Picking the best training data for a model isn't a fixed rulebook. It depends on which model you're training.
Researchers studying coreset selection, the practice of choosing a smaller, budget-limited subset of training samples, compared two common strategies: picking the easiest examples first versus picking examples that give broad geometric coverage of the data. Depending on the budget size, one strategy beats the other, with a crossover point separating the two regimes. By freezing the exact same selected subsets and only swapping the learner, the team found that doubling the width of a ResNet-18 model moved that crossover point from 57 to 85 samples per class on a downsized ImageNet-100 test. Changing image resolution, adjusting stride settings, or replacing the convolutional network with a Vision Transformer also shifted or erased the boundary, and coverage-based selection won across the board once a ViT was used, even when the easy-example subsets came from a convolutional model.
That is a problem for a popular efficiency shortcut: pick a smaller training set once, then reuse it across different models to save on compute. This research suggests that shortcut can quietly break the moment you change a model's width or swap architectures, because the subset that works best for one network is not necessarily best for another.
So before trimming a dataset to cut training costs, check whether you are also changing models. The researchers are upfront that they have not found a universal scaling law, just evidence that what counts as the best data is a moving target.