跳转至

文章背景与核心概要

现代智能体AI系统通常部署多个具备异构技能的大语言模型(LLM)智能体。然而,当前的系统架构通常依赖于预定义的通信通道或完全广播拓扑。随着活跃智能体数量和通信链路的增加,这两种方法在Token成本、执行延迟、信息冗余以及错误传播方面都表现出极差的可扩展性。

为了解决这一瓶颈,本文将智能体选择和通信建模为一个由任务条件净效用函数所支配的合作博弈。通过引入边际价值激活规则与贪婪路由器,该研究能够动态选择最优的智能体子集,并优化具有独立边际成本的通信链路。实验表明,该方法在大幅减少激活智能体数量的同时,能够保持接近最优的效用,为构建高效、可扩展的多智能体AI系统提供了重要的理论与实践基础。


基于技能的智能体AI系统中的动态联盟形成与通信定价 (Dynamic Coalition Formation and Communication Pricing in Skill-Based Agentic AI Systems)

作者: Mojtaba Eslami
主题: 人工智能 (cs.AI); 机器学习 (cs.LG); 理论经济学 (econ.TH)
引用: arXiv:2608.07532 [cs.AI] | DOI: 10.48550/arXiv.2608.07532
提交日期: 2026年7月24日


摘要 (Summary)

Modern agentic AI systems deploy multiple Large Language Model (LLM) agents possessing heterogeneous skills. However, current system architectures typically rely on either predefined communication channels or complete broadcast topologies. Both approaches are inherently inefficient, as token costs, execution latency, information redundancy, and error propagation scale poorly with the number of active agents and communication links.

现代智能体AI系统部署了多个具备异构技能的大语言模型(LLM)智能体。然而,当前的系统架构通常依赖于预定义的通信通道或完全广播拓扑。这两种方法在本质上都是低效的,因为Token成本、执行延迟、信息冗余以及错误传播随着活跃智能体数量和通信链路的增加,其扩展性极差。

To resolve this bottleneck, this paper models agent selection and communication as a cooperative game governed by a task-conditioned net utility function:

\[U(C\mid x) = V(C\mid x) - \sum_{i\in C} c_i\]

This formulation clearly separates coalition-level performance values (\(V\)) from individual agent activation costs (\(c_i\)).

为了解决这一瓶颈,本文将智能体选择和通信建模为一个合作博弈,该博弈受任务条件净效用函数控制:

\[U(C\mid x) = V(C\mid x) - \sum_{i\in C} c_i\]

这种公式化表达清晰地将联盟级别的性能价值(\(V\))与个体智能体的激活成本(\(c_i\))区分开来。


核心贡献与方法论 (Key Contributions & Methodology)

  • Algorithmic Routing: Introduces a marginal-value activation rule alongside a greedy router to dynamically select optimal agent subsets.
  • Granular Edge Optimization: Extends the framework to optimize communication edges with distinct, per-edge costs.
  • Predictive Shapley Valuation: Leverages estimated Shapley values to evaluate which agents are essential to contact prior to and during task execution.
  • Theoretical Guarantees:
  • Connects the routing problem to submodular maximization.
  • Proves a curvature-refined bound for a monotone, cardinality-constrained special case.
  • Establishes a tight \(\frac{1}{2}\)-approximation via double greedy methods for an unconstrained non-monotone case (including a correction for signed objectives). Note: These theoretical guarantees do not apply directly to the main router, which functions as a heuristic.
  • Derives a Shapley-submodularity sandwich bound linking the routing error of marginal-value selection to a per-agent diminishing-returns quantity.
  • 算法路由: 引入了边际价值激活规则以及贪婪路由器,以动态选择最优的智能体子集。
  • 细粒度边缘优化: 扩展了该框架,以优化具有不同、每条边独立成本的通信边缘。
  • 预测性Shapley估值: 利用估计的Shapley值来评估在任务执行前和执行期间哪些智能体是必须联系的。
  • 理论保证:
  • 将路由问题与次模最大化(submodular maximization)联系起来。
  • 证明了针对单调、基数约束特殊情况的曲率精细化界限。
  • 通过双贪婪方法为无约束非单调情况(包括对有符号目标的修正)建立了紧密的 \(\frac{1}{2}\)-近似。注:这些理论保证并不直接适用于作为启发式算法运行的主路由器。
  • 推导出了Shapley-次模夹逼界(sandwich bound),将边际价值选择的路由误差与每智能体收益递减量联系起来。

实验结果 (Experimental Results)

In synthetic evaluations: - Efficiency: Greedy routing achieves 99.5% of the brute-force-optimal utility while activating an average of only 1.96 out of 8 agents, compared to just 38.8% utility retention for full broadcast approaches. - Robustness: Performance remains stable across varying activation costs and redundancy weights. - Failure Modes: Efficiency drops to 66% when submodularity assumptions are heavily violated or when value estimates incorporate significant noise.

在合成评估中: - 效率: 贪婪路由实现了暴力最优效用的 99.5%,同时平均只激活了 8个智能体中的 1.96个,相比之下,完全广播方法的效用保留率仅为 38.8%。 - 鲁棒性: 在不同的激活成本和冗余权重下,性能保持稳定。 - 失效模式: 当严重违反次模假设或价值估计包含显着噪声时,效率会下降到 66%


区别与未来方向 (Distinction & Future Directions)

The authors carefully differentiate this framework from conventional Shapley pricing, hedonic coalition formation, and basic communication-graph pruning. Future work emphasizes empirical evaluation on real-world multi-agent LLM benchmarks.

作者仔细将该框架与传统的Shapley定价、享乐联盟形成(hedonic coalition formation)以及基础通信图剪枝区分开来。未来的工作将着重于在现实世界的多智能体LLM基准上进行实证评估。