LLM-as-judge
Using a language model to grade another model's output against a rubric.
LLM-as-judge scales evaluation where exact-match scoring is impossible (open-ended text, multi-step traces). A judge model receives the task, the output and a rubric and returns a score or pass/fail. Risks are judge bias (position, verbosity, self-preference) and drift; calibration against a human-labelled sample is standard practice. Used by 53.3% of teams that run evals.