How Benchmarks Get Made

When you read in the news that "some model scored 90 on some benchmark", you probably assume the 90 is as objective as a thermometer reading. Turn the exam paper over and look at the back: a benchmark goes through writing questions, finding annotators, publishing, then being ground out by every lab, saturating, and finally being quietly retired. Run that whole sequence together, and it is an industrial pipeline with a full life cycle. Every stage of the pipeline is rushed through by people, and once people rush, things can go wrong. Once you understand this pipeline, you will look at any benchmark score and first ask: is this paper itself clean, and can it still measure anything?

You learned in Y01 that repeatedly picking models on the same batch of questions overfits to the luck of those questions. Further upstream is where that batch of questions came from and who decides what is right. This is also the foundation of Y03's contamination problem: you have to know how questions are made before you can talk about how they leak.

Start at birth. A benchmark is born carrying an inescapable original sin called the construct validity problem. Raji et al. punctured it in 2021 by borrowing from psychometrics: benchmarks like ImageNet and GLUE, declared to measure "general capability", are actually products of specific data, specific metrics and specific annotation practices, and can only measure the narrow construct they really instantiate. Using one as a proxy for "general intelligence" is a validity fallacy in itself. In plain terms, an exam paper only ever tested the questions it asked, while you use its score to talk about a much broader capability, and that gap has been there since the day it was born.

Next comes annotation, deciding a gold answer for every question, the most labour-intensive step and the most error-prone. There is a number here that stops people short: MMLU is one of the most mainstream knowledge benchmarks, yet the MMLU-Redux project in 2024 manually relabelled 5700 of its questions across all 57 subjects and estimated that about 6.49% of the questions are themselves wrong, with mislabelled answers, truncated stems, or several options being correct. What that 6.49% means: the ceiling on model accuracy is nailed at around 93.5% by those bad questions, and no model however strong can pass it, because the "correct answers" in those remaining percentage points are themselves wrong. You thought the model sprinting for full marks had hit a capability ceiling; what it actually hit was an annotation ceiling. And this is not MMLU's private defect: an audit of a batch of open-source Text-to-SQL benchmarks found 66.1% of 121 gold answers mislabelled, badly misestimating model capability. Broken exam papers are a systematic cross-domain disease.

Past birth and annotation, a benchmark enters its most glorious and shortest-lived stretch: release, leaderboard, grinding. The macro fingerprint here is a shrinking useful lifespan. The Stanford AI Index 2025 gives a striking set of figures: MMMU, GPQA and SWE-bench, all introduced in 2023, saw their scores jump by 18.8, 48.9 and 67.3 percentage points respectively just a year later. What that means: the window in which a benchmark goes from "can separate frontier models" to "everyone is near full marks and nothing distinguishes them" is being compressed to a matter of months. But the popular strong narrative also needs discounting here. "Saturation is accelerating" sounds alarming, but Akhtar, Reuel et al. systematically analysed 60 mainstream benchmarks in 2026, defining saturation as "losing statistically distinguishable resolution among top models", and found the saturated fraction rising from 42.9% for benchmarks under two years old to 54.5% for those over five, so it rises with age, but the authors describe the trend as gentle rather than steep. The more counterintuitive line: public test sets and privately held test sets show no statistical difference in saturation rate. Hiding the questions does not stop them saturating. The steep early climb of a few new benchmarks and the gentle long-run trend across all benchmarks are two different things, so do not mix them and frighten yourself.

The last stop in the life is retirement, and retirement is usually undignified. The coding benchmark SWE-bench's whole arc is a specimen-grade death record. The original had models solve real GitHub issues and was for a while the gold standard; problems were found quickly, producing the human-screened clean edition, Verified. But in 2026 OpenAI itself announced it would no longer use Verified, on the grounds that at least 16.4% of task tests are themselves faulty, about 59% of the hardest 138 items, the ones that best separate top models, are broken, and frontier models across labs can recite the human-written fix patch of the day almost word for word, which is reciting an answer rather than reasoning. An earlier audit had also found that about 32.67% of "successful" patches in the original had the answer written directly in the issue description or comments. Then came the anti-contamination successor, Pro, which hid part of its questions as a private bank, and a 2026 audit found about 30% of its items broken, with pass rates shooting from 23.3% to 80.3% in eight months, and the recommendation was simply withdrawn. What this arc means: each fixed generation gets ground out, then a new version is built, and each new version's lifespan is shorter than the last. The speed of building new benchmarks is falling behind the speed at which they are broken. Retirement is not all bad, of course: approaches like Dynabench have humans and models write questions against each other, deliberately hunting samples the current model fails, turning the question pool into a living thing and hedging short lifespans with continuous transfusion.

One open question: if every large-scale knowledge benchmark has a natural floor on its error rate (human annotation quality only goes so high), is the measurable ceiling on accuracy stuck around 95% all along, and has the arms race towards 100% been a measurement illusion from the start? If so, then what should really be compared is not who scores higher but who still stands up on cleaner, uncontaminated questions, which hands the thread straight to Y03's contamination.

The one-line takeaway: a benchmark is a living thing built by people against the clock, it can be wrong, it can be ground out, it can retire; before reading the score, check whether the paper is clean and whether it can still measure anything.