RLVR-based ReViSQL-K2.6 Achieves Human-Level Text-to-SQL Performance Without Scaffolding

Existing Text-to-SQL systems have relied on agent scaffolding consisting of multiple stages, such as schema linking, query generation, and self-correction, to handle complex schemas and ambiguous questions. However, this approach merely compensates for the model's reasoning limitations through prompt structure, resulting in performance approximately 11 points lower than the BIRD benchmark score recorded by human experts (92.96%).

A research team from UIUC and Bridgewater AIA Labs developed the ReViSQL-K2.6 model applying Reinforcement Learning with Verifiable Rewards (RLVR) to overcome these limitations. This approach focuses on enhancing the model's own reasoning capabilities without separate scaffolding. The core lies in two key improvements.

First, they constructed an expert-verified training set that eliminated label errors which could hinder RLVR training. Second, they introduced a reward-shaping technique targeting common failure modes of RLVR. Through this, the model became able to understand complex database schemas more accurately and generate queries.

ReViSQL-K2.6 recorded performance exceeding the human-level 92.96% in the SC-16 (majority voting among 16 samples) setting. This result was achieved without additional model calls or complex orchestration, with a cost of only $0.56 per task. This study demonstrates that model learning through iterative experience is more effective than static prompt engineering for Text-to-SQL tasks.