title: Constrained decoding
aliases: [constrained decoding, structured decoding, structured output]
tags: [term]
category: AI & NLP
summary: Forcing a language model's output to fit a fixed structure or set of allowed choices instead of letting it generate free text.

Constrained decoding

Also called structured decoding, constrained decoding is a technique for restricting what a language model is allowed to produce. Instead of letting the model write any text, the system requires the output to fit a predefined structure (a fixed set of fields, a list of allowed values, or a strict format) and the model can only fill in the blanks within those limits.

This is useful when free-form generation is risky. A general-purpose model asked to translate will happily invent a fluent-sounding answer whether or not it is correct (see Ojibwe Chat). Constraining the model instead means it can only choose among options that someone has vetted in advance (see Yaduha).