Selectively disabling certain experts on a per-token basis to reduce the effective routing space while maintaining model expressiveness.