Unified Framework for On-Policy Self-Distillation in LLMs

A new arXiv paper (2608.08176) proposes a unified optimization framework for on-policy self-distillation (OPSD) in large language models (LLMs). The authors argue that two recent research lines—one focusing on token selection and the other on controlling privileged information given to the teacher—are coupled through the student's learning capacity. They formalize these into a single framework that maximizes aggregate teacher-student divergence subject to a budget on learning difficulty. The paper introduces 'Unified On-' (likely a method name) to address suboptimal solutions from optimizing each variable independently. This work is relevant to AI research on improving reasoning abilities in LLMs through self-distillation.