GRASP: Geometry-aware Residual Alignment for Scalable Pretraining Data Attribution

In three linesGRASP is a data attribution method that models interactions between training subsets via quadratic geometric penalty, surpassing isolated additive approaches. Evaluated on pretraining-scale tasks, it doubles rank correlation for counterfactual subset fidelity and reduces artifact construction costs by nearly an order of magnitude.

Summary generated by Claude — human-verified