Yijiang Li, Yijun Liang, Yunjie Tian, Bingyang Wang, Ke Zhang, Zhenfei Yin, Di Fu, Philip Torr, Nuno Vasconcelos
5 min
Abstract
Visual on-policy distillation relies heavily on an informative teacher-student asymmetry, through either a larger, stronger teacher or privileged supervision, such as reference answers or ground-truth regions of interest. This raises a fundamental question: where can informative asymmetry come from when nothing privileged is available? We answer this by inverting where the asymmetry comes from. Rather than adding privileged information to the teacher, we subtract information from the student. This asymmetry creates the same effective learning signal for free as a teacher with access to information unavailable to the student, without ground-truth annotations, rewards, or a separate stronger teacher model. Building on this principle, we introduce Self-Supervised Visual On-Policy Distillation (S$^2$VOPD), a simple yet effective method that constructs on-policy learning signals from asymmetric augmented views. S$^2$VOPD distills the teacher's distribution conditioned on the original image on-policy into the student distribution conditioned on a strongly augmented view of the same image. We systematically explore a broad design space of visual augmentations and uncover that (1) asymmetry matters: all four augmentation families improve performance, while symmetric self-distillation degrades it; (2) strength matters: performance peaks at a moderate strength; and (3) the gap must remain task-consistent: augmentations that completely remove the question-relevant evidence can induce large but uninformative discrepancies. Across six fine-grained perception benchmarks, S$^2$VOPD improves Qwen3.5-4B from 70.7% to 77.4%, above all open-source models compared, up to Qwen3-VL at 235B, and surpasses GPT-5.4. While holding training data the same, it recovers 96% of the improvement achieved by methods with privileged information. Website is at https://williamium3000.github.io/s2vopd
Sam: What about the degraded view the student gets? Why blur and downscale rather than, say, just cutting out parts of the image?
Alex: The researchers tested several approaches, and cropping — removing chunks of the image — actually hurt performance. Here's why: if you crop too aggressively, you might erase the exact region the question is asking about. The student isn't learning to recover missing detail; it's just guessing blindly because the relevant evidence is gone entirely.
Sam: So the gap is large, but it's not informative. It's just noise.
Alex: Right. The perturbation has to reduce quality without destroying the underlying context. Blurring and downscaling preserve the overall spatial layout — you can still tell where things are — while removing the fine-grained detail the student needs to recover. That creates a stable, meaningful learning signal. It also has a practical benefit: a smaller, lower-resolution student input means less computation during training.
Sam: So it ends up being more efficient and more effective at the same time.
Alex: That's what the results suggest. And the choice of how to measure the gap between teacher and student also matters. The authors compared several mathematical formulas for calculating that distance. A symmetric, balanced measure — one that doesn't privilege the teacher's perspective over the student's — consistently outperformed the alternatives. The intuition is that a one-sided measure can penalize the student for failing to reproduce details that were never visible to it in the first place. That's not a fair or useful signal.
Sam: You can't learn from a standard you were never given the tools to meet.
Alex: Exactly. The balanced measure avoids that trap while still transferring genuinely useful information from teacher to student.
Sam: Stepping back — what does this mean for where the field is heading?
Alex: The broader implication is that models may not need a constant stream of human feedback to keep improving. If a system can generate its own learning signal by comparing a sharp view to a degraded one, it opens the door to continuous self-improvement without expensive external supervision. The next question the field will likely ask is whether this same principle — creating an asymmetry in what two versions of a model can see — can be extended beyond images to other kinds of data.
Sam: It's a meaningful shift in how we think about what teaching an AI actually requires.
Alex: It is. And it suggests that some of the most useful supervisory signals might already be present in the data itself, waiting to be surfaced through the right kind of structural difference between what a teacher sees and what a student sees.
Sam: Thanks for walking us through it.
Alex: Thanks for listening to ResearchPod.