RCLA: Resonance-Coded Language Architecture — Template Routing as an Alternative to Transformer Feed-Forward Layers

We present RCLA, a language model architecture replacing transformer feed-forward layers with sparse template routing over a learned codebook. RCLA v2 (699M parameters) achieves 13.4% lower cross-entropy loss than a 838M-parameter transformer baseline on identical training data, while training 1.61x faster.