RP-OPSD: Resolution-Privileged On-Policy Self-Distillation for Multimodal Large Language Models
2026-07-27 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors propose a method called RP-OPSD that helps large multimodal language models learn better by comparing how they perform on low-resolution versus high-resolution images. The student model works with blurry images, while the teacher model uses clear images to guide the student. This way, the student learns to imitate the teacher's behavior without needing extra annotations or external help. Their experiments show that this approach improves performance and speeds up training compared to previous methods. Overall, the authors demonstrate that using resolution differences is a practical way to teach these models.
On-Policy Self-DistillationMultimodal Large Language ModelsResolution-Privileged InformationTrajectory SupervisionImage ResolutionKnowledge DistillationQwen3.5-9BToken-Level SupervisionTraining Efficiency
Authors
Qihui Zhu, Yuchen Wang, Zijian Wen, Tao Zhang, Mengjie Zhang, Yang Liu, Shuangwu Chen, Siying Wu, Jian Yang, Xiaofeng Jiang
Abstract
On-Policy Self-Distillation (OPSD) uses privileged information available only to the teacher to provide dense token-level supervision on trajectories generated by the student. However, existing methods often rely on verified solution traces, explanations generated by external models, or manually localized visual evidence, which limits their scalable application to multimodal large language models. To address this issue, we exploit the information gap between high- and low-resolution views of the same image and propose RP-OPSD (Resolution-Privileged On-Policy Self-Distillation for Multimodal Large Language Models). During training, the student policy generates on-policy trajectories from images at one-quarter of the original resolution, while the teacher policy provides supervision using the original-resolution images. By minimizing the divergence between their output distributions along the student trajectories, the student learns the predictive behavior of the teacher under high-resolution inputs, thereby strengthening its low-resolution capability and transferring the learned improvement to original-resolution inference. RP-OPSD requires neither additional human annotations nor external models to generate solution traces but only image--question pairs. Experiments on Qwen3.5-9B show that RP-OPSD achieves a 5.45\% relative improvement in average performance at the original resolution and a $1.78\times$ training speedup over OPSD. These results demonstrate that resolution differences can serve as a simple and scalable source of privileged information, providing an effective and efficient approach to on-policy self-distillation for multimodal large language models.