OPSD uses a single LLM's shared weights for both teacher and student roles. The student sees only the question, while the teacher is conditioned on privileged information such as verified reasoning traces. Training minimizes per-token divergence between the two policies on rollouts sampled by the student.
The authors report results on multiple mathematical-reasoning benchmarks, claiming better token efficiency than the reinforcement-learning methods they compare against and better performance than off-policy distillation. The arXiv record says code has been released.
The paper does not establish that the method generalizes beyond its evaluated reasoning tasks, nor does it provide evidence here for production deployment benefits or comparisons to every alternative training method.