Beyond Self-Talk: A Communication-Centric Survey of LLM-Based Multi-Agent Systems

Bingyu Yan, Xiaoming Zhang, Litian Zhang, Lian Zhang, Ziyi Zhou, Dezhuang Miao, Chaozhuo Li

TL;DR

This paper surveys the rapidly growing field of Large Language Model (LLM)-based Multi-Agent Systems (MAS) by focusing on their communication mechanisms. It proposes a novel communication-centric framework to analyze how agents interact, coordinate, and achieve collective intelligence, highlighting key challenges and future research directions.

What Problem Does This Paper Address?

While LLM-based MAS show immense promise for complex, collaborative tasks, existing research often overlooks the crucial role of communication in their success. Many surveys focus on architectures or application domains, leaving a gap in understanding how agents effectively communicate, negotiate, and coordinate their actions. This paper aims to fill this gap by providing a structured, communication-centric view of LLM-MAS, enabling a deeper understanding of their internal workings and facilitating the design of more robust and intelligent systems.

Methods

The paper introduces a comprehensive framework that dissects LLM-MAS communication into two levels:

  • System-Level Communication: This examines the macro-level aspects of how agents are organized and interact. It covers:
  • Communication Architectures: Flat, Hierarchical, Team, Society, and Hybrid structures.
  • Communication Goals: Cooperation (direct and debate-based), Competition, and Mixed goals.
  • Communication Protocols: Emerging standards like Model Context Protocol (MCP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP).
  • System-Internal Communication: This delves into the micro-level interactions within agents. It analyzes:
  • Communication Strategies: One-by-One, Simultaneous-Talk, and Simultaneous-Talk with Summarizer.
  • Communication Paradigms: Message Passing, Speech Act, and Blackboard.
  • Communication Objects: Communication with Self, Other Agents, Environment, and Humans.
  • Communication Content: Explicit (Natural Language, Code/Structured Data) and Implicit (Behavioral Feedback, Environmental Signal).

The authors systematically review existing literature, categorizing studies based on these communication elements and discussing their strengths and limitations.

Datasets

Not discussed in the paper. The paper is a survey and does not present new experimental results or introduce new datasets.

Key Insights and Ideas

  • Communication is King: The paper strongly emphasizes that effective communication is the bedrock of successful LLM-MAS, enabling collective intelligence and complex problem-solving that surpasses individual agent capabilities.
  • A Structured Framework is Needed: Existing surveys are too fragmented. A unified, communication-centric framework is essential to understand the interplay of various communication aspects.
  • Architecture Dictates Interaction: The choice of communication architecture (flat, hierarchical, etc.) profoundly impacts scalability, flexibility, and operational efficiency.
  • Goals Drive Behavior: Understanding why agents communicate (cooperation, competition, or a mix) is critical for shaping their interactions and achieving desired outcomes.
  • Protocols for Interoperability: Emerging protocols like MCP, A2A, and ANP are crucial for standardizing communication, enhancing security, and enabling seamless integration across diverse LLM-MAS.
  • Internal Dynamics Matter: Strategies, paradigms, objects, and content of communication within agents are key to their ability to reason, plan, and collaborate effectively.
  • The Rise of Multimodality: Future LLM-MAS will increasingly need to handle multimodal communication (text, images, audio, video) for more natural and context-aware interactions.

Key Results

As this is a survey paper, it does not present quantitative results from experiments. Instead, its key "results" are the insights derived from analyzing existing research:

  • Categorization of LLM-MAS: The paper provides a structured taxonomy of LLM-MAS based on communication architectures, goals, strategies, paradigms, objects, and content, offering a comprehensive overview of the field.
  • Identification of Trends: It highlights the growing importance of standardized communication protocols and the need for multimodal communication capabilities.
  • Summary of Strengths and Limitations: For each communication aspect, the paper outlines the advantages and disadvantages of different approaches, guiding future design choices.

Limitations and Future Work

The paper identifies several significant challenges and proposes future research directions:

  • Optimizing System Design: The need for efficient and scalable communication paradigms, especially for hybrid architectures handling increasing complexity and agent numbers. Optimizing computational resource allocation and ensuring agents can correctly interpret vast amounts of internal system information are also critical.
  • Advancing Agent Competition: Finding the optimal balance between competition and cooperation, developing scalable competition strategies, and safely integrating competition into real-world applications.
  • Unified Communication Protocol: The proliferation of diverse protocols creates redundancy and interoperability issues. Developing a unified, standardized protocol (akin to HTTPS) is crucial for seamless integration and scalability.
  • Multimodal Communication: Effectively presenting and coordinating different modalities (text, image, audio, video) in a coherent manner for agents is a key challenge. Future work should focus on improving multimodal data fusion and designing robust agents for handling this content.
  • Communication Security: Adversarial attacks (man-in-the-middle, data tampering) pose significant risks. Future research must embed robust defense mechanisms within the communication layer, including end-to-end encryption, fine-grained authentication, and adaptive key management.
  • Benchmarks and Evaluation: Existing evaluation suites are limited and often focus on individual agent performance rather than system-level properties like coordination efficiency, robustness, and emergent collaboration. A next-generation benchmark suite is needed to provide a holistic assessment of LLM-MAS.

Summary updated

Abstract

Large Language Models (LLMs) have recently demonstrated remarkable capabilities in reasoning, planning, and decision-making. Building upon these strengths, researchers have begun incorporating LLMs into multi-agent systems (MAS), where agents collaborate or compete through natural language interactions to tackle tasks beyond the scope of single-agent setups. In this survey, we present a communication-centric perspective on LLM-based multi-agent systems, examining key system-level features such as architecture design and communication goals, as well as internal mechanisms like communication strategies, paradigms, objects and content. We illustrate how these communication elements interplay to enable collective intelligence and flexible collaboration. Furthermore, we discuss prominent challenges, including scalability, security, and multimodal integration, and propose directions for future work to advance research in this emerging domain. Ultimately, this survey serves as a catalyst for further innovation, fostering more robust, scalable, and intelligent multi-agent systems across diverse application domains.
BibTeX
@misc{yan2025beyond,
  author = {Bingyu Yan and Xiaoming Zhang and Litian Zhang and Lian Zhang and Ziyi Zhou and Dezhuang Miao and Chaozhuo Li},
  title = {{Beyond Self-Talk: A Communication-Centric Survey of LLM-Based Multi-Agent Systems}},
  year = {2025},
  eprint = {2502.14321},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.14321},
}
Figure