LLM-as-a-Judge

Programs like ChatGPT write texts that can be good or bad in very different ways. There is rarely a single correct answer, which makes evaluation difficult. In the past, humans read two answers and decided which one was better. That is expensive and takes a long time. With LLM-as-a-Judge, another text program takes on this role: it reads the answers, compares them, and explains which one it considers better. LLM stands for Large Language Model, i.e. a large language model trained on huge amounts of text.

Why companies replace human reviewers

Anyone developing an AI product must test constantly. Every small change to the model or the instructions can improve or worsen quality. Without measurement, you wouldn’t even notice. Human reviewers, however, quickly cost several euros per evaluated answer. With ten thousand test cases, that becomes unaffordable.

A model acting as judge, on the other hand, costs fractions of a cent per case. It works around the clock and delivers results within minutes. This allows a team to check several times a day whether a change has helped. This fast feedback is the real reason for the popularity of the method.

Still, it’s important to remember: the goal remains the satisfaction of real humans. A model’s judgment is only a proxy measure. Good teams therefore regularly check whether the model’s judgments still align with those of human testers. If they diverge significantly, the entire test setup becomes worthless.

From evaluation sheet to verdict

In practice, the judging model receives a precise instruction. It contains the question, the answer to be evaluated, and the criteria. Typical criteria are correctness, completeness, comprehensibility, and politeness. Often a sample solution is also provided. The model is then supposed to give a grade from 1 to 5 and briefly justify it.

Direct comparison is also common: the model sees two answers and says which one is better. Such pairwise comparisons are more reliable than absolute grades, because a comparison is easier than an abstract score. From many pairwise comparisons, a ranking can be calculated, similar to chess ratings.

The method has known weaknesses. Models often prefer the answer that comes first, which is why the order is reversed and both runs are evaluated. They also prefer long answers with many bullet points, even if these add nothing extra. And a model recognizes its own style and rates it more favorably. Anyone who has GPT-4 both answer and judge is thus partly measuring itself.

Rankings, chatbots, and audit reports

The method is most visible in public rankings for language models. There, models compete against each other, and part of the judgments come from other models. When the news reports that a new model has surpassed another, such an automatic evaluation is often behind it. It’s almost always worth checking who did the judging.

This also happens in the background at companies. An insurer with a customer service chatbot has samples of conversations automatically graded. A judging model then checks whether the answer matched the policy and whether it was friendly. Conspicuous cases are passed on to a human. The same technology is found in many tools sold as AI monitoring or evaluation.

Related is the idea of guardrail models, which check answers for dangerous or forbidden content before they are sent. The difference lies in the purpose: a guardrail blocks, a judge grades. But both show the same trend. AI systems are increasingly controlling other AI systems, and humans only check samples.