From Sampling to Grammars: Making LLMs Reliably Output Structured Data (Even for Thinking Models)

Use efficient sampling plus grammar constraints to guarantee format today, but expect models to natively emit structured outputs tomorrow—especially when you let them think first, then constrain.