New Math Trick Cuts AI Training Costs by Nearly Half

Some AI work has to be done by trial and error, because you cannot look inside the model to see what is going wrong. Each guess costs what researchers call a "query" — a bit of computing power. A new paper by Yangyang Feng and Yao Shu introduces ZoAQ, a method that keeps the answers from earlier guesses and reuses them for the next step. Instead of paying extra to check whether a guess was reliable, the method lets old answers do double duty. The result: roughly 43–48% fewer queries than the standard approach.

The numbers are specific. On test problems, ZoAQ cut queries by 43–48% compared with fixed-budget methods that used 1.2 million queries. In "black-box attacks" (tricking an AI by poking it like a locked box, without seeing inside), it hit 100% success using about 320 tries on handwritten-digit images and 625 tries on small colour photos. And when fine-tuning OPT language models across four settings, it needed 43–46% fewer evaluations, with accuracy changing by less than two hundredths of a point.

Why should you care? Training and customising AI is expensive, and that cost shows up in subscription prices, product features, and how much energy data centres burn. If a technique can deliver nearly the same quality for roughly half the computing, it could make custom AI cheaper and faster to build — smaller companies could afford it, not just the giants. The same efficiency, though, also helps anyone probing an AI system for weaknesses, since attacks get cheaper too.

The honest catch: this is a 40-page theory paper tested on synthetic problems, small image datasets, and mid-sized language models. It has not been proven on the giant systems people actually use, and "fewer queries" does not automatically mean a smaller bill in real data centres. Treat it as a promising cost-cutting idea that still needs real-world proof.