arXiv Artificial Intelligence· Hongxing Li, Yixin Li, Dingming Li, Zixuan Wang, Yuchen Yan, Wenqi Zhang, Weiming Lu, Yongliang Shen·· 19 小时前AI 评分24
GPD:用路由式 3D 特权蒸馏提升视觉语言模型空间推理
Distilling Routed 3D Privilege for Spatial Reasoning in Vision-Language Models
AI 导读
GPD(Geometry-Privileged Distillation)把深度、语义和鸟瞰图(BEV)线索渲染为紧凑文本,与参考答案一起路由给教师模型,仅对错误轨迹施加特权 KL 来增强 GRPO,部署模型仍只用 RGB 输入。
来源:arXiv Artificial Intelligence · arxiv.org