What is LLM as Judge?
Also called model-graded evaluation, AI grader.
LLM as judge is an evaluation technique in which a language model scores the output of another system against a written rubric, standing in for a human reviewer. It makes open-ended qualities such as helpfulness, tone, or faithfulness measurable at scale, and is used where no exact reference answer exists.
A judge setup consists of a rubric, an input format, and an output schema. The rubric states the criteria and what each score level means. The input typically includes the original request, the response under review, and sometimes a reference answer or retrieved source documents. Constraining the judge to emit a structured verdict, such as a label plus a short justification, makes results parseable and gives a reviewer something to audit.
The technique has documented biases and the research literature is still active on how severe they are. Judges tend to prefer longer responses, favor text stylistically similar to their own output, and are sensitive to the order in which candidates are presented in pairwise comparisons. Position bias is usually mitigated by running both orderings and discarding disagreements. Verbosity bias is mitigated by rubrics that score specific required elements rather than overall impression.
Validating the judge is a separate task from using it. The standard approach is to have humans label a sample, then measure agreement between the judge and those labels on the same items. Agreement well below human-to-human agreement means the judge is not usable as a substitute yet. Reporting that agreement number alongside any judge-derived metric is what keeps the metric honest.
Cost and independence are practical concerns. Judging adds a model call per item, so scoring every output can rival the cost of producing it, and sampling is common. Using the same model family to both generate and judge introduces correlated blind spots, so teams sometimes use a different provider or a smaller specialized grader, accepting that no arrangement fully removes the shared-failure risk.
Key points
- A model scores another system's output against a written rubric
- Known biases include verbosity, position, and self-preference
- Must be validated against human labels before it is trusted
- Structured verdicts with justifications make results auditable
- Judging adds a model call per item, so sampling is common
In practice
To evaluate a research agent, a team writes a rubric with four binary checks: every claim cites a retrieved source, no cited source is fabricated, the answer addresses the actual question, and no forbidden speculation appears. A judge model returns the four booleans plus a one-line reason. Two reviewers label 150 of the same outputs by hand, and the judge agrees on 91 percent of items.