PulseAugur
实时 14:56:22

DART架构通过融合Transformer和SSM来增强长上下文序列建模

研究人员推出了一种新颖的架构DART(Decoded Attention over Recurrent States),它结合了Transformer和状态空间模型(SSMs)的优势,用于高效的长上下文序列建模。DART基于Mamba-2,通过从SSM的压缩状态解码token条件化的键和值来实现状态-记忆注意力(SMA)。与传统的注意力机制相比,这种方法显著降低了推理缓存需求,并在保持语言建模质量的同时增强了联想回忆和检索能力。 AI

影响 增强了长上下文建模的效率和检索能力,可能提高复杂NLP任务的性能。

排序理由 该集群描述了一篇详细介绍序列建模新颖架构的研究论文。

在 arXiv cs.LG 阅读 →

AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →

DART架构通过融合Transformer和SSM来增强长上下文序列建模

报道来源 [2]

  1. arXiv cs.LG TIER_1 English(EN) · Yixiao Qian, Song Chen, Pengkai Wang, Jiaxu Liu, Shengze Cai, Chao Xu ·

    DART:用于高效长上下文序列建模的循环状态解码注意力

    arXiv:2608.02032v1 Announce Type: new Abstract: Modern language models are built primarily from Transformers, recurrent models, and their hybrid architectures. Transformers rely on token-level attention memories, while recurrent models such as state space models (SSMs) and linear…

  2. Hugging Face Daily Papers TIER_1 English(EN) ·

    DART:用于高效长上下文序列建模的循环状态解码注意力

    Modern language models are built primarily from Transformers, recurrent models, and their hybrid architectures. Transformers rely on token-level attention memories, while recurrent models such as state space models (SSMs) and linear attention maintain compact recurrent states. Th…