Evals and benchmarks

In Model differences and harnesses, you saw that there are many models and that each one has strengths. The obvious next question is: who decides which one is "best"? The answer is a pile of tests, a few public scoreboards, and a lot of marketing. Once you know how the tests work, a launch chart reads like a sales flyer: interesting, not proof.

An eval is a test plus a scoring rule

The word you will hear in AI circles is eval, short for evaluation. An eval is two things bolted together: a set of test questions, and a rule for deciding whether each answer is right. A spelling quiz with an answer key is an eval. So is a driving test.

The scoring rule matters as much as the questions. Some rules are easy for a computer to apply: did the model pick choice C, and was C correct? Some rules need a program to run: did the code the model wrote actually work? And some rules need a person: which of these two answers is more helpful? Easy rules make tests cheap and repeatable. Human rules capture things a checklist cannot, like tone and usefulness, but they are slow and people disagree.

A benchmark is an eval that has been published so everyone can run the same one and compare. When a lab says its new model "scored higher on X," X is almost always a benchmark.

Three benchmarks you will see named

There are hundreds of benchmarks. These three show up constantly, and each one tests something different.

• MMLU (Massive Multitask Language Understanding). Multiple choice questions across 57 subjects, from elementary math and US history to computer science and law, each with four possible answers. It was built in 2020 to measure broad school-style knowledge.
• GPQA (a Graduate-level, Google-proof Q&A test). 448 multiple choice questions in biology, physics, and chemistry, written by experts in those fields. "Google-proof" means smart people outside the field could not look the answers up: in the paper, skilled non-experts with unlimited web access, spending over half an hour on average, got only about a third right, while PhD experts in the field got about two thirds.
• SWE-bench (software engineering). 2,294 real problems pulled from public bug reports in 12 popular open source projects written in the Python programming language. The model gets the project's code and a description of the bug, and has to change the code to fix it. The scoring rule is strict and automatic: the project's own tests are run, and the fix only counts if the tests that used to fail now pass and nothing else broke.

See the pattern? A model can ace a science exam and still write a clumsy text to a customer. A model can fix bugs in Python and still give you a bad hiking plan. A score is real information about that test, not about your job.

Leaderboards that rank by human votes

Fixed tests have an obvious gap: they cannot measure "is this answer actually good to a person?" So some leaderboards skip the answer key and ask people instead. A leaderboard is just a public ranking that updates as new scores come in.

The best known one is Arena, at arena.ai. As of October 2026 it goes by Arena; it was called LMArena before that and started out as Chatbot Arena, and the old lmarena.ai address now forwards to the new one. The idea is simple. You type a question, two anonymous models answer side by side, and you vote for the better answer. Only after you vote do you see which models they were. Arena's own FAQ says only votes cast while the models are still anonymous count toward the rankings. All those head-to-head votes are turned into a rating, much like a chess rating: beat a strong opponent and your number goes up a lot, beat a weak one and it barely moves.

Blind voting fixes a real problem. People are swayed by brand names, and hiding the names removes that. But it brings new problems. The people voting are mostly curious tech fans asking whatever they felt like asking, not a landscaper pricing a job in Washington. Voters can reward answers that are longer or nicer looking, even when a shorter answer is more correct. A human vote measures what a crowd preferred in the moment, which is useful, and also not the same thing as "right."

Why scores mislead

Three problems show up again and again. None of them mean benchmarks are useless. They mean a single number deserves a second look.

Test questions leak into training. As How AI is trained covers, models learn from enormous scrapes of the public internet. Popular benchmarks are on the public internet. So the questions, and sometimes the answers, can end up in the training data. This is called contamination: the model has seen the test before, the way a student who found last year's exam might score well without understanding the material. Researchers tested this in 2024 by writing a fresh set of grade school math problems in the same style as a popular math benchmark. Some models lost as much as 8 percent of their accuracy on the new problems, a hint that they had partly memorized the original. The same study also found that many of the strongest models held up fine, so contamination is a real risk, not a blanket verdict.

Benchmarks fill up. When a test is new, scores are spread out and you can tell models apart. As models improve, they crowd toward the top, and the test stops separating them. This is called saturation. The team that built a harder version of MMLU in 2024 said exactly this: performance on the original had begun to plateau, making it hard to tell models apart. Their fix was harder questions and ten answer choices instead of four. That cycle is why the benchmark names on launch charts keep changing.

Labs show their best side. A lab chooses which benchmarks to put on its launch chart, which settings to run them with, and sometimes how many tries to allow. Nobody puts the test they lost on the slide. Human-vote leaderboards are not immune either. A 2025 study called "The Leaderboard Illusion" found that some providers tested many private versions of a model on Chatbot Arena before launch and could choose which results to make public, which tilts the rankings toward whoever can afford the most tries.

There is a quieter fourth issue. Models are a little random from one run to the next, as Temperature and sampling explains, so a single run of a test is a single roll of the dice. Good evals run each question several times. Launch charts rarely say how many times they did.

Your own eval beats theirs

Here is the good news. Writing a useful eval for yourself is easy, and it answers the question no leaderboard can: which tool is better for me? You need three things, and they are the same two parts every eval has, plus a judge.

• Real questions. Not trivia, and not trick questions designed to catch the AI out. The actual things you ask, or wish you could ask. A mix of easy and hard, short and long.
• A scoring rule you decide in advance. Write it down before you read any answers, so you are not swayed by the one that sounds most confident. "Correct, short, and no made-up facts" is a perfectly good rule.
• You as the judge. You know your work. For questions where you are not sure of the right answer, check against a source you trust, the same habit Trust and limits teaches for anything that matters.

Try the idea in miniature below. Five questions, two answers each, labeled only A and B so you judge the words and not the brand, just like a blind leaderboard. There is no answer key. Notice how your tally depends on what you care about: some answers are safer, some are more useful, some are shorter. That is the point. Your scoring rule is part of the test.

Try this yourself

Build a ten question personal eval. You only need two chat apps you already use (free versions are fine) and a notes app or a sheet of paper. No installs, no accounts beyond the ones you have.

1. Write down ten real questions from your own life or work. Use the template below to get a good mix.
2. Write your scoring rule before you ask anything. Keep it to three marks: good as is, good after a small fix, or not usable.
3. Start a fresh chat in each app and paste the same question, word for word. One question per chat, so earlier answers do not color later ones.
4. Score each answer by your rule. For any question where you are unsure, ask the same question a second time in a new chat and see if the answer holds.
5. Count up. Then look past the total at the pattern. Which tool won the questions that matter most to you?

My personal eval (fill in your own questions)

1. A fact I already know the answer to: ____
2. A how-to question from my work: ____
3. A short message I actually need to write: ____
4. A question about my town or area: ____
5. Something I would normally ask a friend for advice on: ____
6. A question with numbers in it (a price, a measurement, a date): ____
7. A long document or email I want summarized (paste it in): ____
8. A question where the honest answer is "it depends": ____
9. A question about something that happened in the last month: ____
10. A question I suspect it will get wrong: ____

My scoring rule (decide before reading answers):
Good as is / Good after a small fix / Not usable
Extra point off for any made-up fact.

Keep the list. When a new model launches with a chart that says it beat everyone, run your ten questions again. That hour tells you more than any leaderboard about whether to switch.

That is the last lesson of Under the Hood. The habit to keep: when someone tells you a model is the best, ask "best at what, scored by whom?" Then test it on your own questions.