DualSQL: Text-to-SQL with Multi-Agent Reinforcement Learning
Shijie Chen and Yu Su (Ohio State) with Yu Gan, Fatma Ozcan and colleagues at Google train both halves of a Text-to-SQL pipeline, schema linking and SQL generation, as two agents that share one model backbone and learn jointly through multi-agent RL.
Ask this paper
One backbone, two roles. The schema-linking agent and the SQL-generation agent share weights and scaffold, so a single RL run optimizes both instead of training two separate models.
Database tools in the loop. Three database access tools let the agents query the live database during multi-step reasoning rather than generating SQL from the schema text alone.
Rollout guardrails. A set of guardrail mechanisms on rollouts prevents the model collapse that multi-agent RL tends to hit, so accuracy keeps rising through training.
REX reward. Robust execution match is a new correctness metric that judges SQL equivalence more accurately than exact execution match and is used as the reward signal.
Small data, strong result. Trained on 3,755 examples, DualSQL-4B reaches 68.0% execution accuracy on BIRD dev, matching prior 7B models, and DualSQL-8B reaches 71.1%, above prior single-model systems with 32B parameters.
Abstract
State-of-the-art Text-to-SQL systems are typically multi-agent pipelines centered around two fundamental tasks: schema linking and SQL generation. However, existing work trains separate models for each task, failing to leverage the synergy between these interrelated tasks. In this work, we propose DualSQL, a new Text-to-SQL system consisting of two agents powered by a single model backbone. The agents share the same model weights and agentic scaffold, enabling joint optimization through a robust multi-agent reinforcement learning (RL) framework. We design three database access tools to facilitate effective multi-step reasoning grounded to interactions with the databases. To improve training and avoid model collapse, we introduce a set of rollout guardrail mechanisms that stabilizes multi-agent RL training, supporting DualSQL to keep improving during training. We also introduce a new SQL correctness metric, robust execution match (REX), to more accurately judge SQL correctness and assign reward signals. Being trained on only 3755 examples, DualSQL-4B achieves an impressive 68.0% execution accuracy on the BIRD development set, matching previous 7B models. DualSQL-8B further improves to 71.1%, outperforming previous state-of-the-art single-model solutions with 32B parameters. These results demonstrate the strength of joint multi-agent reinforcement learning for building high performance Text-to-SQL pipelines.