Learnable placeholder tokens added to transformer inputs to absorb and stabilize problematic activations without affecting semantic content.