EvaluationAdvanced

LLM-as-judge pitfalls

“What are the pitfalls of using an LLM as a judge, and how do you mitigate them?”

What this tests

  • Knowledge of concrete biases: position, length, self-preference, verbosity
  • Calibration against human labels
  • Rubric design and pairwise versus absolute scoring
  • Knowing when not to use a judge

Answers by level

Read the beginner answer first and notice what is missing.

Known biases: position bias (prefers the first or second option in pairwise comparisons), length and verbosity bias (longer looks better), self-preference (rates outputs from its own model family higher), sycophancy toward confident tone, and insensitivity to factual errors when the text reads well. Rubric-free 1–10 scores are also poorly calibrated and drift between runs. See LLM-as-Judge.

Mitigations: a specific rubric with observable criteria per score level; pairwise judgements with order swapped and results averaged; a different model family than the one under test; provide the reference answer or the source documents so the judge checks facts rather than fluency; and calibrate by measuring agreement with a human-labeled set, reporting it, and re-checking when the judge model or prompt changes.

And use the judge only where a deterministic check cannot work. Format, presence of ids, forbidden actions, numeric results are all coded checks. See Deterministic Evaluators.

Green flags · Red flags

Green flags
  • Names position, length, self-preference, fluency bias
  • Order swapping, different model family, reference-grounded judging
  • Binary rubric criteria over holistic scores
  • Measures judge agreement with human labels and tracks it over time
  • Versioned judge prompts and models
  • Uses deterministic checks wherever possible
Red flags
  • Trusts a strong model with a vague prompt
  • No calibration against humans
  • Holistic 1–10 scores as the primary metric
  • Unaware of injection into judged content

Follow-up questions

F1
How would you detect position bias in your judge?
F2
Judge agreement with humans is 65%. What now?

Practical scenario

Your RAG eval uses the same model as judge and generator with the prompt "rate the answer's quality from 1 to 10". Scores are consistently 8–9 while users report frequent factual errors. Diagnose the judge setup, redesign it, and describe how you would prove the new judge is trustworthy.

Related concepts · Learn this topic