Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Adaptive, Scalable, and Robust Coordination of LLM Agents: A Dynamic Ad-Hoc Networking Perspective

Rui Li, Zeyu Zhang, Xiaohe Bo, Quanyu Dai, Chaozhuo Li, Feng Wen, Xu Chen

TL;DR

This work reframes LLM-based multi-agent coordination as dynamic ad-hoc networking and introduces RAPS, a decentralized reputation-aware publish-subscribe framework. RAPS decouples agents via a content-centric substrate and augments it with two overlays: Reactive Subscription for online intent refinement and Bayesian Reputation for decentralized trust. Empirical results across five benchmarks show state-of-the-art performance, strong scalability with larger agent pools, and robust resilience to adversarial agents. The approach offers a practical, training-free pathway to open, self-organizing, and trustworthy multi-agent systems with broad applicability to adaptive collaborative tasks.

Abstract

Multi-agent architectures built on large language models (LLMs) have demonstrated the potential to realize swarm intelligence through well-crafted collaboration. However, the substantial burden of manual orchestration inherently raises an imperative to automate the design of agentic workflows. We frame such an agent coordination challenge as a classic problem in dynamic ad-hoc networking: How to establish adaptive and reliable communication among a scalable number of agentic hosts? In response to this unresolved dilemma, we introduce RAPS, a reputation-aware publish-subscribe paradigm for adaptive, scalable, and robust coordination of LLM agents. RAPS is grounded in the Distributed Publish-Subscribe Protocol, allowing LLM agents to exchange messages based on their declared intents rather than predefined topologies. Beyond this substrate, RAPS further incorporates two coherent overlays: (i) Reactive Subscription, enabling agents to dynamically refine their intents; and (ii) Bayesian Reputation, empowering each agent with a local watchdog to detect and isolate malicious peers. Extensive experiments over five benchmarks showcase that our design effectively reconciles adaptivity, scalability, and robustness in a unified multi-agent coordination framework.

Towards Adaptive, Scalable, and Robust Coordination of LLM Agents: A Dynamic Ad-Hoc Networking Perspective

TL;DR

This work reframes LLM-based multi-agent coordination as dynamic ad-hoc networking and introduces RAPS, a decentralized reputation-aware publish-subscribe framework. RAPS decouples agents via a content-centric substrate and augments it with two overlays: Reactive Subscription for online intent refinement and Bayesian Reputation for decentralized trust. Empirical results across five benchmarks show state-of-the-art performance, strong scalability with larger agent pools, and robust resilience to adversarial agents. The approach offers a practical, training-free pathway to open, self-organizing, and trustworthy multi-agent systems with broad applicability to adaptive collaborative tasks.

Abstract

Multi-agent architectures built on large language models (LLMs) have demonstrated the potential to realize swarm intelligence through well-crafted collaboration. However, the substantial burden of manual orchestration inherently raises an imperative to automate the design of agentic workflows. We frame such an agent coordination challenge as a classic problem in dynamic ad-hoc networking: How to establish adaptive and reliable communication among a scalable number of agentic hosts? In response to this unresolved dilemma, we introduce RAPS, a reputation-aware publish-subscribe paradigm for adaptive, scalable, and robust coordination of LLM agents. RAPS is grounded in the Distributed Publish-Subscribe Protocol, allowing LLM agents to exchange messages based on their declared intents rather than predefined topologies. Beyond this substrate, RAPS further incorporates two coherent overlays: (i) Reactive Subscription, enabling agents to dynamically refine their intents; and (ii) Bayesian Reputation, empowering each agent with a local watchdog to detect and isolate malicious peers. Extensive experiments over five benchmarks showcase that our design effectively reconciles adaptivity, scalability, and robustness in a unified multi-agent coordination framework.
Paper Structure (40 sections, 10 equations, 5 figures, 7 tables, 1 algorithm)

This paper contains 40 sections, 10 equations, 5 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. LLM Agent Coordination as Dynamic Ad-Hoc Networking
  4. Problem Formulation
  5. Networking Perspective
  6. RAPS: Reputation-Aware Publish-Subscribe Communication
  7. Overview
  8. Substrate: Publish-Subscribe Protocol
  9. Overlay I: Reactive Subscription
  10. Overlay II: Bayesian Reputation
  11. Experiments
  12. Experimental Setup
  13. Main Results
  14. Scalability and Efficiency
  15. Robustness Analysis
  16. ...and 25 more sections

Figures (5)

  • Figure 1: Illustrations of different MAS coordination paradigms.
  • Figure 2: The overall framework of RAPS. The architecture consists of three layers: (a) A distributed publish-subscribe substrate that decouples agents into publishers and subscribers for MAS coordination; (b) Reactive subscription that allows agents to adaptively refine their intents based on the message flow; and (3) Bayesian reputation that employs a local watchdog to assess peer reliability for robustness.
  • Figure 3: Scalability and Efficiency analysis with varying # agents.
  • Figure 4: Cost-Performance analysis of different MAS coordination methods. $k$ represents the communication rounds for RAPS.
  • Figure 5: Impact of agent pool quality on HumanEval benchmark.