On-policy self-distillation (OPSD) is a training strategy described in a new arXiv preprint. In OPSD, a large language model is trained using supervision from a teacher instantiated from the same model as the student, rather than from a fixed external model.
The teacher is conditioned on privileged information (PI); the abstract begins to specify this as "such as verified…" before the text is cut off. The paper's title refers to "verbal regularization," but the abstract excerpt does not explain that term. Since the source is only a brief abstract, details about the training procedure, experiments, and results are not available here.
The approach appears to be a variant of self-distillation that gives the teacher access to information not available to the student at inference time. As with any preprint, the findings have not yet been peer-reviewed.