TL;DR: LLM-as-judge means using a language model to grade another model's outputs against a written rubric. It makes evaluation cheap enough to run on thousands of cases instead of the dozens a human reviewer can handle. The catch: judge models have known, measurable biases, so a judge you have not checked against human labels is just a second opinion of unknown quality.
How it works
You hand the judge three things: the original input, the output being graded, and a rubric that spells out what to reward and what to penalize. The judge returns a score - a number on a scale, a pass/fail verdict, or a choice between two candidate answers - usually with a short justification. Grading a support bot's reply for "resolved the actual question, cited only real policies, stayed polite" is a typical rubric; so is comparing two drafts and picking the more faithful one.
The appeal is scale. Human review is the gold standard for quality, but nobody hand-grades ten thousand outputs per eval run. A judge model does it in minutes for a few dollars, which is what makes nightly regression runs and large golden datasets practical at all. Judges are also how fuzzy criteria - helpfulness, tone, whether an answer contains a hallucination - become numbers you can track over time.
The biases are real and documented. Position bias: in pairwise comparisons, judges favor whichever answer appears first, so good harnesses grade both orderings and average. Verbosity bias: longer answers score higher even when the extra words add nothing. Self-preference: a model rates output from its own family more generously, which is one reason to judge with a different model than the one being tested. Rubrics that demand a justification before the score, and scales with few, well-defined points, all reduce the noise.
The discipline that makes it trustworthy is calibration: take a sample of judge verdicts, have humans grade the same cases blind, and measure agreement. High agreement means you can lean on the judge; low agreement means the rubric needs rewriting, not that the numbers are close enough. Skipping this step is how teams end up optimizing for a judge's taste instead of their users' - the same trap public benchmarks fall into when models are tuned to the test.
Where it sits in the AI stack
A judge is the scoring stage of an eval pipeline. Outputs from the system under test flow through it, and its verdicts become the metrics everything downstream reads:
Key tools and implementations
-
DeepEval
Open-source eval framework with ready-made judge metrics, including the G-Eval rubric pattern.
-
Ragas
Judge-based metrics specialized for RAG - faithfulness and relevance scored by a grading model.
-
Braintrust autoevals
A library of prebuilt judge scorers - factuality, closeness, moderation - usable in any harness.
-
Arize Phoenix
Open-source observability with built-in judge templates for scoring traces at scale.
Related entries
- Evals (AI evaluations) Structured tests that score an AI system's outputs so teams can measure quality and catch regressions.
- Prompt testing Running prompt changes against a fixed set of test cases so an edit cannot silently degrade quality.
- Hallucination When a model states false information fluently and confidently because it generates plausible text, not verified facts.
- Model benchmarks Standardized tests that score model capabilities, useful for rough comparison but easy to overfit and game.