Using one language model to grade another is cheap and scalable. It is also a measurement instrument that needs validation like any other.
Why people use it
Human evaluation is slow and expensive. A strong model can score thousands of responses quickly, often agreeing closely with people on straightforward criteria.
Known biases
- Position bias: the judge may prefer whichever answer appears first.
- Length and verbosity bias: longer answers can win regardless of quality.
- Self-preference: a model may favor outputs similar to its own style.
- Format sensitivity: small changes to the rubric can change scores.
Validate before trusting
- Collect a sample of human labels.
- Measure agreement between the judge and humans, and compare it with agreement among humans.
- Check calibration: when the judge reports high confidence, is it right more often?
- Test swaps, such as reversing answer order, to detect position bias.
Reduce errors
Randomize order, use clear rubrics, average multiple judgments, and consider multiple judge models. Where a ground-truth check exists, such as unit tests or exact answers, prefer it.
Combining judge and human data
Statistical methods can use a small set of human labels to correct for judge error and give valid confidence intervals on the final estimate.
The rule of thumb
A judge model is evidence about quality, not quality itself. Report its validation alongside every result that depends on it.
Continue with the book
Explore the full treatment, worked examples, and exercises in Statistical Foundations for LLM Development and Evaluation.
Adapted from Statistical Foundations for LLM Development and Evaluation, published by Choral Software. Article text and original examples © 2026 Choral Software. All rights reserved.