Task-Agnostic Gradient-Based Method Measures Training Data Influence

A new metric estimates training data influence by measuring how much a gradient update reduces the squared distance to final parameters. This approach allows for influence estimation from intermediate checkpoints without requiring specific downstream tasks or validation sets.