LLM-as-Judge

LLM-as-judge is an eval method where a language model grades another model's outputs against a defined rubric, standing in for a human reviewer at a fraction of the cost and time — as long as the judge itself has been checked against real human judgments first.

Code graders can't assess open-ended quality — was this summary actually helpful, was this tone appropriate — and human review, the gold standard for exactly those judgments, doesn't scale to thousands of examples on every CI run. LLM-as-judge sits in the gap: it can evaluate nuance the way a human can, and it can run at the speed and volume of code.

How it works

A judge prompt typically supplies the rubric explicitly (what separates a 5 from a 3), one or two worked examples of each score level, and the response to grade, then asks for a score and a short justification — the justification matters because it makes disagreements auditable rather than a black-box number. The judge is usually a stronger, more careful model than the one being graded, since grading quality is itself bounded by the judge's own judgment.

The step teams most often skip, to their cost: validating the judge. Before trusting an LLM judge's scores, a team scores a sample against real human labels and checks agreement — not just raw percentage agreement, which is inflated by how often both raters simply agree by chance, but a chance-corrected measure like Cohen's kappa. Known judge biases compound the need for this check: position bias (favoring whichever answer appears first), verbosity bias (rewarding longer answers regardless of quality), and self-preference bias (a model judge favoring outputs in its own family's style).

Why it matters

LLM-as-judge is what makes evaluating subjective qualities — helpfulness, groundedness, tone — practical at the scale a CI-gated eval suite needs. Skip the validation step and a team ships an unvalidated number that feels rigorous and measures nothing but the judge's own quirks; do the validation and an LLM judge becomes a genuinely trustworthy proxy for the human review it's standing in for.

A worked example

A team validates a pass/fail judge against 300 human-labeled outputs (200 humans marked pass, 100 marked fail). The judge agrees with the human label on 180 of the 200 passes and 81 of the 100 fails — 261 of 300 correct, 87 percent raw agreement. But chance agreement alone, given each rater's marginal pass/fail rates, would already produce about 55.4 percent agreement — so Cohen's kappa, which corrects for that, works out to (0.87 − 0.554) / (1 − 0.554) ≈ 0.71. A kappa around 0.7 is generally read as "substantial" agreement — good enough to trust the judge for this task, but not so perfect that spot-checks stop.

How Miatz teaches it

Learners build a small LLM judge and validate it themselves against hand-labeled examples, computing raw agreement and Cohen's kappa by hand on a confusion matrix before trusting the judge on a single ungraded example — so "the judge agrees with humans" is a computed number, not an assumption.