RIG-RoPE：关系与实例门控的旋转位置编码
原标题：RIG-RoPE: Relation- and Instance-Gated Rotary Positional Encoding with Duration-Aware Temporal Coordinates
AI 摘要
该论文提出 RIG-RoPE，一种关系与实例门控的旋转位置编码机制，用于改进多模态大语言模型中的位置编码。它通过模态指示符、视觉实例标识符和持续时间坐标，解决静态多维位置分配中的跨模态空间干扰和时间步长不均问题。该方法无需新增参数，可在平铺注意力内核中实现，但尚未声称经验上的优越性。
正文节选
Computer Science > Computation and Language Title:RIG-RoPE: Relation- and Instance-Gated Rotary Positional Encoding with Duration-Aware Temporal Coordinates View PDF HTML (experimental) Abstract:Rotary positional encoding (RoPE) is a core component of modern language models and has been extended to multimodal LLMs through multidimensional variants such as multimodal RoPE (M-RoPE), which split positional channels into temporal, height, and width subspaces. This report identifies two l