Presentation description

We propose a reinforcement learning based approach to improve NL-to-SQL translation by embedding an SQL engine within the training loop, enabling LLM's to explore query formulations interactively. Using Group Relative Policy Optimization (GRPO), the model learns from execution feedback, optimizing for syntactical correctness, Relaxed Exact Match Score, and Exact Match Score. This neuro-symbolic framework enhances SQL generation accuracy and reasoning capabilities for structured data tasks.