Structured Output

Structured output constrains a language model response to a schema such as JSON, YAML, a table, a function call, or a domain-specific format. It makes outputs easier to parse, validate, test, and feed into tools.

If ${\mathcal{Y}}_{\mathrm{valid}}$ is the set of valid outputs, constrained generation aims for:

$$\hat{y}\in {\mathcal{Y}}_{\mathrm{valid}}$$

One way to view decoding is:

$$\hat{y}=\arg \underset{y\in {\mathcal{Y}}_{\mathrm{valid}}}{\max }{p}_{\theta }(y\mid x)$$

where invalid outputs are excluded or rejected.

Key Ideas

• A schema is a contract between the model and downstream code.
• Valid syntax is not the same as correct semantics.
• Structured output reduces parsing ambiguity but does not remove the need for validation.
• Tool arguments should be checked before execution.
• For public wiki workflows, structured output can turn paper reading, claim extraction, and link suggestions into reviewable artifacts.

Practical Checks

• What schema must the output satisfy?
• Are required fields, enums, ranges, and nested objects explicit?
• Are invalid outputs rejected, repaired, or retried?
• Does validation check semantics, not only syntax?
• Is untrusted model output isolated before it affects tools or files?
• Are evidence fields required when the output contains factual claims?

Related

• Inference contract
• Structured prediction
• Evidence-grounded generation
• Decoding
• Tool calling
• Tool contract
• Tool use
• Verification loop