Reproduce “Self-Supervised Visual On-Policy Distillation” on OpenResearch

About this paper

Yijiang Li, Yijun Liang, Yunjie Tian, Bingyang Wang, Ke Zhang, Zhenfei Yin, Di Fu, Philip Torr and 1 more

Self-Supervised Visual On-Policy Distillation (S²VOPD) introduces a method to generate an informative learning signal for vision-language models by having a teacher observe a clean image while a student sees an augmented version. This approach achieves a 6.76 percentage point gain on fine-grained perception benchmarks for a Qwen3.5-4B model, surpassing larger open-source and several proprietary models without relying on privileged information or external rewards.

Key insights

• Informative teacher–student asymmetry can be generated purely through self-supervision using asymmetric visual augmentations, eliminating the need for external annotations, rewards, or stronger teacher models.
• The effectiveness of visual augmentation for this self-supervised distillation is dependent on creating a 'task-consistent' information gap, where moderate augmentation strength is optimal and overly aggressive augmentations (e.g., heavy cropping) can degrade performance by removing task-relevant evidence.
• A balanced divergence measure, such as the generalized Jensen–Shannon divergence, is crucial for effectively transferring knowledge from a clean-view teacher to an augmented-view student, outperforming unidirectional KL divergences.

Summary by alphaXiv.