On-policy self-distillation aims to incorporate session-learned information into model weights.

“The whole point of this procedure is to distill what the model learned in a session back into the weights themselves.”