Rethinking Multi-Agent Intelligence Through the Lens of Small-World Networks

Rethinking Multi-Agent Intelligence Through the Lens of Small-World Networks Boxuan Wang, Zhuoyun Li, Xiaowei Huang, Yi Dong∗ School of Computer Science and Informatics, University of Liverpool Liverpool, United Kingdom Abstract Large language models (LLMs) have enabled multi-agent systems (MAS) in which multiple agents argue, critique, and coordinate to solve complex tasks, making communication topology a first-class design choice. Yet most existing LLM-based MAS either adopt fully connected graphs, simple sparse rings, or ad-hoc dynamic selection, with little structural guidance. In this work, we revisit classic theory on small-world (SW) networks and ask: what changes if we treat SW connectivity as a design prior for MAS? We first bridge insights from neuroscience and complex networks to MAS, highlighting how SW structures balance local clustering and long-range integration. Using multi-agent debate (MAD) as a controlled testbed, experi- ment results show that SW connectivity yields nearly the same accuracy and token cost, while substantially stabilizing consensus trajectories. Building on this, we introduce an uncertainty-guided rewiring scheme for scaling MAS, where long-range shortcuts are added between epistemically divergent agents using LLM-oriented uncertainty signals (e.g., semantic entropy). This yields controllable SW structures that adapt to task difficulty and agent heterogene- ity. Finally, we discuss broader implications of SW priors for MAS design, framing them as stabilizers of reasoning, enhancers of ro- bustness, scalable coordinators, and inductive biases for emergent cognitive roles. Keywords Multi-Agent Systems, Small-World Network, Multi-Agent Debate, Large Language Models 1 Introduction Large language models (LLMs) have enabled a new generation of multi-agent systems (MAS) in which multiple LLM agents inter- act, critique, and collaborate to solve complex tasks. Frameworks such as AutoGen [33] and CAMEL [14] demonstrate that organiz- ing agents into conversational roles can enhance reasoning, while benchmarks like AgentBench [17] reveal persistent limitations in long-horizon planning, coordination, and self-correction. These observations underscore that the structure of inter-agent communi- cation is becoming a central design axis for multi-agent intelligence. A growing line of work investigates how communication topol- ogy affects collective reasoning. In multi-agent debate (MAD), a widely studied testbed for collaborative reasoning, agents itera- tively exchange arguments and refine their answers [6, 13, 16, 18]. Although early MAD implementations commonly adopted a fully- connected topology, recent work challenges this assumption. Li et al. [15] show that even a simple ring topology, where each agent ∗Corresponding Author: Yi.Dong@liverpool.ac.uk Preprint, Under Review P=0 (Regular Lattice) 0…(Full text truncated)…

This content is AI-processed based on ArXiv data.