Choosing which tokens to update during training based on per-token confidence or convergence signals.