Self-Supervised Visual On-Policy Distillation Method Introduced

A new research paper on arXiv (2608.14144) introduces Self-Supervised Visual On-Policy Distillation (S$^2$VOPD), a method for visual on-policy distillation that does not require privileged information. Traditional approaches rely on a teacher-student asymmetry, often using a larger teacher or privileged supervision like reference answers or ground-truth regions. The paper inverts this by subtracting information from the student instead of adding to the teacher, creating an effective learning signal without annotations, rewards, or a separate stronger teacher. The method constructs on-policy learning signals from asymmetric augmented views. The paper is authored by researchers and was announced as a cross-type submission. The work addresses a fundamental question in reinforcement learning and computer vision about the source of informative asymmetry when no privileged data is available.