aiqbench analysis
SR-OPSD introduces a novel reinforcement learning technique that converts sparse outcome rewards into dense token-level supervision by constructing a self-referenced geometric target. By minimizing forward Rényi divergence between this evolving teacher and the student policy, the method effectively addresses the moving target problem common in on-policy distillation. Experiments across scientific reasoning, tool use, and code generation confirm its robustness across various model families and scales.
SR-OPSD: Self-Referenced On-Policy Self-Distillation
arXiv:2608.09745v2 Announce Type: replace Abstract: On-policy self-distillation (OPSD) converts feedback into dense token-level supervision on student-generated trajectories, complementing reinforcement learning with sparse outcome rewards. Its self-teacher, derived from the student's current or exponentially averaged parameters and conditioned on additional context, evolves alongside the student and its rollout context distribution. The benefit of modifying this moving target depends on how target--student probability mismatches translate into updates. We propose \emph{Self-Referenced On-Policy Self-Distillation (SR-OPSD)}, which constructs a normalized geometric target from the self-teacher and a frozen initial policy, then minimizes the forward R\'enyi divergence from this target to the student. The interpolation coefficient controls the self-teacher's contribution, while the R\'enyi order controls the power weighting of target-to-student probability ratios in the gradient. For fixed contexts and target components, we establish a conditional variational characterization and derive the exact token-logit gradient, revealing how anchoring and projection jointly shape the effective update target. Experiments across scientific reasoning, tool use, mathematical reasoning, and code generation demonstrate strong performance across multiple model families and scales. Ablations further show that reference anchoring can improve or degrade performance depending on the projection objective, supporting the joint design of target construction and projection geometry.
Abstract:On-policy self-distillation (OPSD) converts feedback into dense token-level supervision on student-generated trajectories, complementing reinforcement learning with sparse outcome rewards. Its self-teacher, derived from the student's current or exponentially averaged parameters and conditioned on additional context, evolves alongside the student and its rollout context distribution. The benefit of modifying this moving target depends on how target--student probability mismatches translate into updates. We propose \emph{Self-Referenced On-Policy Self-Distillation (SR-OPSD)}, which constructs a normalized geometric target from the self-teacher and a frozen initial policy, then minimizes the forward Rényi divergence from this target to the student. The interpolation coefficient controls the self-teacher's contribution, while the Rényi order controls the power weighting of target-to-student probability ratios in the gradient. For fixed contexts and target components, we establish a conditional variational characterization and derive the exact token-logit gradient, revealing how anchoring and projection jointly shape the effective update target. Experiments across scientific reasoning, tool use, mathematical reasoning, and code generation demonstrate strong performance across multiple model families and scales. Ablations further show that reference anchoring can improve or degrade performance depending on the projection objective, supporting the joint design of target construction and projection geometry.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
arXiv:2608.09745 [cs.LG]
(or arXiv:2608.09745v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2608.09745 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 287 words
Related coverage
On-Policy Self-Distillation without Any Supervision
arXiv · Aug 7, 2026
$\beta$-OPSD: Deriving with Policy Optimization, Training with Self-Distillation
arXiv · Jul 31, 2026
Train Ahead, Distill Back: Bootstrapping On-Policy Self-Distillation for Large Language Models
arXiv · Sep 30, 2026
Projection sampling preserves capabilities but diminishes gains from alignment fine-tuning
LessWrong · Oct 16, 2026