Lower teacher predictive entropy corresponds to substantially higher teacher top-1 agreement with the ground-truth token across all data domains and teacher sizes.

来源：source_paper:Figure 3, Page 5