Researchers proposed DualSQL, a Text-to-SQL system with two agents sharing one model. The agents use a multi-agent reinforcement learning framework. Three database access tools enable multi-step reasoning with databases. Rollout guardrails prevent model collapse during training. A new metric, robust execution match, judges SQL correctness for reward signals. DualSQL-4B matches prior 7B models on execution accuracy. DualSQL-8B outperforms prior state-of-the-art single-model solutions with 32B parameters. The paper was submitted to arXiv on 16 September 2026.
What changed
Text-to-SQL systems used separate models for schema linking and SQL generation.
What it unlocks
Joint optimization of schema linking and SQL generation in a single model backbone.
- 68.0% execution accuracy on BIRD dev set
- 71.1% execution accuracy on BIRD dev set
- 3755 training examples
- 4B and 8B parameter models
Sources