active learning

Active Experiment Selection for Cheaper Scaling Laws

A budget-aware method picks the most useful pilot runs for scaling-law fitting, often matching full-data fits with about 10% of the training budget.