On September 26, according to the official website of the globally authoritative evaluation benchmark BIRD-Bench, Ant Digital Technologies' data analysis intelligent agent Agentar-SQL has surpassed numerous domestic and international vendors including AT&T, Google Cloud, Tencent Cloud, and Alibaba Cloud, ranking first globally. This marks the highest achievement by a Chinese company on this leaderboard.
BIRD-Bench is recognized as the world's most authoritative natural language to SQL evaluation benchmark, requiring AI large models to convert natural language queries into Structured Query Language (SQL) and execute them stably in real, complex, large-scale production-grade databases.
The BIRD-Bench dataset covers 37 industry scenarios including finance, power, and healthcare, totaling 33GB and containing over 10,000 high-complexity query tasks. It serves as an authoritative platform for top global AI teams to demonstrate their technical capabilities.
Notably, Ant Digital Technologies' Agentar-SQL achieved first place on both the BIRD leaderboard's execution accuracy ranking (81.67 points) and execution efficiency ranking (77 points). This signifies that Ant Digital Technologies has achieved global leadership in technological innovation in the intelligent data querying field.
According to the introduction, the Agentar-SQL intelligent agent is built on Ant Digital Technologies' SQL large model Agentar-Scale-SQL, designed to enable users to easily complete complex data query tasks through natural language. Through the GSPO (Group Sequence Policy Optimization) reinforcement learning training method, it can enhance SQL intrinsic reasoning, allowing the large model to deeply consider SQL frameworks during the reasoning phase, avoid potential logical errors, and improve SQL logical accuracy.
Additionally, Agentar-SQL possesses multi-round reflective correction capabilities, enabling the model to review and correct generated SQL through multiple rounds, enhancing SQL language precision. Agentar-SQL also employs an innovative two-stage generation method, allowing the large model to generate multiple SQL candidates and then conduct pairwise "tournament" competitions among the SQL options to select the optimal SQL.