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

Abstract: LLM-based multi-agent systems have recently gained significant attention due to their potential for complex, collaborative, and intelligent problem-solving capabilities. Existing surveys typically categorize LLM-based multi-agent systems (LLM-MAS) according to their application domains or architectures, overlooking the central role of communication in coordinating agent behaviors and interactions. To address this gap, this paper presents a comprehensive survey of LLM-MAS from a communication-centric perspective. Specifically, we propose a structured framework that integrates system-level communication (architecture, goals, and protocols) with system internal communication (strategies, paradigms, objects, and content), enabling a detailed exploration of how agents interact, negotiate, and achieve collective intelligence. Through an extensive analysis of recent literature, we identify key components in multiple dimensions and summarize their strengths and limitations. In addition, we highlight current challenges, including communication efficiency, security vulnerabilities, inadequate benchmarking, and scalability issues, and outline promising future research directions. This review aims to help researchers and practitioners gain a clear understanding of the communication mechanisms in LLM-MAS, thereby facilitating the design and deployment of robust, scalable, and secure multi-agent systems.

• The paper establishes a framework for dissecting system-level and internal communication, categorizing architectures and strategies in LLM-MAS.
• It details various architectures—from flat to hybrid—and communication paradigms like message passing, speech act, and blackboard models.
• It highlights challenges and opportunities including secure communication, game theory integration, and multimodal interactions for effective collaboration.

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

Introduction

This paper undertakes a comprehensive survey of LLM-based Multi-Agent Systems (LLM-MAS) from a communication perspective, focusing on system-level and internal communication mechanisms. LLMs have demonstrated exceptional capabilities in reasoning, planning, and decision-making. By integrating LLMs into MAS, agents can collaborate or compete to accomplish complex tasks beyond the scope of single-agent configurations. The research provides an in-depth examination of LLM-MAS, elucidating the architecture design, communication strategies, paradigms, objects, and content that enable collective intelligence and collaboration.

System-Level Communication

The paper establishes a framework to dissect communication in LLM-MAS into two distinct levels: system-level and internal communication.

System-Level Communication

System-level communication focuses on the architecture and overarching goals of the MAS. The study categorizes system architectures into several types, each suiting different application requirements and collaboration dynamics:

• Flat Architecture: Agents operate at the same level without a central controller, promoting flexibility in small-scale tasks but potentially leading to coordination overhead in larger setups.
• Hierarchical Architecture: Higher-level agents oversee lower-level ones, thus optimizing task allocation at the expense of potential bottlenecks at upper echelons.
• Team Architecture: Agents are grouped into specialized teams addressing facets of complex tasks, encouraging synergy but requiring significant coordination efforts.
• Society Architecture: Emulates human societal interactions, with agents adhering to norms and conventions, ideal for simulating broad social dynamics.
• Hybrid Architecture: Combines elements from other architectures for complex problem-solving, posing challenges in coordination.

  Figure 1: The macro to micro framework from a communication perspective applicable to all LLM-MAS workflows. (a) denotes system-level communication and will be discussed in Section 3. (b) denotes system internal communication and will be discussed in Section 4.

Communication goals are characterized as cooperation, competition, or a mixture of both, guiding agent interactions and decision-making processes.

System Internal Communication

The paper explores internal communication mechanisms within LLM-MAS, focusing on strategies, paradigms, objects, and content.

Communication Strategies

• One-by-One: Agents communicate sequentially, maintaining context but potentially increasing time complexity.
• Simultaneous-Talk: All agents communicate simultaneously, fostering brainstorming but requiring careful synchronization.
• Simultaneous-Talk-with-Summarizer: Introduces a summarizer to re-align global contexts, balancing parallelism with coherence.

Communication Paradigms

• Message Passing: Direct communication suitable for rapid information dissemination.
• Speech Act: Emphasizes language as a tool for performing actions, crucial in negotiations and dynamic adjustments.
• Blackboard: Offers a centralized repository for information sharing, enhancing coordination in complex tasks.

Communication Objects

Characterizes the entities interacted with during communication, including self, other agents, the environment, and humans, each influencing system dynamics.

Communication Content

Differentiates between explicit (natural language, code) and implicit communication (behavioral feedback, environmental signals), dictating how information is shared and actions are coordinated.

Challenges and Opportunities

The research identifies several challenges and prospects in optimizing LLM-MAS:

• Optimizing System Design: Developing hybrid architectures and scalable communication paradigms that balance resource allocation and communication efficiency.
• Advancing Agent Competition Research: Integrating game theory to balance competition and cooperation, fostering strategic innovation.
• Multimodal Content Communication: Incorporating multimodal data to enhance the versatility of interactions.
• Communication Security: Ensuring secure interactions to protect against malicious activities in diverse applications.
• Benchmarks and Evaluation: Developing comprehensive evaluation benchmarks to assess collaborative and interactive capabilities across multiple domains.

Conclusion

The survey proposes a nuanced framework for understanding and advancing LLM-MAS through a communication-centric lens. By dissecting system-level and internal communication constructs, the research broadens the conceptualization of LLM-MAS, identifies pivotal challenges, and motivates future exploration in enhancing scalability, security, and robustness in multi-agent settings.