Researchers have developed a new approach to translating natural language questions into SQL queries, called Text-to-SQL. They fine-tuned a large language model using reinforcement learning on a dataset with verified data, without building complex multi-stage pipelines. This method achieved human-level accuracy on benchmarks, outperforming other systems by 10-22%. The team also identified two common failure modes of standard reinforcement learning and proposed a new reward shaping method to address them. --- Why it matters: This matters because it shows that current approaches to Text-to-SQL may be overcomplicating the problem. By fine-tuning a language model on verified data, researchers can achieve human-level performance without building complex pipelines. This could simplify the development of Text-to-SQL systems and improve their accuracy. Source: https://arxiv.org/abs/2603.20004