xin-icl-calibration-function-formula

Summary

This formula defines the baseline cost function that the in-context learning theory uses to measure how much information is lost when the model's confidence across classes is imbalanced; the more the class probabilities drift away from uniform, the higher the calibration cost becomes. Because it is taken as a fixed starting point rather than derived from anything else, every downstream calibration bound or guarantee in the ICL framework ultimately traces back to this particular shape of the penalty.