Judge Model

When a computer program writes texts, someone has to check whether the answers are good. For a long time, humans did this: they read two answers and said which one was better. That is expensive and takes a long time. A judge model takes over this role. It is itself a text program, but it is not tasked with writing an answer, but rather with grading someone else’s answer. The name comes from the English word for a legal judge.

Why evaluating is more expensive than answering

A modern language program is measured against millions of sample answers. If one were to check each of these by hand, one would need entire halls full of test readers. A judge model does the same work in minutes and for a fraction of the cost. That is precisely why it has become a standard tool in recent years.

The second reason is development itself. Companies change their models almost daily. After every change, one needs to know whether the result has gotten better or worse. Human tests would be far too slow for this. Automatic evaluation makes it possible to build in a change in the morning and know by midday whether it helped.

However, there is a risk one needs to be aware of. When one AI grades another, an error can double. If the judge model considers something false to be correct, this mistake gets passed on during training. That is why companies regularly check whether the judge model’s verdicts match those of real humans.

What is in the evaluation instruction

A judge model is usually a perfectly ordinary language program. The difference lies in the instruction it receives. This instruction contains the original question, the answer to be checked, and a list of criteria. Typical points include: Is the answer factually correct? Does it answer the question? Is it understandable? In the end, the model is supposed to output a grade from one to five or write a short justification.

Often, it is not a single answer that gets graded, but two that get compared. The model receives answer A and answer B and is supposed to say which one is better. Comparisons are more reliable than absolute grades, because a benchmark is missing when one sees only a single answer. From many such comparisons, a ranking can then be formed, similar to chess ratings.

Judge models have typical weaknesses that are well studied. They prefer longer answers, even if these contain no more content. They often choose the answer that comes first. And they like texts that sound like their own. A simple trick helps against these biases: one asks the same question twice, swapping the order of the answers each time. Only if the verdict comes out the same both times does it count.

Judge models in rankings and products

Judge models are most visible in the major rankings for AI systems. When the news reports that a new model achieved certain scores in a test, a judge model is often behind it. Some tests combine both: part of the questions is evaluated automatically, a small part is checked by humans.

Such evaluators also work within products, often invisibly. A customer service chatbot can have its own answer checked by a second model before sending it. If the evaluation turns out poorly, the answer is discarded or forwarded to a human. Companies that use AI in customer contact almost always build in such control loops.

A common misconception is that a judge model is especially objective because it is a machine. That is not true. It judges based on what it saw during training and adopts its preferences. It is a fast and cheap tool, but not a neutral referee.