Draft-Conditioned Constrained Decoding for Structured Generation in LLMs

Researchers introduce Draft-Conditioned Constrained Decoding (DCCD), a training-free method that improves structured output generation in large language models by up to 24 percentage points. The technique uses a two-step process that first generates an unconstrained draft, then applies constraints to ensure valid outputs like JSON and API calls.

• →DCCD improves structured output accuracy by up to 24 percentage points over standard constrained decoding methods.
• →The method enables smaller model pairs to match or exceed the performance of much larger constrained baselines.
• →DCCD requires no additional training and works as a training-free inference procedure.
• →The approach separates semantic planning from structural enforcement to reduce generation errors.
• →Results show significant parameter efficiency gains, making structured generation more accessible.