AIDualSQL’s two-agent approach to turning questions into SQL
A post describing a paper from Google and colleagues says DualSQL divides the work: one agent links a question to the right tables and columns, while another writes the SQL.
TLDR
According to the post, DualSQL’s agents share model weights and train together using multi-agent reinforcement learning. They can query the database through three tools while reasoning. The post says the authors added training guardrails and a reward called “robust execution match” to judge SQL correctness more accurately. It reports that, after training on 3,755 examples, DualSQL-4B reached 68.0% execution accuracy on BIRD dev, matching earlier 7B models, while DualSQL-8B reached 71.1%, ahead of previous single-model systems with 32B parameters.
