | execution accuracy (%): 53.1 -> 58.1 | Improved |
| add SFT warm-up before DPO (ablation) | LearNAT · BIRD-dev total (7B) | execution accuracy (%): 54.5 -> 58.1 | Improved |
| add margin-aware DPO stage (ablation) | LearNAT · BIRD-dev total (7B) | execution accuracy (%): 53.7 -> 58.1 | Improved |
| use AST margin in DPO instead of standard DPO (ablation) | LearNAT · BIRD-dev total (7B) | execution accuracy (%): 56.6 -> 58.1 | Improved |
| use margin-aware DPO instead of KTO or IPO (ablation) | LearNAT · BIRD-dev total (7B) | execution accuracy (%): KTO 54.6, IPO 53.7 -> 58.1 | Improved |
| use task decomposition instead of CoT reasoning (ablation) | LearNAT · BIRD-dev total (7B) | execution accuracy (%): 49.3 -> 58.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Alpha-SQL — Alpha-SQL: Zero-Shot Text-to-SQL using Monte Carlo Tree Search (ICML 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ROUTE: Robust Multitask Tuning and Collaboration for Text-to-SQL (ICLR 2025 · similarity 0.91)
- SQL-R1: Training Natural Language to SQL Reasoning Model By Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- CogSQL: A Cognitive Framework for Enhancing Large Language Models in Text-to-SQL Translation (AAAI 2025 · similarity 0.89)
- Alpha-SQL: Zero-Shot Text-to-SQL using Monte Carlo Tree Search (ICML 2025 · similarity 0.88)
- Structure-Guided Large Language Models for Text-to-SQL Generation (ICML 2025 · similarity 0.87)
- SQLens: An End-to-End Framework for Error Detection and Correction in Text-to-SQL (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
