arXiv · cs.AI · Vimal William, Yuanzhe Li, Danella Zhao, Huanrui Yang2026-11-04 13:00 · 4 小时前模型评分43

Anchor：用锚点层与跨层循环状态提升长上下文检索的测试时训练方法

Anchor: Augmented Cross-Layer Recurrent State for Long-Context Retrieval in Test-Time Training

导读

研究提出 Anchor，一种面向测试时训练（TTT，Test-Time Training）的混合注意力方法，只在少数锚点层放置循环记忆，并让多层共享一份跨层循环状态、各层通过自己的读取映射使用它。

分类置信度

中 · 待人工复核

来源级别：媒体与个人

事件类型：研究发布

证据摘录

“We exploit both observations in Anchor: TTT memory placed only at the anchor layers, and a cross-layer recurrent state written once and read by many layers, each through its own learned read map that realizes a layer-specific stretch and rotation, along with demonstrated generalization to other hybrid attentions.”

来源：arXiv · cs.AI · arxiv.org