arXiv cs.AI· Yuqi Xiong, Chunyi Peng, Zhipeng Xu, Qing Yang, Zhenghao Liu, Zulong Chen, Yu Gu, Ge Yu·· 2 天前AI 评分22
SCOPED：面向统一视觉语言模型的技能解耦同策略专家蒸馏
Skill-Decoupled On-Policy Expert Distillation for Unified Vision-Language Models
AI 导读
研究者提出 SCOPED（技能解耦同策略专家蒸馏），通过训练能力专属专家并经奖励引导反思提取显式技能，在同策略蒸馏中仅用技能监督专家、不向统一学生模型暴露技能，使其在推理时内化过程性知识。在八个多模态基准上，SCOPED 持续优于骨干模型与混合能力 RL。分析显示技能增强可提升专家轨迹一致性、对任务相关视觉证据的关注度，并更好保留统一学生模型的专项能力。代码与数据已公开。
正文 · 原文
Abstract:Reinforcement learning (RL) can elicit specialized visual capabilities in vision-language models (VLMs), yet jointly optimizing heterogeneous objectives often introduces cross-capability gradient conflicts. Training experts independently mitigates this interference, but direct reward optimization may entangle specialized knowledge with reward-hacking biases. This paper proposes Skill-Decoupled On-Policy Expert Distillation (SCOPED), which trains capability-specific experts and extracts explicit skills through reward-guided reflection. During on-policy distillation, these skills guide expert supervision while remaining unavailable to the unified student, encouraging it to internalize procedural knowledge without explicit skills at inference time. Across eight multimodal benchmarks, SCOPED consistently outperforms the backbone and mixed-capability RL, demonstrating the effectiveness of skill-decoupled on-policy expert distillation for integrating independently acquired capabilities. Further analyses reveal interference during joint capability optimization and suggest that skill augmentation promotes more consistent expert trajectories, greater attention to task-relevant visual evidence, and better preservation of specialized expertise in the unified student. All code and data are available at this https URL.
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.22217 [cs.CV] |
| (or arXiv:2610.22217v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2610.22217 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Zhipeng Xu [view email]
[v1]
Thu, 15 Oct 2026 17:01:41 UTC (3,785 KB)
来源：arXiv cs.AI · arxiv.org