Meta FAIR's 'Skaling' Law Adds One Interaction Exponent to Chinchilla and Predicts Loss With 1.5–3x Less Error at 10x Less Sweep Compute

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

arXiv 2608.07222 (Aug 7) from Mathurin Videau, Badr Youbi-Idrissi, David Lopez-Paz, and Kartik Ahuja argues that existing scaling laws systematically under- and overestimate loss at both the data-scarce and overtraining extremes because they treat model capacity and data as independent. Skaling introduces a single interaction exponent coupling the two, cutting mean absolute percentage error 1.5–3x across both interpolation and extrapolation regimes. Paired with a sparse grid strategy it achieves full-grid extrapolation using roughly 10x less compute than uniform sweeps, which is the actionable bit: reliable performance prediction from small-scale experiments before committing a training budget.