Benchmaxxing is a measurement failure

Benchmaxxing is optimizing a benchmark score instead of the capability it is meant to reflect, and it works because a raw benchmark score is a count, not a measurement. NIST AI 800-3 (February 2026) puts the gap plainly: intervals on a latent scale have meaning whichever tasks are chosen, and "Raw benchmark scores do not have this property."

Run your favorite model against your favorite benchmark twice. Same model, same items, same harness. Two scores. Run it ten times and you get a band, not a point. Every practitioner has seen this. Few ask the question it raises: if the number moves when nothing moved, what exactly was it measuring?

A benchmark score is a count of correct answers. A count tells you something, and it usually ranks strong systems above weak ones. It is not yet a measurement, because the distance between two counts depends on which items happened to be in the set. A benchmark is an unmarked stick. It can show you that one thing is longer than another, sometimes. It cannot tell you how much longer, and it changes length when the items change.

That gap is what makes benchmaxxing possible: optimizing the score instead of the capability. Engineers are not careless. The score is the only instrument they were handed, and the score rewards its own optimization. Prompt tuning, item selection, decoding tricks, a favorable retry policy: the number climbs while the capability underneath may not move at all. Once a number becomes the goal, it stops being a clean observation. That finding is old in the measurement literature, and the AI field is its newest tenant. Contamination pushes the same way. Golchin and Surdeanu (ICLR 2024) found GPT-4 had seen test data from AG News, WNLI and XSum, and Sainz et al. (Findings of EMNLP 2023) showed contamination causes "overestimation of the performance of a contaminated model."

Three failures follow, and any one of them should stop a decision of consequence.

First, instability. A score that moves across runs with no change in the performer comes from an instrument with no calibration. If a thermometer read 21, then 24, then 19 in a stable room, you would not argue about the room. You would replace the thermometer.

Second, incomparability. Raw scores share no common scale across benchmarks, versions, or systems. A 0.85 on one suite and a 0.72 on another are not amounts of the same thing. They are rankings with decimal points, and rankings built from different item mixes do not subtract, average, or trend.

Third, error that is missing or misplaced. Many leaderboards still print a bare score. When an interval does appear, it usually describes resampling of that one item set, not the error in where the model sits on a scale that other item sets share. The number looks precise to two decimals and says little about the uncertainty a decision needs.

NIST reached the same point in its report on expanding the AI evaluation toolbox with statistical models (NIST AI 800-3, February 2026): "the intervals between LLM capabilities θ on the latent scale have meaning independent of the difficulties of the specific tasks selected into the benchmark. Raw benchmark scores do not have this property." That is the national measurement institute of the United States saying that the difference between two raw scores depends on which items were selected.

The way out is old and published. Model item difficulty and performer capability on one latent scale, so that right-and-wrong patterns become interval-scale estimates, each with its own standard error. That is Rasch measurement, the probabilistic form of additive conjoint measurement. A calibrated ruler behaves differently from an unmarked stick for two reasons: its intervals mean the same thing across the scale whichever items you use, as long as the model fits, and every value carries its uncertainty on its face.

That is the standard we hold our own work to, and our results are early evidence and not yet certified. The calibration-validity program behind them is pre-registered on the Open Science Framework (https://osf.io/j95ef, registered 2026-07-05), and its confirmatory study is not yet complete. On the AIM ruler, the precision of record is published with its unit and its date on our Evidence page (corrected 2026-09-26: an earlier figure here lacked both, so it is withdrawn). The method is published: the approach to metrology across physical and human measurement is Chapter 3 of Models, Measurement, and Metrology Extending the SI (De Gruyter, 2024), open access at https://library.oapen.org/bitstream/handle/20.500.12657/94377/1/9783111036496.pdf#page=114 , and the Cambridge Handbook of Technology and Employee Behavior (2019, pp. 38-56) sets out the role of AI in workplace measurement.

The AI evaluation community can do better, and we will argue for it here, post by post. Benchmark teams do not lack talent. They were handed an unmarked stick and asked to make decisions with it. Reading the stick more carefully will not help. A ruler will.

Future posts cover why we publish the error bar on every value, what it takes to put humans, AI agents, and robots on one calibrated scale, and what evaluation looks like when an agent acts with no human watching.

If a number is going to carry a release, a vendor choice, or an audit answer, it should be a measurement. If your team has a score change it cannot explain, the Design-Partner Readout is the place to start: send a redacted slice of the evaluation data you already keep, traces included, and in about 30 days you get a written early read of how much of the change came from the model, the judge, the dataset, the task, and the prompt. Ask for one at xlnc.co/solutions. An email address is enough to start.

Different teams need different things, from a hosted service to a self-hosted harness that keeps your data on your side. Tell us which fits yours at xlnc.co/solutions.