Cheap, fast, biased
Using an LLM to grade other LLM outputs is the cheapest way to scale evaluation. It's also the most easily fooled. Used carefully, it accelerates work; used carelessly, it produces a comforting number that doesn't predict real quality.
Where LLM judges work
- Format checking — is this JSON valid? does the response include all required fields?
- Pairwise comparison with a clear rubric.
- Categorical classification (refund vs complaint vs feature request).
- Hallucination detection against a reference ("does the response contradict the source?").
Where LLM judges fail
- Numerical scoring on subjective qualities (1–5 helpfulness scores are noisy).
- Anything where the judge has the same blind spots as the model under test.
- Adversarial situations — a judge model can be prompt-injected too.
Calibrate the judge
Spot-check 20% of judge decisions with a human. If agreement rate is below 85%, the judge isn't usable for that task; tighten the rubric or do more human eval.