On-policy distillation (OPD) is a technique for improving language models by having a student learn from a teacher while sampling prefixes from the student's own policy. This self-policy sampling makes the supervision more relevant to the student's actual behavior. However, the abstract notes that OPD's supervision remains confined to the output space, meaning it only transfers what the teacher predicts rather than deeper reasoning processes.
The new method, called On-Policy Mechanism Distillation (OPMD), is positioned as going beyond output supervision. Instead of only matching the teacher's predictions, OPMD aims to distill the underlying mechanisms that produce those predictions. This could allow the student to learn more robustly from the teacher, though the abstract does not provide experimental details.
Because the source is a single arXiv abstract, the article is limited to the stated motivation and proposed direction. No comparison with other methods or quantitative results are available in the provided text.