Putting Task Expertise into RL Achieves Performance on Text-to-SQL (thinkingmachines.ai)

A new model, ReViSQL-K2.6, has achieved human-level accuracy in text-to-SQL tasks, surpassing the human benchmark of 92.96% on the BIRD leaderboard, marking a significant breakthrough for AI in natural language processing and database querying. Unlike previous methods that relied on scaffolding—breaking down tasks into separate stages—this approach employs reinforcement learning with verified rewards (RLVR) and leverages a curated training dataset known as BIRD-Platinum, which addresses the high noise levels found in conventional SQL datasets. This enhanced model not only outperforms leading AI models like GPT-5.6 Sol Ultra and Claude Fable 5 but does so at a fraction of their operational costs, making it a viable option for high-volume applications. Key advancements of ReViSQL-K2.6 include an expert-verified training set, which minimizes errors that can skew learning outcomes, and an innovative reward-shaping technique designed to ensure that the model emphasizes semantic equivalence and incorporates external knowledge effectively. By fine-tuning the reward signals and refining the data inputs, the model achieves superior reasoning capabilities in SQL query formulation. As a result, it can process complex queries more reliably, demonstrating the potential for AI systems to replicate human performance in intricate database tasks without traditional instructional scaffolding.