跳转至

文章背景与核心概要

大语言模型(LLMs)越来越依赖外部工具来检索最新数据、执行繁重计算以及与外部系统交互。传统的自回归模型以顺序方式处理工具调用(即发出调用、暂停等待响应、然后恢复生成),而扩散语言模型(dLLMs)则是通过在其输出的多个部分进行并行的迭代精炼来运作的。

将传统的“停止-恢复”交互模型应用于dLLMs具有局限性且效率低下。它可能会在模型的内部推理稳定之前迫使其做出过早的工具决策、延迟有用的观察结果,并触发冗余的工具执行。

为了解决这种不匹配,本文引入了连续交互扩散(Continuous Interaction Diffusion, CID),这是一种扩散原生的模型-运行时架构,它将工具交互直接集成到了迭代去噪过程中。


连续交互扩散:面向异步工具增强推理的扩散原生运行时

作者: Yuhang Cao
提交时间: 2026年8月11日
研究领域: 人工智能 (cs.AI)
arXiv ID: 2608.10438
DOI: 10.48550/arXiv.2608.10438


摘要

Large language models (LLMs) increasingly rely on external tools to retrieve up-to-date data, perform heavy computations, and interact with external systems. While traditional autoregressive models handle tool use sequentially (emitting a call, halting to wait for the response, and then resuming generation), diffusion language models (dLLMs) operate through parallel, iterative refinement across multiple parts of their output.

大语言模型(LLMs)越来越依赖外部工具来检索最新数据、执行繁重计算以及与外部系统交互。传统的自回归模型以顺序方式处理工具调用(即发出调用、暂停等待响应、然后恢复生成),而扩散语言模型(dLLMs)则是通过在其输出的多个部分进行并行的迭代精炼来运作的。

Applying the traditional stop-and-resume interaction model to dLLMs is restrictive and inefficient. It can force premature tool decisions before the model's internal reasoning stabilizes, delay useful observations, and trigger redundant tool executions.

将传统的“停止-恢复”交互模型应用于dLLMs具有局限性且效率低下。它可能会在模型的内部推理稳定之前迫使其做出过早的工具决策、延迟有用的观察结果,并触发冗余的工具执行。

To resolve this mismatch, this paper introduces Continuous Interaction Diffusion (CID), a diffusion-native model-runtime architecture that integrates tool interactions directly into the iterative denoising process.

为了解决这种不匹配,本文引入了连续交互扩散(Continuous Interaction Diffusion, CID),这是一种扩散原生的模型-运行时架构,它将工具交互直接集成到了迭代去噪过程中。


CID 的核心架构亮点

  • Tri-Channel Separation: CID divides processing streams into:
  • A model-read-only fact channel
  • A thought channel represented by a Typed Cognitive Tensor
  • A display channel
  • 三通道分离: CID 将处理流划分为:
  • 一个模型只读的事实通道
  • 一个由类型化认知张量(Typed Cognitive Tensor)表示的思考通道
  • 一个显示通道
  • Early Perceptual Binding: Information needs can be identified before text or JSON tool calls are fully serialized. This allows external reads to launch concurrently while the denoising process continues.
  • 早期感知绑定: 在文本或 JSON 工具调用完全序列化之前,就可以识别出信息需求。这允许外部读取操作在去噪过程持续进行的同时并发启动。
  • Dynamic State Integration: Returned results are projected directly into the evolving thought state, enabling the model to dynamically revise earlier cognitive steps and display regions.
  • 动态状态集成: 返回的结果被直接投影到演进中的思考状态中,使模型能够动态修正先前的认知步骤和显示区域。
  • Persistent Bindings: Static results can be cached and reused to prevent redundant external executions, while dynamic data sources can be refreshed on demand.
  • 持久化绑定: 静态结果可以被缓存和复用以防止冗余的外部执行,而动态数据源则可以按需刷新。
  • Optimized Efficiency: CID is engineered to expose evidence earlier, overlap tool latency with model computation, minimize duplicate work, and preserve useful reasoning paths when new evidence is introduced.
  • 优化效率: CID 旨在更早地暴露证据、将工具延迟与模型计算重叠、最小化重复工作,并在引入新证据时保留有用的推理路径。

文档详情与全文链接

  • Format / Length: 15 pages, 5 figures, 1 table
  • 格式 / 篇幅: 15 页,5 个图表,1 个表格

(Note: This initial paper formally establishes the architecture, runtime, and training objectives along with evaluation protocols for task quality and end-to-end efficiency, focusing exclusively on read-only tools without making empirical performance claims yet.)

(注:这篇初始论文正式确立了架构、运行时和训练目标,以及任务质量和端到端效率的评估协议,目前仅专注于只读工具,暂未提出经验性能声明。)