In the context of attention, the temperature parameter controls the sharpness of the softmax: low temperature makes attention concentrate on a few keys (low entropy, possibly overfitting), while high temperature makes the weights more uniform (high entropy, possibly under-focusing).