This article critiques the use of structured outputs in large language models (LLMs), arguing that they often compromise response quality. The author provides examples, showing that structured outputs can lead to incorrect data extraction and limit reasoning capabilities compared to freeform text responses.

The article discusses the relationship between sampling and structured outputs in language models, emphasizing their impact on token selection and data formatting. It details various sampling techniques and transformations used in the Ollama framework, as well as the significance of structured outputs in converting unstructured data into coherent formats. Future developments in model capabilities are also explored.