Human review is the gold standard for judging AI output, but it is slow and expensive. Simple automatic metrics — exact match, word overlap — miss most of what makes an answer good. LLM-as-a-judge uses a capable language model to grade outputs according to instructions, offering a scalable middle ground. It is now a standard part of AI evaluation, from research leaderboards to product testing.

The research basis

The approach was studied systematically in the 2023 paper Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena by Lianmin Zheng and colleagues. They found that a strong judge model agreed with human preferences at a level comparable to the agreement between humans themselves — over 80% in their setting — while identifying several systematic biases. The G-Eval paper (2023) showed that asking the judge to follow evaluation steps and a form improved alignment with human ratings for text generation tasks.

Three common setups

  1. Single-answer grading: the judge scores one answer on a scale (for example 1–5) against a rubric.
  2. Pairwise comparison: the judge sees two answers to the same question and picks the better one, or declares a tie.
  3. Reference-guided grading: the judge compares an answer with a reference answer or required facts.

Pairwise comparison is often more reliable for subjective quality; reference-guided grading is better when there is a correct answer.

Known biases

  • Position bias: in pairwise comparisons, judges may favor the answer shown first (or second).
  • Verbosity bias: longer, more detailed-looking answers tend to be rated higher even when they are not better.
  • Self-preference: a model may favor outputs that resemble its own style or come from the same model family.
  • Limited ability: a judge cannot reliably grade math or code it cannot solve itself.
  • Leniency: judges often give high scores unless instructed to be strict.

Making judges reliable

  • Write a specific rubric. “Is the answer good?” is too vague. Instead: “Score 1 if the answer contradicts the policy document; 3 if correct but incomplete; 5 if correct, complete and cites the relevant section.”
  • Ask for reasoning before the score so the judgment is explained and auditable.
  • Swap positions in pairwise comparisons and count only consistent judgments.
  • Grade one criterion at a time rather than everything at once.
  • Provide references where a correct answer exists.
  • Use code for what code can check: formats, required fields, exact values.
  • Calibrate against humans: have people label a sample, measure agreement with the judge, and refine the rubric until agreement is acceptable.
  • Use a strong judge, ideally from a different model family than the system being evaluated.

Where it fits

LLM-as-a-judge is ideal for fast iteration: comparing prompt versions, regression testing and monitoring production samples. Keep human review for final decisions, high-stakes domains and periodic calibration. Together with code-based checks, it forms the backbone of a practical evaluation set.

Common mistakes to avoid

  • Trusting absolute scores across judges. Different judge models score on different implicit scales; compare systems with the same judge and rubric.
  • Never checking agreement with humans. Without calibration, you don’t know whether the judge measures what you care about.
  • Rewarding length. Instruct judges to ignore length and penalize unnecessary verbosity.
  • Grading tasks beyond the judge’s ability. Use tests, execution or experts for hard math and code.

Calibrating your judge

Have humans score a sample of outputs, compare with the judge’s scores, and adjust the rubric until agreement is acceptable before relying on automated judging.

Related: building an evaluation set.

Key takeaways

  • LLM-as-a-judge uses a strong model to grade outputs with a rubric.
  • Research found strong judges can agree with humans about as often as humans agree with each other.
  • Known biases include position, verbosity and self-preference.
  • Specific rubrics, reasoning-first scoring, position swapping and human calibration improve reliability.
  • Combine AI judges with code checks and periodic human review.

Frequently asked questions

Is LLM-as-a-judge accurate enough?

For many quality criteria, a well-designed judge is accurate enough for iteration and regression testing. Calibrate it against human labels for your specific task.

Which model should be the judge?

Use a strong model, ideally from a different family than the one you evaluate, to reduce self-preference.

Does it cost a lot?

It costs model calls for every graded output. Use code checks where possible and sample production traffic rather than grading everything.