NIST puts two different error bars on the same AI benchmark score

A figure caption states the finding plainly: some pairs of the tested language models differ significantly in benchmark accuracy but not in generalized accuracy ^[7]. An ordering can therefore be statistically solid about the exact items that were scored and say nothing defensible about the wider domain those items were drawn to represent.

That gap follows from where the randomness lives. Repeat trials on the same item move a score, which is why NIST says precision on the generalized figure improves by running more trials per item ^[10]. But if the questions themselves are treated as a sample from a larger population of similar questions ^[4], then which questions got drawn is a second source of variation, and only the generalized estimand has to carry it. NIST also notes that evaluators frequently do want to treat benchmark items as representative of a larger set, which makes generalized accuracy the reasonable target ^[14]. For anyone reading a vendor deck, the claim being made is usually the general one, while the interval being quoted, when one is quoted at all, is the narrow one.

The precision arithmetic is unforgiving in the ordinary way. If the standard error is the standard deviation of results divided by the square root of the number of questions ^[8], then halving an interval takes about four times as many questions ^[15]. More trials per item is often the cheaper purchase ^[10]. A generalized linear mixed model is cheaper still in items and buys further precision, at the cost of additional assumptions ^[11]. NIST compares the regression-free approaches against GLMMs on both real and simulated benchmark data ^[16] and declines to crown a method, putting the choice on the evaluator and the goals of the evaluation ^[13].

This framework does not address whether the benchmark was the right instrument. Generalized accuracy is defined against the population of questions similar to those in the benchmark ^[4], and that population is a statistical construct, not your ticket queue. If the quantity being estimated is the wrong one, adding uncertainty bounds around it does not fix that; it just attaches error bars to the same mistake. Nothing in this framework touches latency, the unit cost of inference, or behaviour on inputs nobody wrote a graded item for.

What it does give a buyer is two short questions that need no statistics to ask: which accuracy is this, and by which estimator. NIST names procurers alongside evaluators and developers as its audience ^[12], which is unusual candour about who is actually harmed by an uninterpretable number. Evaluators, meanwhile, still typically report a single proportion of correct outputs ^[5]. By the report's own standard, results whose assumptions are implicit and whose uncertainty is unquantified are difficult or impossible to make decisions from ^[3]; a stated estimand and a stated method are the minimum for treating a score as evidence.