Empirical scaling laws for autoregressive language models relate prediction loss to model size, data size, and optimization compute, but their theoretical origin is still poorly understood in sequential pretraining settings. We study this question in a tractable teacher--student model where a stable latent linear RNN generates trajectories and a sketched linear recurrent student is trained by safeguarded full-batch WSD gradient descent on next-token prediction. The sketch dimension $M$ plays the role of model size, while $N$ independent trajectories of length $P$ provide the training tokens. We allow the innovation and initialization covariances to have different power-law exponents $α$ and $θ$. The induced design spectrum produces explicit approximation, optimization, and statistical scaling laws separated by spectral crossovers. When $θ\geα$, the original one-scale rates $M^{1-β_α}$, $R^{(1-β_α)/α}$, and $(NP)^{-1}\min\{M,R^{1/α}\}$ are recovered. When $α-2r\leθ<α$, the heavier initialization tail changes the rates beyond $P$-dependent model and optimization crossovers. The proof uses a covariance event only internally and a globally safeguarded step size on its complement. The variance retains the factor $(NP)^{-1}$, while sequence length also suppresses the initialization transient, so $N$ and $P$ cease to be fully interchangeable in the two-scale regime.

翻译：暂无翻译