DAIR.AI推文
亚马逊新工作：CTIFoundry 在索引时构建知识结构，提升智能体性能
亚马逊及同事提出 CTIFoundry，在索引时构建知识结构，将四个权威安全知识库的官方交叉引用转化为类型化可遍历边，并添加跨度级报告层保留来源。在开源 harness 上，仅更换动作表面，四模型整体 F1 提升 0.19-0.28，小模型在脚手架上的表现优于旗舰模型在平面基座上的表现，且工具调用减半。
译文
亚马逊及同事们非常有趣的新工作。（请收藏）这与将语料库视为智能体框架一部分的新兴主题相呼应。规划循环、工具协议和上下文管理已迅速成熟，而智能体所调研的知识仍以不透明的分块形式存在于嵌入索引之后。CTIFoundry 转而将结构构建在索引阶段。四个权威安全知识库之间的官方交叉引用成为类型化的可遍历边，而基于跨度锚定的报告层则确保每个分块都附带来源信息。七个类型化工具和三个程序性技能在标准开源框架上暴露了这种结构。仅更换操作表面，就在四模型面板上使相同框架的智能体整体 F1 提升了 0.19 至 0.28。在该脚手架上的小模型在平面基底上以大约一半的工具调用次数击败了旗舰模型。论文：https://arxiv.org/abs/2608.18613 在我们的学院中追踪更多热门 AI 论文：https://academy.dair.ai/
Very interesting new work Amazon and colleagues. (bookmark it) This connects to the emerging theme of treating the corpus as part of the harness. Planning loops, tool protocols, and context management have matured fast, while the knowledge an agent investigates over still sits behind an embedding index as opaque chunks. CTIFoundry builds structure at index time instead. Official cross-references across four authoritative security knowledge bases become typed traversable edges, and a span-grounded report layer keeps provenance attached to every chunk. Seven typed tools and three procedural skills expose that structure on a stock open-source harness. Swapping only the action surface lifts the identically-harnessed agent by 0.19 to 0.28 overall F1 across a four-model panel. A small model on the scaffold beats a flagship on the flat substrate at roughly half the tool calls. Paper: https://arxiv.org/abs/2608.18613 Track more trending AI papers in our academy: https://academy.dair.ai/
