Thinking Before Constraining: A Unified Decoding Framework for Large Language Models
This addresses the problem of balancing reasoning flexibility and output reliability in large language models for tasks like classification and reasoning, representing an incremental improvement over existing constrained decoding methods.
The paper tackles the trade-off between natural generation's rich reasoning and structured generation's guaranteed-parsable outputs by proposing a unified decoding framework that allows LLMs to reason freely until trigger tokens, then switch to structured generation, achieving up to 27% accuracy gain over natural generation with only 10-20 extra tokens.
Natural generation allows Language Models (LMs) to produce free-form responses with rich reasoning, but the lack of guaranteed structure makes outputs difficult to parse or verify. Structured generation, or constrained decoding, addresses this drawback by producing content in standardized formats such as JSON, ensuring consistency and guaranteed-parsable outputs, but it can inadvertently restrict the model's reasoning capabilities. In this work, we propose a simple approach that combines the advantages of both natural and structured generation. By allowing LLMs to reason freely until specific trigger tokens are generated, and then switching to structured generation, our method preserves the expressive power of natural language reasoning while ensuring the reliability of structured outputs. We further evaluate our approach on several datasets, covering both classification and reasoning tasks, to demonstrate its effectiveness, achieving a substantial gain of up to 27% in accuracy compared to natural generation, while requiring only a small overhead of 10-20 extra tokens.