LLM-as-judge
Open-ended answers have no single reference. A common shortcut is to ask a strong model to judge them, which works until its own biases leak into the scores.
Pairwise judging
Show the judge a question and two answers, A and B, and ask which is better. Repeat over many questions to get a win rate. It scales far better than human review, and on many tasks it agrees with people fairly often.
Position bias
Judges tend to favour the answer in a particular position, often the first. If model A is always shown first, part of its “win rate” is just the slot it sat in.
The lab simulates a judge with an adjustable position bias. The two models are truly equal unless you change the quality gap. Compare judging in one order with judging in both orders.
A worked example
Suppose A and B are truly equal but the judge favours the first position by 15 points. With A always shown first, A appears to win about 65% of the time. Judging each pair twice, A first and then B first, and counting only consistent verdicts as wins, brings the estimate back near 50%.
In practice
Strong models used as judges agree with people about as often as people agree with each other on many tasks, which makes them useful for fast iteration. Known biases include position, length (longer answers look better) and self-preference (a judge favouring its own model’s style). Write a clear rubric, swap positions, and check a sample of verdicts by hand.
Easy to confuse
- A judge score is a measurement with its own error, not ground truth.
- Pairwise judging (which is better) is usually more reliable than asking for an absolute score from 1 to 10.
A judge model prefers whichever answer it sees first. How do you control for that?
Swapping positions cancels a position bias; disagreements between the two orders become ties.
Which of these is a known bias of LLM judges?
Verbosity bias, position bias and self-preference (favouring outputs that resemble its own) are all documented.
What should you check before trusting an LLM judge on a new task?
A judge is a measuring instrument: calibrate it against human judgements on your own data before relying on it.
Read more in the books
- AI Engineering, Chip Huyen: Ch. 3, “Evaluation Methodology”: AI as a judge.
- AI Engineering, Chip Huyen: Ch. 4, “Evaluate AI Systems”: design your evaluation pipeline.