Attention-Aware Routing: Coupling Routing and Attention in MoEs
Attention-Aware Routing (AAR), which augments the router with temporal and spectral features extracted from a sliding window of attention weights that represent a summary of the model's contextual state, disentangled from the hidden state, is proposed.
Despoina Kosmopoulou, Anastasios Tsetsilas, Efthymios Georgiou et al.
· 0 citations