When developing and evaluating new Large Language Models (LLMs), human evaluation can be slow and expensive. To address this, researchers use LLM judges, which are powerful LLMs themselves, to evaluate the outputs of other models. These judge models are typically given the original prompt, the responses from one or more "candidate" LLMs, and a set of criteria or a rubric. They then "score" or "rank" the responses based on factors like helpfulness, accuracy, coherence, or safety. This automates part of the evaluation process, allowing for faster iteration and comparison of models.