scalingfactbullishThe three-term scaling law can be robustly fit with significantly fewer training runs because it uses data from suboptimal batch sizesMachine Learning (Statistics)27 Jul 2026http://arxiv.org/abs/2607.01487v1