Proposal: Validating Scaling Law HP Predictions via Small Scale Sweeps

JJitsev · x · 2026-08-20

Proposes a follow-up on scaling laws: show that given extensive hyperparameter sweeps at smaller scales, the predictions for larger scales land close enough to optimal hps. This relies on the existence of an extensive neighborhood around the optimal that forgives prediction errors.

Related event: Scaling Laws Debate: Easy Hyperparameter Identification Is Not the Same as Predictability(7 posts)→

Original post →

More from Research

Research channel →