Tag
This paper characterizes the unique parameter space dynamics of on-policy distillation (OPD) for large language models, showing that it exhibits relaxed off-principal updates and subspace locking, distinguishing it from supervised fine-tuning and reinforcement learning with verifiable rewards.