New Rule Tells AI When to Stop Guessing — Saving 62% of the Work

Imagine a security scanner, a medical imaging device, or a factory camera that decides for itself where to look next. Each time it takes a look, an AI produces a pile of possible answers — basically educated guesses about what it's seeing. The obvious question is: how many of those guesses do you need before you can confidently announce an answer? Stop too soon and you get it wrong. Keep going and you burn time, computing power and money for nothing.

The researchers focused entirely on that stopping decision. The natural approach is to declare an answer once most of the AI's guesses agree — say, 80% of them. But they showed this rule carries a hidden trap: when the true level of agreement sits right at the threshold, the system declares an answer roughly half the time. That's closer to a coin flip than a guarantee. So they designed replacement rules that are calibrated to a target error rate you actually choose.

Then they tested it. Using a classic dataset of handwritten digits and a fixed sequence of looks, their "stop as soon as the answer is locked in" approach saved up to 62% of the AI's guessing work. A smarter sequential version performed best overall. There's a bonus trade-off too: at high accuracy, that same rule can spend more computing power to take fewer physical measurements — useful when each measurement is costly, slow, or exposes someone to radiation.

The honest caveat is that this is a statistics paper tested on simple images, with a pre-set order of where to look, not on a real medical device. But the core idea — teaching AI to recognise when it already knows enough — is exactly what's needed to make smart scanners cheap enough to use everywhere.