文章背景与核心概要
随着多厂商、自主化的大语言模型(LLM)智能体开始在开放的“智能体网络(Agentic Web)”中协作,如何安全地共享工具选择与路由知识(即何时调用哪个工具、如何调用)成了一个亟待解决的关键挑战。传统的扁平化文本(如提示词、示例池)以及模型权重或适配器,在跨异构平台进行安全知识传输时往往存在局限性与隐私风险。
为了克服这些局限,本文作者引入了类型化联邦构件(typed federated artifacts)——这是一种经过模式验证的对象,具有结构清晰的字段,专为实现字段级隐私保护、争议解决和跨模型迁移而设计。他们将这一概念实例化为 SYNAPSE1,即一个用于工具路由知识的通用框架。研究表明,通过规范的数据清洗和类型化字段设计,联邦知识汇编(compendium)的路由性能能够逼近中心化方案,同时在处理矛盾信息和跨模型评测中展现出显著优势。此外,作者也指出了当前标准评测基准中存在的数据泄露与检索局限等警示性发现。
Typed Federated Artifacts for the Agentic Web: Sharing Tool-Routing Knowledge Across Frozen, Heterogeneous LLM Agents
面向智能体网络的类型化联邦构件:在冻结的异构LLM智能体间共享工具路由知识
arXiv ID: 2609.06815 [cs.CL]
DOI: 10.48550/arXiv.2609.06815
Submitted: September 6, 2026
Authors: Abhijit Chakraborty, Ni Trieu, Vivek Gupta
Summary
摘要
The paper explores the challenges of building an open, networked "agentic web" where autonomous Large Language Model (LLM) agents from multiple vendors must collaborate, maintain privacy, and share knowledge about tool selection and routing (i.e., which tool to call and when). 本文探讨了构建开放且互联的“智能体网络”所面临的挑战。在该网络中,来自多个厂商的自主大语言模型(LLM)智能体必须相互协作、保持隐私,并共享有关工具选择和路由(即何时调用哪个工具)的知识。
To address the limitations of flat text (prompts, example pools) and model weights/adapters in transferring routing knowledge safely across heterogeneous platforms, the authors introduce typed federated artifacts—schema-validated objects featuring well-defined fields designed for per-field privacy, dispute resolution, and cross-model transfer. They instantiate this concept as SYNAPSE1, a common framework for tool-routing knowledge. 为了解决扁平化文本(提示词、示例池)以及模型权重/适配器在跨异构平台安全传输路由知识时的局限性,作者引入了类型化联邦构件(typed federated artifacts)——这是一种经过模式验证(schema-validated)的对象,具有定义明确的字段,专为实现字段级隐私、争议解决和跨模型迁移而设计。他们将这一概念实例化为 SYNAPSE1,即一个通用的工具路由知识框架。
Key Contributions & Findings:
核心贡献与发现:
- Data Cleanup: Removing 192 garbage entries and 1,916 training items that overlapped with test queries revealed that a federated compendium can route within 1.1 points of a centralized one, consuming around 20 MB of JSON per client per round on StableToolBench (3,180 tools).
数据清洗: 在 StableToolBench(包含 3,180 个工具)上,通过移除 192 个垃圾条目以及与测试查询重叠的 1,916 个训练样本发现,联邦知识汇编(compendium)的路由性能可以达到仅比中心化方案落后 1.1 个百分点的水平,每个客户端每轮消耗约 20 MB 的 JSON 数据。
Format vs. Flat Strings: Presenting shared experience to the router as typed fields rather than a flat string yields significant performance gains (8.5 points on clean data and 7.4 under 60% injected contradiction). Furthermore, combining the typed merge with proper rendering proves that the two halves are inseparable.
格式对比与扁平字符串: 将共享经验作为类型化字段而非扁平字符串提供给路由器,可带来显著的性能提升(在干净数据上提升 8.5 个百分点,在注入 60% 矛盾数据的情况下提升 7.4 个百分点)。此外,将类型化合并与适当的渲染结合使用,证明了这两部分密不可分。
GPT-4o Evaluation: On the \(\tau\)-bench retail benchmark, every compendium arm improves GPT-4o agents' per-step tool-call accuracy by at least 6.7 points, though this is primarily attributed to output format rather than federated experience itself.
GPT-4o 评估: 在 \(\tau\)-bench 零售基准测试中,所有知识汇编分支都使 GPT-4o 智能体的单步工具调用准确率提高了至少 6.7 个百分点,不过这主要归功于输出格式,而非联邦经验本身。
Cautionary Findings: On a topic-labeled math proxy and StableToolBench, a simple TF-IDF classifier over the labeled experience outperforms every LLM routing arm (by 48 and 26 points, largely due to retrieval recall). The authors note that standard benchmarks frequently contain labeled queries for supposedly "unseen" tools and test queries verbatim prior to filtering, preventing accurate measurement of true zero-shot tool routing.
- 警示性发现: 在带主题标签的数学代理和 StableToolBench 上,针对带标签经验的简单 TF-IDF 分类器表现优于所有 LLM 路由分支(分别高出 48 和 26 个百分点,主要由于检索召回率的优势)。作者指出,标准基准测试经常包含针对所谓“未见”工具的带标签查询,并且在过滤前直接包含测试查询,这阻碍了对真正零样本(zero-shot)工具路由的准确测量。
Metadata & Additional Links
元数据与其他链接
- Primary Subject: Computation and Language (
cs.CL)- Secondary Subject: Artificial Intelligence (
cs.AI)- Full-Text & Source Access:
- View PDF
- HTML Version
- TeX Source
- License: Creative Commons Attribution 4.0 International
- 主要学科: 计算与语言 (
cs.CL)- 次要学科: 人工智能 (
cs.AI)- 全文与源码访问:
- 查看 PDF
- HTML 版本
- TeX 源码
- 许可证: 知识共享署名 4.0 国际许可协议 (Creative Commons Attribution 4.0 International)
