LLM-as-Judge
A judge is a second LLM call that scores the first one. It is the cheapest way to scale a rubric past what one human reviewer can read. It is also the technique most teams get wrong, because they treat it as a vibes machine instead of a classifier with measurable error rates against a human gold set.
The nine chapters here are opinionated. We use binary pass-fail with a written critique, not 1-to-5 Likert. We pick the judge bigger than the generator. We measure TPR and TNR, not agreement percentage. We swap positions on pairwise calls. And we do not let a judge near hard reasoning, math, or code correctness without an executable verifier behind it. On the techniques side, the condensed references are LLM-as-Judge, rubric design, and pairwise comparison and Elo.
Chapters:
- Introduction. What an LLM judge is, when it works, when it does not, and the three biases everyone hits.
- The four-part judge prompt. A template you can copy: role and definitions, criteria, few-shot examples from your train set, output format.
- Calibrating your judge against humans. Train, dev, test split, TPR and TNR over agreement percent, the 100-example minimum, iteration loop.
- Biases you must defuse. Position, verbosity, self-preference, format sensitivity, sycophancy. Each with the empirical evidence and the mitigation.
- Pairwise vs direct scoring. When relative comparison beats absolute scores, and the decision rule for choosing between them.
- Distilled judges. Small specialized judge models, when they pay for themselves, and the cost math against frontier judges.
- Where LLM judges fail. The JudgeBench result that GPT-4o is roughly random on hard correctness judgments, and what that means for your pipeline.
- Decision models as judges. Jev and the typed-verdict judge: the cost math for scoring every trace, where two independent studies say it breaks, and why a guardrail cannot also be its own eval.
- OpenAI's Decisions API vs Jev. The second decision model, from OpenAI: how its typed answers differ from Jev's, why its score field is an average and not a level, and a bake-off to run on your own gold set.
The honest take page is the one you should not skip. Strong judges agree with humans well above 80 percent on MT-Bench style preference data. On JudgeBench style hard correctness, GPT-4o scores 56.6 percent, barely above a coin flip. Reasoning models close much of that gap (o3-mini at high effort scores 80.9 percent), but not enough to replace an executable check when one exists. The gap between preference data and hard correctness matters; you cannot generalize one to the other.