TEXT-TO-SQL MODELS MAKE CHOICES WITHOUT SAYING SO

Summary from the TLDR Data newsletter

SpecBridge shows that text-to-SQL models often silently choose one interpretation of ambiguous questions, while answering each interpretation separately improves coverage by 29-63 percentage points without reducing precision on clear requests.