Unified Neural Scaling Laws

In three linesUnified Neural Scaling Law (UNSL) functional form models simultaneous variation of model parameters, training dataset size, training steps, inference steps, compute, and hyperparameters on performance. Validated across vision, language, math, and reinforcement learning with more accurate extrapolations than existing scaling laws.

Summary generated by Claude — human-verified