跳转至

文章背景与核心概要

随着大型语言模型（LLM）越来越多地被部署并用于访问庞大的工具库、函数、API和专用代理，它们面临着巨大的性能障碍。不断增长的候选动作空间迫使模型处理更多的架构模式、消耗过多的提示词Token，并且难以在相似或不相关的候选工具之间进行区分。

为了解决这一问题，作者推出了 AgentWeave，这是一个确定性的推理前路由层。AgentWeave 没有修改下游的语言模型，而是通过利用适用性、需求性、能力和路由信号，预先构建出一个有边界且模型可见的动作空间，从而精简候选工具集。

在使用公开的 MadeAgents/Hammer2.1-1.5b 模型、基于固定的 BFCL 衍生路由压力协议进行评估时： * 成功率： 在 48 个新的多函数任务中，AgentWeave 取得了 12.5% (6/48) 的原生 BFCL 成功率，而全工具（all-tools）、确定性随机前 8（deterministic random top-8）以及语义前 8（semantic top-8）基线的成功率均为 0% (0/48)（成对成功率差异为 +12.5 个百分点；95% 自助法置信区间：+4.17 至 +22.92 个百分点；精确 McNemar 检验 \(p = 0.03125\)）。 * 效率： 与暴露全工具相比，AgentWeave 减少了 70.18% 的工具，消耗了 61.70% 的输入 Token，并实现了 50.95% 的更低本地模型平均延迟。

尽管绝对任务成功率仍然较低，且本研究主要作为基于 BFCL 的路由压力评估，而非官方排行榜评分，但这些发现表明，候选空间的构建对固定模型的行为具有实质性影响，并将推理前路由确立为实现高效工具使用的关键阶段。

论文元数据 (Paper Metadata)

  • arXiv ID: arXiv:2608.23078 [cs.AI]
  • 学科分类: 人工智能 (cs.AI)；计算与语言 (cs.CL)
  • ACM 类别: I.2.11; I.2.7
  • DOI: 10.48550/arXiv.2608.23078
  • 提交时间: 2026年8月24日
  • 作者: Saurav Singla, Aarav Singla, Advik Gupta, Parnika Gupta

AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models

AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models

license icon

license icon

Summary

Summary

As large language models (LLMs) are increasingly deployed with access to vast collections of tools, functions, APIs, and specialized agents, they face significant performance hurdles. A growing candidate action space forces models to process more schemas, consume excessive prompt tokens, and struggle to distinguish between similar or irrelevant alternatives.

As large language models (LLMs) are increasingly deployed with access to vast collections of tools, functions, APIs, and specialized agents, they face significant performance hurdles. A growing candidate action space forces models to process more schemas, consume excessive prompt tokens, and struggle to distinguish between similar or irrelevant alternatives.

To combat this, the authors introduce AgentWeave, a deterministic pre-inference routing layer. Instead of modifying the downstream language model, AgentWeave reduces the candidate set beforehand by building a bounded, model-visible action space using eligibility, requirement, capability, and routing signals.

To combat this, the authors introduce AgentWeave, a deterministic pre-inference routing layer. Instead of modifying the downstream language model, AgentWeave reduces the candidate set beforehand by building a bounded, model-visible action space using eligibility, requirement, capability, and routing signals.

Evaluated using a frozen BFCL-derived routing-pressure protocol on the public MadeAgents/Hammer2.1-1.5b model: * Success Rate: AgentWeave achieved 12.5% (6/48) native BFCL successes on 48 fresh multiple-function tasks, compared to 0% (0/48) for all-tools, deterministic random top-8, and semantic top-8 baselines (paired success difference of +12.5 percentage points; 95% bootstrap CI: +4.17 to +22.92 points; exact McNemar \(p = 0.03125\)). * Efficiency: Relative to full all-tools exposure, AgentWeave presented 70.18% fewer tools, consumed 61.70% fewer input tokens, and achieved 50.95% lower mean local-model latency.

Evaluated using a frozen BFCL-derived routing-pressure protocol on the public MadeAgents/Hammer2.1-1.5b model: * Success Rate: AgentWeave achieved 12.5% (6/48) native BFCL successes on 48 fresh multiple-function tasks, compared to 0% (0/48) for all-tools, deterministic random top-8, and semantic top-8 baselines (paired success difference of +12.5 percentage points; 95% bootstrap CI: +4.17 to +22.92 points; exact McNemar \(p = 0.03125\)). * Efficiency: Relative to full all-tools exposure, AgentWeave presented 70.18% fewer tools, consumed 61.70% fewer input tokens, and achieved 50.95% lower mean local-model latency.

While absolute task success remains low and the study serves primarily as a BFCL-derived routing-pressure evaluation rather than an official leaderboard score, the findings demonstrate that candidate-space construction materially impacts fixed-model behavior and establishes pre-inference routing as a critical stage for efficient tool use.

While absolute task success remains low and the study serves primarily as a BFCL-derived routing-pressure evaluation rather than an official leaderboard score, the findings demonstrate that candidate-space construction materially impacts fixed-model behavior and establishes pre-inference routing as a critical stage for efficient tool use.

Paper Metadata

Paper Metadata

  • arXiv ID: arXiv:2608.23078 [cs.AI]
  • Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
  • ACM Classes: I.2.11; I.2.7
  • DOI: 10.48550/arXiv.2608.23078
  • Submitted: August 24, 2026
  • Authors: Saurav Singla, Aarav Singla, Advik Gupta, Parnika Gupta
  • arXiv ID: arXiv:2608.23078 [cs.AI]
  • Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
  • ACM Classes: I.2.11; I.2.7
  • DOI: 10.48550/arXiv.2608.23078
  • Submitted: August 24, 2026
  • Authors: Saurav Singla, Aarav Singla, Advik Gupta, Parnika Gupta