自監督視覺同策略蒸餾

摘要

視覺在策略蒸餾高度依賴於具有資訊性的師生不對稱性，無論是透過更大的、更強的教師模型，還是透過特權監督（例如參考答案或真實標註的感興趣區域）。這提出了一個基本問題：當沒有任何特權資訊可用時，資訊性不對稱性從何而來？我們透過逆轉不對稱性的來源來回答這個問題。我們不是向教師添加特權資訊，而是從學生端減去資訊。這種不對稱性無需真實標註、獎勵或單獨的更強教師模型，即可免費產生與能夠獲取學生無法獲得之資訊的教師相同的有效學習信號。基於這一原則，我們提出了自監督視覺在策略蒸餾（S²VOPD），這是一種簡單而有效的方法，從不對稱的增強視圖中構建在策略學習信號。S²VOPD 以在策略方式將教師在原始圖像條件下的分佈蒸餾至學生在該圖像的強增強視圖條件下的分佈。我們系統性地探索了視覺增強廣闊的設計空間，並發現：(1) 不對稱性至關重要：所有四類增強家族均提升了性能，而對稱自蒸餾則使其下降；(2) 強度很重要：性能在中等強度時達到峰值；(3) 差異必須保持任務一致性：完全移除與問題相關證據的增強會導致巨大但無資訊性的差異。在六個細粒度感知基準上，S²VOPD 將 Qwen3.5-4B 從 70.7% 提升至 77.4%，超過了所有被比較的開源模型（包括高達 235B 的 Qwen3-VL），並超越了 GPT-5.4。在保持訓練數據相同的情況下，它恢復了使用特權資訊之方法所取得改進的 96%。網站位於 https://williamium3000.github.io/s2vopd

English

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^2VOPD), a simple yet effective method that constructs on-policy learning signals from asymmetric augmented views. S^2VOPD 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^2VOPD 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