Scaling Laws: New Axes and Coupling

A Harvard paper introduces exploration as a third scaling axis for generative models, achieving 4.1x FLOP and 6.2x sample efficiency across image, video, NLP, and robotics. Skaling introduces a coupled scaling law linking model size and data through a single interaction exponent, achieving 1.5-3x MAPE improvement and 10x compute savings for extrapolation. Together, these challenge conventional scaling narratives and offer practical guidance for training strategies.