VASPO: Improving Self-Distillation via Privileged Value Estimation
October 13, 2026
VASPO optimizes on-policy self-distillation by integrating a self-teacher and a critic within a PPO framework. It leverages the fact that teacher guidance is strongest at different generation stages compared to the critic's value predictions.
HOW THIS AFFECTS YOU
●
builderThis technique may improve the convergence and quality of reasoning models.
●
researcherThis identifies complementary positional dynamics in reasoning model training.