What is LLM-as-a-Judge and how is it used for automated evaluations?

Short answer: LLM-as-a-Judge uses a powerful LLM (e.g. GPT-4o or Claude 3.5 Sonnet) with structured rubrics to automatically grade, score, or compare candidate model outputs for qualitative criteria like accuracy, helpfulness, tone, safety, and format adherence.

Detail

Traditional NLP metrics like BLEU or ROUGE perform exact n-gram matching, failing to assess semantic correctness or reasoning quality. Human evaluation is slow and unscalable. LLM-as-a-Judge bridges this gap.

Common Judge Setup Configurations

1. Single-Answer Scoring: The judge rates a generated answer on a Likert scale (1-5 or 1-10) against an explicit grading rubric.
2. Pairwise Comparison (A/B Testing): The judge receives two outputs (Model A vs Model B) for a given prompt and selects the winning response (or tie).
3. Reference-Based Evaluation: The judge compares the generated response against a human-written ground truth reference answer.

[Prompt + Candidate Output + Scoring Rubric] ──► [Judge LLM] ──► Structured Grade + Rationale

Example

Python prompt template for a single-answer judge evaluation:

judge_prompt_template = """You are an expert evaluator scoring an AI assistant response.

[User Question]
{question}

[AI Response]
{response}

[Scoring Rubric]
Score 1: Completely inaccurate or unhelpful.
Score 3: Partially correct but missing key detail.
Score 5: Exceptionally clear, fully accurate, and directly answers the question.

Provide your evaluation in valid JSON with fields "explanation" and "score" (integer 1-5)."""

Interview tips

• Highlight position bias (in pairwise comparison, judges often favor whichever response appears first as Candidate A) and verbosity bias (judges favor longer outputs).
• Explain swap evaluation: running pairwise comparisons twice with candidate order swapped to eliminate position bias.

⬅ Back to Evaluation and Testing · All topics