Definition
Plain language
The shared core inside an LLM judge that does the actual judging, before any specific answer format is chosen.
As stated in the literature
In the Judge Circuits framing, the format-agnostic sub-network at intermediate transformer layers that encodes an abstract judgment along a roughly one-dimensional axis, shared across rating and classification prompts.
Why it matters: It suggests that LLM judgments across different prompt formats may not be as independent as they look, since they share a common internal verdict signal.
For example, the same internal direction inside the judge model fires when scoring 'rate this answer 1-5' and when scoring 'is this answer correct yes/no', then gets converted into whichever output format the prompt requested.
Heard on the show
“They call the shared core the Latent Evaluator — the part of the model that actually does the judging.”Episode 055 — Why LLM Judges Flip Their Verdicts When You Change the Question Format