Bootstrapping Semantic Layer from Execution for Text-to-SQL

In three linesGATE (Grounding After Test from Execution) bootstraps missing groundings in text-to-SQL via execution feedback. The system keeps multiple grounding hypotheses open, executes already-grounded parts to obtain observations, then validates and stores confirmed hypotheses. This execution-grounded memory accumulates and improves performance on real-world and controlled benchmarks.

Summary generated by Claude — human-verified