面向大语言模型智能体工作流中并行分支的直接隐空间合成框架
英文摘要
Shikun Liu, Mufei Li, Dongqi Fu, Haoyu Wang, Yinglong Xia, Hong Li, Hong Yan, and Pan Li propose a framework that synthesizes latent representations directly to enable parallel branches in LLM-agent workflows. This method reduces the computational overhead of orchestrating multiple LLMs by avoiding explicit token-level communication, instead fusing latent-space paths for simultaneous execution. The approach improves responsiveness and scalability for complex, multi-agent tasks. The paper demonstrates how latent-space synthesis can redefine collaboration among LLMs in automated decision-making and content generation systems.
中文摘要
Shikun Liu、Mufei Li、Dongqi Fu、Haoyu Wang、Yinglong Xia、Hong Li、Hong Yan 和 Pan Li 提出一种通过直接合成隐表征来实现大语言模型智能体工作流中并行分支的框架。该方法通过融合隐空间路径进行同步执行，避免了显式 token 级通信，从而降低了多个大语言模型协同的计算开销。该方案提升了复杂多智能体任务的响应速度和可扩展性，并展示了隐空间合成如何重塑自动化决策和内容生成系统中 LLM 的协作方式。
关键要点
Introduces a framework for direct latent-space synthesis that allows parallel execution of branches in LLM-agent workflows.
提出一种直接隐空间合成框架，允许在 LLM 智能体工作流中并行执行多个分支。
Reduces computational overhead by fusing latent representations instead of managing explicit text-level communication between LLMs.
通过融合隐表征代替协调 LLM 之间显式的文本通信，降低计算开销。
Enhances scalability and responsiveness for complex multi-agent tasks such as automated content generation and decision-making systems.
提升复杂多智能体任务（如自动化内容生成和决策系统）的可扩展性与响应速度。
Authored by Shikun Liu, Mufei Li, Dongqi Fu, Haoyu Wang, Yinglong Xia, Hong Li, Hong Yan, Pan Li; paper available on arXiv.
作者包括 Shikun Liu、Mufei Li、Dongqi Fu 等；论文发布于 arXiv。