AgentNLQ achieves 78.1% accuracy on BIRD benchmark with multi-agent NL2SQL

Olena Bogdanov and colleagues have released AgentNLQ, a general-purpose multi-agent framework for converting natural language questions into SQL queries. The system uses a novel orchestrator powered by large language models (LLMs) that plans query generation, orchestrates sub-tasks, reflects on intermediate results, and self-corrects errors before final output. This iterative process helps AgentNLQ reach 78.1% semantic accuracy on the challenging BIRD (Big Bench for LaRge-scale Database) benchmark, significantly outperforming prior end-to-end approaches.

The method introduces two key innovations. First, an advanced schema enrichment technique that transforms raw database schemas into context-aware metadata by incorporating user-provided business rules and domain-specific annotations. This gives the LLM a richer understanding of column meanings and relationships. Second, the orchestrator’s self-correction loop allows the system to detect and fix common SQL mistakes, such as incorrect joins or missing filters, without human intervention.

AgentNLQ was evaluated across multiple domains and datasets from the BIRD-SQL benchmark, demonstrating consistent generalization. The authors note that while current LLMs have made remarkable progress, NL2SQL still lags behind expert human performance. AgentNLQ’s multi-agent design—where separate agents handle schema understanding, query planning, and validation—represents a practical step toward closing that gap. The paper is available on arXiv under ID 2605.19010.