mixture-of-experts-adaptive-capacity

Optimize Mixture-of-Experts efficiency by decoupling token-level expert activation from layer architecture. Use dynamic threshold routing where expert count per token varies by input complexity, and apply layer-wise capacity scheduling to match representational diversity patterns.

Independent GitHub snapshot · Automated classification · Original documentation governs