DualSQL 8B tops 32B Text-to-SQL models
DualSQL is a novel Text-to-SQL framework that decomposes complex natural language database querying into two collaborative agents sharing a single model backbone: one dedicated to schema linking (identifying relevant tables and columns) and another focused on SQL synthesis. By jointly optimizing both agents through multi-agent reinforcement learning equipped with database interaction tools, rollout guardrails, and a Robust Execution Match (REX) evaluation metric, DualSQL enables an 8B model to achieve 71.1% execution accuracy on the BIRD benchmark, outperforming previous 32B parameter single-model baselines.
Multi-agent RL with shared backbone weights demonstrates that targeted agentic decomposition and environment feedback beat brute-force parameter scaling for structured enterprise data tasks.
* Decoupling schema linking from SQL writing mirrors human analyst workflows and prevents context pollution during query generation.
* DualSQL-8B surpassing 32B models highlights significant efficiency gains for deploying database agents in production.
* Rollout guardrails and the REX metric provide critical training stability, mitigating common failure modes in code-oriented RL.
DISCOVERED
2h ago
2026-09-22
PUBLISHED
2h ago
2026-09-22
RELEVANCE
AUTHOR
omarsar0