What Are AI Benchmarks? How to Read Model-Comparison Numbers Like 'MMLU 90%'

Every time a new AI model is announced, numbers line up: “achieved 88% on MMLU,” “beat the previous model on SWE-bench.” These benchmark scores have become the common language for discussing which model is better. Yet what those numbers actually measure, and how far you can trust them, is surprisingly rarely explained.

Chase only the announced scores without knowing how to read benchmarks, and you may jump to the conclusion that “a high number means it’ll be excellent for our business, too.” In reality, there are pitfalls where scores swell beyond a model’s true ability. This article organizes what AI benchmarks are—through the contents of major benchmarks and the reasons you can’t take scores at face value—from a decision-maker’s perspective. It’s a foundation for correctly reading news about large language models (LLMs) and reasoning models.

Why Are Benchmarks Needed?

When we say “this AI is smart,” how should we measure that smartness? If everyone has a different impression, you can’t compare whether model A or model B is better. So you need a common test that gives every model the same problems and grades them by the same standard. That is a benchmark.

Wikipedia defines a language model benchmark as “a standardized test designed to evaluate the performance of language models on various natural language processing tasks”¹. Like a standardized school exam or an automobile fuel-economy test, it’s a yardstick for lining things up and comparing them under equal conditions—that is the role of a benchmark in AI.

What Is a Benchmark Made Of?

A benchmark is made of two main parts. Wikipedia explains that it consists of “a dataset and corresponding evaluation metrics”¹. The dataset provides “text samples and annotations,” while the metrics “measure a model’s performance on tasks like answering questions, text classification, and machine translation”¹.

The grading method (evaluation metric) differs by task. For multiple-choice questions, accuracy is the basis. On the other hand, in situations like programming where “you may try several times and just need one correct attempt,” the metric pass@n is used. Wikipedia explains: “the model is given n attempts to solve each problem. If any attempt is correct, the model earns a point”¹. Even for the same “score,” the meaning changes depending on how many attempts produced it—a point worth noting.

Major Benchmarks: What Do They Measure?

There are hundreds of benchmarks, but the ones that frequently appear in the news can be broadly divided into “those measuring knowledge and reasoning” and “those measuring coding.” Let’s look at representative examples.

Benchmark	What it measures	Scale¹	What it asks
MMLU	Broad academic knowledge	16,000 Q / 57 subjects	Math, history, law, medicine, etc.²
GPQA	Graduate-level scientific reasoning	448 Q	Expert-grade questions in biology, physics, chemistry³
SWE-bench	Practical coding	2,294 tasks	Real bug fixes from GitHub⁴
HumanEval	Basic coding	164 Q	Write a Python function¹
GSM8K	Math word problems	8,500 Q	Elementary-level multi-step arithmetic¹

Measuring Knowledge and Reasoning: MMLU and GPQA

The most frequently cited is MMLU (Measuring Massive Multitask Language Understanding). Published by Hendrycks et al. in 2020, this test spans “57 tasks including elementary mathematics, US history, computer science, law, and more”², demanding broad knowledge and problem-solving ability. At the time of release, the paper reported that “while most recent models have near random-chance accuracy, the very largest GPT-3 model improves over random chance by almost 20 percentage points on average”². In other words, it was a hard test when it appeared, but as models advanced, each company came to score highly.

Newer and harder is GPQA (Graduate-Level Google-Proof Q&A). Published in 2023, this benchmark consists of “448 multiple-choice questions written by domain experts in biology, physics, and chemistry”³. Its distinguishing claim is being “Google-proof.” According to the paper, even PhD-holding experts reach only 65% accuracy (74% when discounting clear mistakes the experts identified in retrospect), while skilled non-experts “only reach 34% accuracy, despite spending on average over 30 minutes with unrestricted access to the web”³. By collecting hard problems that web search can’t crack, it tries to measure whether an AI can truly perform deep reasoning.

Measuring Coding: SWE-bench

Measuring ability in the development field is SWE-bench. Published in 2023, this benchmark uses “2,294 software engineering issues sourced from GitHub across 12 popular Python repositories”⁴, having the AI solve real GitHub issues (bug reports and feature requests). Rather than a quiz, it’s a sturdy test of whether the model can understand changes spanning multiple functions, classes, and files, and fix real code while interacting with an execution environment.

Worth noting is the score at the time of release. The paper reported that “the best-performing model, Claude 2, is able to solve a mere 1.96% of the issues”⁴. Even the cutting-edge model of the time could barely solve real development tasks. This extremely low score conversely means it was “a benchmark with great headroom that’s hard to game,” and it went on to become an important yardstick for measuring the subsequent progress of AI coding ability.

Three Reasons You Can’t Take Scores at Face Value

Benchmarks are useful, but taking announced scores directly as “true ability” is dangerous. There are three pitfalls a decision-maker should know.

1. Saturation

The first is saturation. Wikipedia describes it as the state in which “many models reach the highest performance level practically possible, and so the benchmark can no longer differentiate these models”¹. Even a once-hard test loses almost all meaning in a one-point gap once every company scores 95% or 98%. In fact, the once-standard benchmark GLUE became saturated, and the harder SuperGLUE was created¹. Marketing that competes over a few points on a “near-perfect” benchmark warrants careful skepticism.

2. Data Contamination

The second, and most important to watch, is data contamination. Wikipedia also calls this “training on the test set,” occurring when “benchmark questions may have answers already present in the training set”¹. Because an AI ingests vast amounts of web documents during pretraining, if a benchmark’s questions and answers are mixed in, the AI is not “solving by reasoning” but merely “remembering the answers.”

Research on benchmark contamination frames it as the phenomenon in which evaluation data leaks into a model’s training and fine-tuning data, “resulting in overfitting to supposedly held-out test sets and compromising test validity”⁵. As a result, scores swell beyond true ability. Some benchmarks take countermeasures—such as embedding markers called “canary strings” to find and exclude contaminated documents¹—but completely preventing contamination remains difficult. Even with a high announced score, whether it reflects “reasoning power” or “rote memorization” can’t be told from the number alone.

3. What Benchmarks Can’t Measure

The third is a more fundamental issue. A benchmark measures a specific ability in isolation; it doesn’t reflect everything needed for real-world work. Even the original MMLU paper points out that models “frequently do not know when they are wrong”². Even with a high score, the tendency to hallucinate—confidently producing errors—as well as response safety and fit with your own data, must be verified separately. It’s the same structure as how a person who scores high on a test doesn’t necessarily deliver results at work.

How to Read Benchmarks Wisely

With all this in mind, a way of relating to benchmark scores comes into view. What matters is neither to “ignore” the numbers nor to “treat them as absolute,” but to read them while being conscious of what test’s score this is.

From a practical standpoint, holding the following questions makes misjudgment less likely. Is the benchmark close to your use case (a knowledge quiz, or coding)? Is the score already saturated with little difference, or is it a domain with headroom like SWE-bench? And ultimately, rather than deciding by public benchmark rankings alone, test with your own real data and tasks—running a small test on “your own problem set,” so to speak, is the most reliable evaluation.

A new model’s splashy scores are an important clue to the direction of progress. But they are just one yardstick for side-by-side comparison, not the value itself for your organization. Read the news knowing how benchmarks work and where they fall short, and you’ll be able to gauge technical progress without being swept along by the numbers. If you want to learn more about what’s inside the models, reading alongside how reasoning models—which have pushed scores up—work, and the parameter count that expresses a model’s size, will sharpen the resolution of the debate over performance.