Evaluation (Eval)
An evaluation, or eval for short, is a standardized test used to measure the performance of an AI system. It consists of tasks with known solutions and a rule for how the answers are scored.
When a company claims its new computer program is especially good, that claim has to be verifiable. This is exactly what evaluations are for, usually called "evals" in industry jargon. An eval is a collection of tasks for which the correct solution is already known. The program is made to work through all the tasks, and one counts how many it solves correctly. The result is a number, often a percentage, and numbers can be compared. Without such tests, the question "is this system better than that one?" would remain mere assertion.
What's at stake when no one measures
Modern AI systems are not created through programming in the classical sense. No one writes down rules for the system to follow. Instead, it learns from vast amounts of example text, and what exactly it learns in the process cannot be read off directly by anyone. So even the developers themselves don't know for certain what their system can do. Measuring is the only way to find out.
For companies and investors, a great deal hinges on these numbers. When a manufacturer reports that its model has improved on a well-known test from 70 to 85 percent, that moves stock prices. Evals also matter in procurement: a hospital purchasing AI for reading X-rays wants documented accuracy rates, not marketing promises.
Evals also serve safety purposes. Systems are specifically tested for whether they provide instructions for weapons, insult people, or state fabricated facts. Such tests are called red-teaming when humans deliberately try to provoke the system into missteps. The European AI regulatory framework also requires documented testing for high-risk applications.
Task collection, scoring rule, leaderboard
An eval consists of two parts: a dataset of tasks and a scoring procedure. For multiple-choice questions, scoring is simple, since only one answer is correct. The well-known test MMLU works this way and contains around 16,000 questions from subjects ranging from law to physics. For coding tasks, the generated code is simply run and checked to see whether it produces the correct output.
Things get difficult with open-ended tasks such as summaries or translations. There, no single correct solution exists. A common workaround: have a second, powerful AI model grade the answers. This is called LLM-as-a-judge and is cheaper than human evaluators, but also more error-prone. An alternative is leaderboards, where real users compare two anonymous answers and choose the better one.
The industry's biggest problem is called contamination. The test tasks are often freely available on the internet, and the training data comes from the internet as well. A model may therefore have already seen the solutions. In that case, the eval measures not ability but memory. A comparable situation would be a class test whose questions and answers had already circulated in the class chat beforehand. This is why new, secretly held test sets are constantly being created.
Evals in product announcements and everyday company life
Every major model launch comes with bar charts. Names like MMLU, GPQA, or SWE-bench show up there, each with percentage figures next to those of the competition. Anyone reading such charts should remain skeptical. Manufacturers like to pick the tests on which they perform well and leave out the ones that don't fit.
In companies that actually deploy AI, evals look less spectacular. An insurer that has customer inquiries sorted automatically collects a few hundred real inquiries along with the correct classification. This internal eval runs after every change to the system. If the accuracy rate drops, the change is rolled back. Such in-house test sets are usually more informative than public leaderboards, because they reflect the actual work being done.
A dedicated market has since formed around evals. Vendors sell software for testing AI applications, and the term regularly appears in job postings. Evaluation should not be confused with monitoring: evals check against known solutions before deployment, while monitoring observes the running system in operation.
