The paper was first submitted on April 1, 2026 and the listed page identifies the current version as v5, revised October 21, 2026.
The survey examines on-policy distillation: training a student language model with guidance on trajectories generated by the student itself. Its stated target is the mismatch in conventional distillation, where a student trains on fixed teacher demonstrations but, at deployment, must continue from its own earlier outputs and errors.
Song and Zheng organize the area around student-generated states, teacher information, and policy updates. They connect the framing to self-distillation, multiple teachers, reinforcement learning, and interactive agents, and discuss supervision reliability, learnability, capability transfer, training efficiency, and computation allocation.
This is a survey and describes itself as work in progress. The source does not establish that one on-policy method is best, report a new model or a common benchmark result, or provide evidence that the approach improves a particular deployed system.