How to read an AI benchmark score

An AI benchmark score is usually the share of a fixed set of test questions that a model gets right, under one set of conditions. It is a result for one test. It is not a general measure of how well a system will work for you.

What the percentage measures

MMLU (Massive Multitask Language Understanding) is a multiple-choice question set. The authors of MMLU-CF, a newer benchmark built from it, say that datasets like MMLU are widely used to evaluate large language models (MMLU-CF, arXiv:2412.15194).

The score depends on the setup as well as the model. The MMLU-CF authors report that GPT-4o scored 73.4% with 5-shot prompting (five worked examples in the prompt) and 71.9% with 0-shot prompting on their test set (MMLU-CF). The same model on the same test therefore gives two different numbers depending on how many examples the prompt includes.

Why a test set can leak into training

A model's training data can overlap with a public test set. This is called benchmark contamination. The PaCoST authors say that training data may include data from commonly used benchmarks, either unintentionally or intentionally (PaCoST, arXiv:2406.18326). The MMLU-CF authors link contamination to the open-source nature of these benchmarks and the broad sources of training data, and say it leads to unreliable evaluation results (MMLU-CF).

Proving that a leak has happened is hard. Sander and colleagues write that "it is difficult to assert whether a model has been trained on a test set" (Detecting Benchmark Contamination Through Watermarking, arXiv:2502.17259).

Why one score does not show real-world performance

The PaCoST authors say this overlap can lead to high scores on leaderboards while performance in real-world applications disappoints (PaCoST). A study of reasoning models by Wang and colleagues reports that, in one of the two scenarios they tested, most contamination detection methods performed near random guessing. They conclude that developers could contaminate such models to inflate leaderboard performance while leaving few traces (On the Fragility of Benchmark Contamination Detection in Reasoning Models, arXiv:2510.02386). A single leaderboard number therefore cannot settle whether a model is good at the task you care about.

Four checks you can make

1. Check the setup. Look for the test name and the number of examples in the prompt. The MMLU-CF figures above differ between 5-shot and 0-shot, so a bare percentage without this detail is incomplete (MMLU-CF).
2. Check whether the test set is public. The MMLU-CF authors split their benchmark into a validation set and a test set to prevent malicious data leakage. They publish the validation set and keep the test set closed (MMLU-CF).
3. Check whether the developer explains how it handled overlap. The MMLU-CF authors describe three decontamination rules used to build their benchmark (MMLU-CF). If a score comes with no such statement, treat it with more caution.
4. Check whether a "clean" claim rests on one detection method. Wang and colleagues report that evading contamination detection for reasoning models is easy (arXiv:2510.02386). A claim that a model is uncontaminated because one detector passed it is weak evidence.

Limits of this guide

All four sources are read from their arXiv pages. The arXiv records note that the PaCoST paper was accepted at EMNLP 2024 Findings and the reasoning-model paper at ICLR 2026; the other two records name no venue, so we treat them as preprints. We read the abstracts, not the full papers. This guide does not say how any particular commercial model was scored beyond the GPT-4o figures reported by the MMLU-CF authors.