Definition
Plain language
Picking practice problems for an AI by checking whether training on them would push the model in the same direction as getting better at the real goal.
As stated in the literature
A data-selection criterion that scores a candidate training example by the agreement (e.g., cosine similarity) between its training gradient and the gradient of a held-out validation objective, keeping examples whose update direction points toward the target distribution rather than selecting by raw difficulty.
Why it matters: It picks training data by how much it actually helps the goal rather than by surface difficulty, avoiding examples that push the model the wrong way.
For example, before adding a practice problem to training, the method checks whether learning it would nudge the model in the same direction as doing better on the real target task.