Evolving Layer-Specific Scalar Functions for Hardware-Aware Transformer Adaptation

Proposes evolving layer-specific scalar functions to replace layer normalization in Vision Transformers for edge deployment, addressing the global reduction bottleneck while avoiding expensive full model retraining through hardware-aware adaptation.