In the channel-blind synthetic superposition task, baseline error rises rapidly with feature load, off-by-one abstention does not help, and only the value-slot norm gate recovers the trained oracle; when routing can see content, routing focuses on relevant items, baseline and oracle coincide, and the gate has no advantage. · CiteArk
In the channel-blind synthetic superposition task, baseline error rises rapidly with feature load, off-by-one abstention does not help, and only the value-slot norm gate recovers the trained oracle; when routing can see content, routing focuses on relevant items, baseline and oracle coincide, and the gate has no advantage.
