What do GPT-4's exam scores actually mean?

When OpenAI introduced GPT-4 on March 14, the headline number was its bar exam result. On a simulated Uniform Bar Exam, a standard US licensing exam for lawyers, GPT-4 scored 298 out of 400, which OpenAI estimates at about the 90th percentile of human test takers. GPT-3.5, the previous model, scored 213, around the 10th percentile. The jump is real and striking.

But a percentile is a comparison with people taking a test designed for people. To know what it means for you, it helps to look at the whole table, at how the tests were run, and at what critics are pointing out this week.

Read the whole table, not just the headline

The technical report lists more than 30 exams. A handful of rows tell a more mixed story than the bar exam alone:

The pattern is worth noticing. GPT-4 does best where questions reward broad knowledge and reading. It does far less well on essay writing about literature and on hard, fresh programming problems. The report also lists 3 of 45 hard problems solved on LeetCode, a coding practice site.

OpenAI itself cautions against overreading. Its announcement says the difference from GPT-3.5 can be subtle in casual conversation and shows mainly on complex tasks. It also says GPT-4 still makes up facts and makes reasoning errors.

How the tests were run

According to the report, OpenAI did no special training for these exams. Questions came from publicly available materials, and the scores were turned into percentiles using published score distributions. The bar exam was run by outside collaborators at Casetext and Stanford CodeX.

The key detail is contamination: test questions that also appear in the model's training data. A model that has seen a question before may be recalling rather than reasoning, like a student who memorized last year's answer key. OpenAI says a minority of exam questions turned up in the training data. For each exam it ran a version without those questions and reported the lower of the two scores.

What the critics say

Computer scientists Arvind Narayanan and Sayash Kapoor published a critique on March 20 on their blog, AI Snake Oil. They make two points.

First, contamination may be worse than the report suggests. They cite an observation by researcher Horace He: on easy problems from the coding site Codeforces, GPT-4 solved 10 out of 10 from before 2021 but none of 10 recent ones. GPT-4's training data runs to September 2021. They also argue that OpenAI's method, which looks for exact matches of short text samples, would miss a question that had been reworded or had its names and numbers changed.

Second, even a clean score may not mean what it seems. Exams for humans assume that a person who answers correctly has the underlying skills a professional needs. A model trained on a huge amount of text may answer through closely similar examples it has already seen, with much shallower reasoning. And the bar exam emphasizes legal knowledge, which language models handle well, more than practical skills like interviewing a client.

A mini-test on your own task

Before you trust any score, run a small test of your own. Collect ten real examples from your work where you already know the right answer, and keep the format the same each time:

You are helping a small accounting office. Read the client email below and list every document the client still needs to send us, as a bulleted list. If the email is unclear, say what you would ask the client.

Email: [paste a real, anonymized email here]

Score each answer as correct, partly correct or wrong, and note the kind of error. Ten examples will not make a scientific study, but they will tell you more about your use case than any bar exam percentile.

Why don't benchmarks predict how it will do on your task?

A benchmark is a fixed set of test questions used to compare models. Benchmarks are useful for tracking progress, but they have three gaps when you want to know whether GPT-4 will help you.

The questions are not your questions. A bar exam item has one correct answer and a tidy setup. Your work probably involves messy documents, unclear goals and information that changes. A model can be excellent at the first and unreliable at the second.

Your task may also involve things the model has never seen. GPT-4 generally knows little about events after September 2021, and it knows nothing about your company's policies, your local rules or your clients. Narayanan and Kapoor point out that the world keeps changing, so a model leaning on familiar patterns has less to draw on when something is new.

Finally, a single score hides how the model fails. An 88th-percentile result tells you nothing about whether its mistakes are harmless typos or confident errors that could cost you money.

That is why the mini-test above matters more than the headline number: it measures the model on your questions, your format and your standard of a good answer.