Google提出DualSQL：多智能体强化学习训练Text-to-SQL
Great paper from Google and colleagues.
多智能体RL解决Text-to-SQL的新范式，小参数实现大模型性能，值得跟进研究。
Great paper from Google and colleagues.
来自 Google 及其同事的出色论文。
Trains Text-to-SQL agents using multi-agent RL.
使用多智能体强化学习训练 Text-to-SQL 代理。
(bookmark it)
(收藏此页)
This work proposes DualSQL, which splits Text-to-SQL into two agents, one that links the question to the right tables and columns and one that writes the SQL.
这项工作提出了 DualSQL，将 Text-to-SQL 任务拆分为两个代理：一个负责将问题链接到正确的表和列，另一个负责编写 SQL。
Both agents run on the same model weights, so a single multi-agent RL run trains both roles together. The agents can query the database through three tools while they reason.
两个代理共享同一组模型权重，因此单次多智能体强化学习运行即可同时训练这两种角色。在推理过程中，代理可以通过三种工具查询数据库。
Multi-agent RL tends to collapse during training, so the authors add guardrails on rollouts and a new reward, robust execution match, that judges SQL correctness more accurately.
多智能体强化学习在训练过程中容易崩溃，因此作者添加了 rollout 护栏以及一种新的奖励机制——鲁棒执行匹配（robust execution match），以更准确地判断 SQL 的正确性。
Trained on only 3,755 examples, DualSQL-4B reaches 68.0% execution accuracy on BIRD dev, matching earlier 7B models. DualSQL-8B reaches 71.1%, ahead of previous single-model systems with 32B parameters.
仅在 3,755 个样本上训练，DualSQL-4B 在 BIRD dev 集上达到了 68.0% 的执行准确率，与早期的 7B 模型持平。DualSQL-8B 达到了 71.1%，优于之前拥有 32B 参数的单模型系统。
Paper: https://arxiv.org/abs/2609.18135
论文：https://arxiv.org/abs/2609.18135
Chat with Paper: https://academy.dair.ai/papers/dualsql-text-to-sql-with-multi-agent-reinforcement-learning-2609.18135
与论文对话：https://academy.dair.ai/papers/dualsql-text-to-sql-with-multi-agent-reinforcement-learning-2609.18135
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力