Definition
Plain language
The small piece at the end of an LLM judge that translates its inner verdict into whatever output format the prompt asked for.
As stated in the literature
In the Judge Circuits framing, format-specific attention heads at the terminal layers of an LLM judge that route the Latent Evaluator's abstract judgment into the specific token demanded by the prompt.
Also called: Task Formatters
Why it matters: Separating formatting from evaluation explains why a judge model can know the right answer yet still output the wrong label when the prompt format changes.
For example, after the model has internally decided one response is better, the Task Formatter is what turns that judgment into the literal token 'A' or 'B' the prompt asked for.
Heard on the show
“And they call the format-specific output machinery the Task Formatters — the parts that route the abstract judgment into "five" or "yes" or "entailment" or whatever the prompt asked for.”Episode 055 — Why LLM Judges Flip Their Verdicts When You Change the Question Format