LLM-as-Judge, With Caution
Use a model to grade answers at scale only after measuring it against trained people, and know the biases to watch for.
Objective
By the end of this lesson, you can write a judge rubric, measure an LLM judge against human labels, and decide where a judge is and is not trustworthy.
Why teams use model judges
Some answers cannot be checked with simple code. "Is this reply kind, accurate, and at a 6th-grade reading level?" needs judgment. People are the best graders, but grading thousands of answers by hand is slow.
LLM-as-judge means using a language model to grade another model's output against a rubric. It scales. It also has known weaknesses.
Known judge biases
Research and practice have documented several:
- Leniency: judges tend to pass answers that sound confident and fluent, even when wrong.
- Position bias: when comparing two answers, judges can favor the one shown first.
- Length bias: longer answers often get higher scores.
- Self-preference: a model may favor answers written in its own style.
- Blind spots: a judge cannot catch errors it does not know are errors, such as an out-of-date local rule.
Calibrate before you trust
Never use a judge until you have compared it with people:
- A trained person grades 50 to 100 answers using the written rubric.
- The judge grades the same answers.
- Measure overall agreement, and agreement on failures: of the answers people failed, how many did the judge fail?
- Read every disagreement. Fix the rubric or the judge prompt.
- Re-check on fresh samples regularly.
The first run downloads Python into your browser (several MB).
Write a judge rubric
Good judge prompts are narrow. Grade one thing at a time, ask for a reason before the verdict, and give examples of pass and fail.
Judge rubric builder
Write a judge prompt that checks one thing: whether an answer is fully supported by the given source passages.
Your prompt
Act as a strict grader who checks answers against source passages.
Where judges do not belong
- As the only check on harmful cases. People grade those.
- On questions where the judge lacks the facts, such as local rules not in the passages.
- As a launch gate without calibration.
- Grading its own model's outputs without human spot checks.
Quiz: Judge check
Go deeper: Machine Learning, Minus the Hype: A Practical Playbook to Ship Useful Models
Discussion
Ask. Answer. Build together.
Loading…