Ranking training samples by how well their gradients align with validation set gradients to prioritize high-impact data.