跳转至

SQL-Zero：自我进化的 Text-to-SQL

文章背景与核心概要

传统的 Text-to-SQL 智能体训练严重依赖昂贵且特定领域的、由人工标注的自然语言与 SQL 配对数据，这在扩展到新数据库时构成了巨大的瓶颈。

SQL-Zero 表明，完全可以通过零标注配对数据来训练具有竞争力的 SQL 求解器。通过利用新颖的“出题者-求解者”自我博弈框架（proposer-solver self-play framework），挑战者和求解者均从同一个基础大语言模型（LLM）出发实现共同进化。真实标签（Ground truth）完全通过针对数据库本身的执行结果来确立。

SQL-Zero: Self-Evolving Text-to-SQL

arXiv: [2609.04697 [cs.AI]]
Submitted: September 4, 2026
Authors: Daniel Machado Pedrozo, Julia Soares Dollis, Bryan Lincoln Marques de Oliveira, Vinicius Alboneti Aguiar, Sávio Salvarino Teles de Oliveira, Telma Woerle de Lima Soares
Full-Text Links: View PDF | HTML Version | TeX Source

执行摘要

Training competitive Text-to-SQL agents traditionally relies heavily on expensive, domain-specific, human-annotated natural-language and SQL pairs, creating a significant bottleneck when scaling to new databases.

SQL-Zero demonstrates that a competitive solver can be trained with zero annotated pairs. By utilizing a novel proposer-solver self-play framework, both a challenger and a solver evolve starting from the same base Large Language Model (LLM). Ground truth is established purely through execution against the database itself.

核心方法论

Key Methodology

  • 出题者-求解者自我博弈： 挑战者动态生成与求解者当前难度水平相匹配的 SQL 对（专门针对“困难但可解”的查询）。
    • Proposer-Solver Self-Play: The challenger generates SQL pairs dynamically calibrated to the solver's current difficulty level (specifically targeting "hard but solvable" queries).
  • 交替更新： 两个角色通过 GRPO（群组相对策略优化，Group Relative Policy Optimization）进行交替迭代更新。
    • Alternating Updates: Both roles are updated iteratively using GRPO (Group Relative Policy Optimization) in alternating turns.
  • 多样性保障机制： 对挑战者应用模板级的重复惩罚，以防止生成过程中出现多样性崩溃。
    • Diversity Safeguards: A template-level repetition penalty is applied to the challenger to prevent diversity collapse during generation.

结果与性能

Results & Performance

  • BIRD 数据库性能提升（零标注）： 在 BIRD 开发集上，自我博弈训练使零样本（zero-shot）基础模型的性能在 3B 参数量下提升了 6.6 个百分点，在 7B 参数量下提升了 7.3 个百分点。
    • BIRD Database Improvements (Zero-Labels): Self-play training improves performance over the zero-shot base model on the BIRD development set by 6.6 points at 3B parameters and 7.3 points at 7B parameters.
  • 与人类标注黄金标准（Gold Standards）的对比： SQL-Zero 的得分高于使用完全相同的配方在相同数据库的人类 BIRD 黄金数据上训练的对照组（尽管精确的配对测试并未明确解决该差距）。
    • Comparison to Human-Annotated Gold Standards: SQL-Zero scores higher than a matched control trained using the exact same recipe on human BIRD gold data across the same databases (though an exact paired test does not definitively resolve that margin).
  • 跨数据库泛化能力： 迁移能力在很大程度上取决于模型规模：
    • Cross-Database Transfer: Transfer capabilities depend heavily on model scale:
  • 在 3B 规模下： 每次迭代在未见过的 Spider 数据库上以及词汇扰动（Spider-Syn）下均优于基础模型，其性能退化幅度小于 BIRD-gold 对照组。 > * At 3B: Every iteration outperforms the base model on unseen Spider databases and under lexical perturbation (Spider-Syn), degrading less than the matched BIRD-gold control.
  • 在 7B 规模下： 仅有第一次迭代成功保持了迁移性能。 > * At 7B: Only the first iteration successfully preserves transfer performance.

附加信息

Additional Information