New Training Trick Makes AI Learn 30% Faster — Same Results

Training an AI model is like baking: you can't rush it, and the ingredients cost real money. One popular ingredient, called SAM (short for sharpness-aware minimization), makes a model more accurate by having it double-check each adjustment it makes. The catch is that this doubles the cost of every single step of training — even the early steps where it barely helps. A new paper from researcher Stamatis Mastromichalakis asks a simple question: what if we only use the expensive ingredient at the end?

The answer is a method called Temperon. For the first 43% of training, the model learns the plain, cheap way. Then it switches once to the expensive SAM method for the final stretch. Across four standard image tests (CIFAR-10, CIFAR-100, SVHN and Tiny ImageNet), Temperon matched the best always-expensive approach on accuracy while reaching the hardest accuracy target 32–35% sooner on three of the four tests.

The trick wasn't invented for images alone. The paper shows it also works for GPT-2, a well-known text model: full quality at 29% less real-world training time. On a set of standard language tests called GLUE, it was never worse than the expensive method while using only a third of the costly steps. The author also ran careful experiments to see what actually mattered — and found the switch itself, not other clever tweaks, did the work. Those extra tweaks were dropped from the paper.

The honest limitation: the win isn't universal. On one test (Tiny ImageNet) a rival method was simply faster, and the author says so plainly. The method also relies on knowing your training schedule in advance, since you can't tell from the model's behavior when to flip the switch. But for teams paying for GPU time, a 30% cut in training hours is a real bill reduction.