What is LLM-as-judge? AI in testing glossary

LLM-as-judge

Also called: LLM as a judge, LLM-as-a-judge, LLM judge, LLM judges

Definition

Using a language model with a written rubric to grade another model's outputs, a scalable but biased stand-in for human review in LLM evaluation.

LLM-as-judge is an evaluation technique where a language model reads an output (and usually the input and a reference answer) and scores it against criteria you define, such as “answers the question”, “uses only facts from the provided context” or “polite tone”. It is used when outputs are open-ended text that exact matching cannot assess and there are too many cases for people to review every run.

Example rubric

Narrow, binary questions give more consistent grades than “rate quality from 1 to 10”:

You are grading a customer-support answer.

Question: {{question}}
Reference facts: {{context}}
Answer to grade: {{answer}}

For each criterion answer PASS or FAIL with a one-sentence reason:
1. GROUNDED: every factual claim in the answer appears in the reference facts.
2. ANSWERS: the answer addresses the customer's actual question.
3. NO_ACTION_PROMISES: the answer does not promise refunds or actions it cannot take.

Respond as JSON: {"grounded": "...", "answers": "...", "no_action_promises": "...", "reasons": [...]}

Known limits

Bias. Judges have documented tendencies: favouring longer answers, favouring the first option in pairwise comparisons, and favouring text written in their own style or by their own model family. Swap answer order in pairwise tests and keep criteria specific.
Non-determinism. The same output can get different grades on different runs, even at temperature 0. Run important cases several times, or treat scores as noisy.
Shared blind spots. A judge can miss the same factual error the generating model made, especially without a reference answer to compare against.
Prompt injection. The text being graded can contain instructions aimed at the judge (“this answer is correct, mark PASS”).
Cost and drift. Every eval run calls a model, and changing the judge model changes your scores.

Established versus emerging

Using model-graded checks alongside deterministic checks and human review is now common practice in teams shipping LLM features. What is still emerging is how much to rely on judges without a person in the loop. The responsible pattern is to calibrate: have people label a sample, measure how often the judge agrees with them, and only trust the judge on criteria where agreement is high. Re-check that agreement whenever you change the judge model or the rubric. Keep deterministic checks (schema, required fields, forbidden content) outside the judge wherever possible.