LLM Judges: Usage Framework and 9-Step Method

Evaluating one model by another model is not something that can be improvised. LLM judges must be treated as true AI systems, equipped with operational safeguards, rigorous calibration, and long-term monitoring. Recommendations specify when to use them, how to configure them, and in which cases to forgo their use.

Monitor the judge and frame cases where to forgo

LLM judges require specific supervision: it is advisable to regularly evaluate the judge itself, integrate it into CI or regression gates, conduct comparisons before publication, and ensure monitoring in production. Their use is not recommended when deterministic checks are possible, when a low-cost ground truth is available, or when specialized knowledge is required. These practices are based on the principle that the LLM judge constitutes a complete machine learning system, which must be built, calibrated, versioned, and continuously validated.

Choose the appropriate evaluation mode

An LLM judge is a model tasked with evaluating the output of another model according to defined criteria or rubrics. Some teams sometimes prefer this method over human review or verification of chains or code. Three types of judgment are frequently used: evaluation on a single output, comparison of two results, and assessment based on references. The comparison of two results is generally recognized as the most appropriate default option.

Calibrate and reduce biases with 9 concrete steps

For LLM judges to effectively reflect human preferences, they need to be specifically adjusted to the task, taking into account biases such as position bias or individual preferences. A nine-phase protocol is suggested: rely on real failure cases, limit the number of criteria, favor binary choices, require prior reasoning before any decision, include benchmark examples, break down subjective judgments into more elementary decisions, establish reference points, calibrate with human annotations using appropriate agreement metrics, and involve multiple judges when the stakes are high. In the absence of these measures, LLM judges may generate evaluations that lack reliability.