LLM-as-Judge · bigonotes

Some qualities cannot be checked in code: whether an answer is helpful, faithful to its sources, or in the right tone. For these you use a second model to grade the output. This is called LLM-as-judge, and it is powerful as long as you remember the judge can be wrong too.

How it works#

You write a rubric that defines what a good answer looks like. Then you send the question, the answer and the rubric to a judge model, and it returns a grade.

The question, the answer and a rubric go into a judge model, which returns pass or fail for each criterion

Pass or fail per criterion#

Ask for pass or fail on each criterion, not a score from 1 to 10. A score of 6 tells you nothing about what to fix. A fail on "the answer is grounded in the retrieved sources" tells you exactly where the problem is.

Grade the answer against each criterion. Reply PASS or FAIL for each,
with one sentence of reason.

1. Grounded: every claim is supported by the provided sources.
2. Complete: the answer addresses every part of the question.
3. Tone: the answer is polite and plain, with no jargon.

Pass or fail also makes judges more consistent, because a yes-or-no decision is easier to make the same way twice than a point on a scale.

Three biases#

Judge models have known biases. Each one can make a worse answer look better.

Bias	What happens	Mitigation
Position	When comparing two answers, the judge prefers the first one	Run both orders and compare
Verbosity	Longer answers get rated higher even when they are not better	Put conciseness in the rubric
Self-preference	A model prefers outputs that look like its own style	Use a judge from a different model family

Validate the judge#

Because of these biases, you check the judge before you trust it:

Take 50 to 100 cases.
Grade them yourself.
Run the judge on the same cases.
Measure how often the judge agrees with you.

If the judge disagrees too often on cases where you are confident, the rubric needs work, or you need a different judge. Validate again whenever you change the rubric or the judge model.

A judge you have not validated is just another unchecked model output.

Short line#

An LLM judge grades fuzzy qualities against a rubric, best as pass or fail per criterion. Judges have position, verbosity and self-preference biases, so validate one against your own grades on 50 to 100 cases before trusting it.

Key takeaways#

Use a judge model for qualities code cannot check: helpfulness, faithfulness, tone.
Send the question, the answer and a rubric. Get a grade back.
Pass or fail per criterion tells you what to fix. A 1 to 10 score does not.
Biases: position (prefers first), verbosity (prefers longer), self-preference (prefers its own style).
Validate the judge on 50 to 100 cases you graded yourself before trusting it.

Core idea#

A judge model grades qualities code cannot check, such as helpfulness, faithfulness and tone, against a rubric.

Key points#

Send the question, the answer and the rubric. Get a grade back.
Ask for pass or fail per criterion, not a 1 to 10 score.
A fail on "grounded in sources" shows what to fix. A score of 6 does not.
Validate the judge: grade 50 to 100 cases yourself, then measure agreement.
Low agreement on confident cases: fix the rubric or change the judge.

Three biases#

Bias	What happens	Mitigation
Position	Prefers the first of two answers	Run both orders
Verbosity	Rates longer answers higher	Put conciseness in the rubric
Self-preference	Prefers its own style	Judge from a different model family

Pitfalls#

Do not trust a judge you have not validated.
Do not use 1 to 10 scores. They hide the cause.

Short line#

Grade fuzzy qualities with a rubric, pass or fail per criterion. Validate the judge, because it has position, verbosity and self-preference biases.