$β$-OPSD: Deriving with Policy Optimization, Training with Self-Distillation
2026-07-30 • Machine Learning
Machine Learning
AI summaryⓘ
The authors study a way to improve reasoning language models by using on-policy self-distillation (OPSD), which copies knowledge from a teacher model to a student model. They noticed that the usual OPSD method is limited because it always treats the student exactly like the reference policy. To fix this, they introduce a parameter β that balances between following the reference policy and learning from the teacher more flexibly. Their method, called β-OPSD, mixes the two policies’ outputs in a smart way to train models better and more stably without extra computational cost. Tests on math reasoning tasks showed that β-OPSD works better than the standard OPSD approach.
on-policy self-distillationpolicy optimizationKL divergencereinforcement learninglanguage modelsteacher-student traininglogits mixingreturn-to-gomathematical reasoning benchmarksregularization parameter
Authors
Jiawei Xu, Minghui Liu, Juzheng Zhang, Tom Goldstein, Furong Huang
Abstract
On-policy self-distillation (OPSD) is a promising approach to improve reasoning language models, but it remains brittle in practice: making it work reliably often requires substantial engineering effort. We identify a structural source of this difficulty: vanilla OPSD is precisely the $β=1$ member of a broader policy-optimization family, where $β$ weights the KL penalty anchoring the student to a reference policy. This equivalence turns $β$ from an implicit value fixed at one into a controllable regularization parameter, yielding a more general formulation that trades off proximity to a reference policy against privileged teacher guidance. We introduce $β$-OPSD and derive its optimal policy as a geometric interpolation between the reference policy and the privileged teacher. Directly optimizing this objective with reinforcement learning, however, would be costly and high-variance. Rather than optimize the RL objective directly, we turn its closed-form solution into a distillation target. Each value of $β$ selects a target along the reference-to-teacher path, which we implement efficiently by mixing their token-level logits. In this way, inexpensive distillation approximates the solution of expensive policy optimization. Return-to-go credit assignment further aligns token updates with the sequence-level objective while retaining the simplicity of OPSD. Experiments on mathematical reasoning benchmarks show that $β$-OPSD consistently outperforms vanilla OPSD, improving optimization stability and downstream reasoning performance. Our results provide a principled route from self-distillation to policy optimization and back without sacrificing the efficiency that makes OPSD practical.