WTF is LLM-as-a-judge?

mainstreamFirst seen 9 Jun 2023

Using one LLM to grade or compare the outputs of another against a rubric.

Also called: llm as judge, model-graded eval, auto-grader

TL;DR

Humans are slow and expensive graders. An LLM judge can score thousands of outputs ("Is this answer correct and polite? 1–5") in minutes. It is not perfect, so you check that its scores agree with human judgments before trusting it.

How it works

Formats: single-answer grading against a rubric, pairwise comparison (A vs B), and reference-guided grading. Known biases include position bias (preferring the first option), verbosity bias (preferring longer answers) and self-preference. Mitigations: swap positions, use specific binary criteria rather than vague scales, ask for reasoning before the verdict, and measure agreement with a human-labelled sample.

Example

A simple judge prompt

You are grading a support reply.
Question: {question}
Reply: {reply}

Does the reply fully resolve the question without
inventing policy? Explain briefly, then answer PASS or FAIL.

Origin

Coined / introduced by: Zheng et al. (LMSYS)
When: 9 Jun 2023
Story: Named and studied in 'Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena'.

Related jargons

Sources

Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (arXiv)