📝 SummarXiv

Abstract

Transformer-based language models traditionally use uniform (isotropic) layer sizes, yet they ignore the diverse functional roles that different depths can play and their computational capacity needs. Building on Layer-Wise Scaling (LWS) and pruning literature, we introduce three new LWS variants - Framed, Reverse, and Crown - that redistribute FFN widths and attention heads via two or three-point linear interpolation in the pre-training stage. We present the first systematic ablation of LWS and its variants, on a fixed budget of 180M parameters, trained on 5B tokens. All models converge to similar losses and achieve better performance compared to an equal-cost isotropic baseline, without a substantial decrease in training throughput. This work represents an initial step into the design space of layer-wise architectures for pre-training, but future work should scale experiments to orders of magnitude more tokens and parameters to fully assess their potential.