Beyond Self-Talk: A Communication-Centric Survey of LLM-Based Multi-Agent Systems (2502.14321v2)

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.

LLM-based Multi-Agent Systems (LLM-MAS) leverage the reasoning and natural language capabilities of LLMs to enable collaboration and competition among agents, addressing tasks beyond the scope of single-agent approaches. The survey "Beyond Self-Talk: A Communication-Centric Survey of LLM-Based Multi-Agent Systems" (Yan et al., 20 Feb 2025 ) introduces a communication-centric perspective to analyze these systems, moving beyond individual agent capabilities to focus on the crucial role of inter-agent communication. It proposes a two-level framework encompassing system-level features and internal communication mechanisms.

System-Level Communication Framework

The system-level framework addresses the overarching structure and purpose of communication within an LLM-MAS.

Communication Architecture

This dimension defines the organizational structure and information flow pathways among agents. Five primary architectures are identified:

• Flat: Agents operate in a decentralized manner, communicating peer-to-peer. This structure is suitable for dynamic tasks or smaller systems where direct interaction is beneficial, such as collaborative dialogue or debate scenarios (e.g., fact-checking systems).
• Hierarchical: Agents are arranged in a tree-like structure, with higher-level agents coordinating or aggregating information from lower-level agents. This facilitates task decomposition and specialization, making it suitable for complex problem-solving (e.g., CausalGPT for reasoning, ChatDev for software development).
• Team: Agents are grouped into specialized teams based on roles or expertise, promoting complementary skill utilization. This structure is effective for tasks requiring diverse capabilities (e.g., MAGIS for GitHub issue resolution, POLCA for political negotiation simulation).
• Society: Designed to simulate larger-scale social environments, focusing on emergent behaviors, social norms, and complex interactions among a population of agents (e.g., Stanford Village for social behavior simulation, EconAgent for economic modeling).
• Hybrid: Integrates elements from multiple architectures to balance flexibility, efficiency, and specialization according to task requirements (e.g., FixAgent for debugging, ChatSim for scene simulation).

Communication Goal

This dimension defines the underlying objective driving agent interactions. Three categories are proposed:

• Cooperation: Agents collaborate towards a shared objective.
  • Direct Cooperation: Agents straightforwardly assist each other (e.g., collaborative code generation).
  • Cooperation Through Debate: Agents engage in critical dialogue or argumentation to refine solutions or reach consensus (e.g., enhancing reasoning, fact-checking).
• Competition: Agents possess conflicting goals or compete for limited resources, necessitating strategic or persuasive communication (e.g., game playing, simulating polarized debates).
• Mixed: Agents exhibit both cooperative and competitive dynamics, reflecting the complexity of many real-world scenarios (e.g., social simulations like Stanford Village, negotiation tasks like POLCA).

Internal Communication Mechanisms

Internal mechanisms detail the specifics of how, when, and what information is exchanged between agents.

Communication Strategy

This governs the timing and sequencing of agent communication:

• One-by-One: Agents communicate sequentially in a turn-taking fashion. This preserves conversational context but can introduce latency, suitable for tasks requiring strict order (e.g., Chain-of-Agents).
• Simultaneous-Talk: Agents communicate concurrently, allowing for parallel idea generation or brainstorming but risking information overload or inconsistency (e.g., AutoAgents).
• Simultaneous-Talk-with-Summarizer: Combines parallel communication with a dedicated summarizer agent to consolidate information and maintain coherence, often employed in hierarchical or team architectures (e.g., CausalGPT, AgentCoord).

Communication Paradigm

This defines the methods for representing, transmitting, and interpreting information:

• Message Passing: Direct exchange of messages (point-to-point or broadcast), typically involving rich natural language content. This is the most common paradigm.
• Speech Act: Utilizes language not just to convey information but to perform actions (e.g., instructing, requesting, persuading). This is crucial for dynamic coordination and negotiation (e.g., POLCA, ChatDev).
• Blackboard: Employs a centralized, shared repository where agents post and retrieve information. This facilitates coordination in complex systems by providing a common ground (e.g., MetaGPT).

Communication Object

This specifies the entity with which an agent interacts:

• Self: Internal monologue or self-reflection for planning, reasoning, or refining thoughts (e.g., AgentCoord's reflection mechanism).
• Other Agents: The standard form of inter-agent communication, central to collaboration and competition in nearly all LLM-MAS.
• Environment: Interaction with external surroundings, system states, or feedback mechanisms, particularly relevant for embodied agents or systems interacting with external tools (e.g., BlockAgents).
• Human: Direct interaction with human users for instructions, feedback, or collaborative task execution (e.g., PeerGPT).

Communication Content

This refers to the actual information being exchanged:

• Explicit Content: Directly conveyed information.
  • Natural Language: Common form for instructions, dialogue, and explanations.
  • Code and Structured Data: Used for precise instructions, configurations, or data exchange.
• Implicit Content: Information conveyed indirectly through actions or observations.
  • Behavioral Feedback: Inferred from the actions or strategic adjustments of other agents.
  • Environmental Signals: Interpreted from changes in the environment or system state.

Challenges in LLM-MAS Communication

Several significant challenges impede the development and deployment of robust LLM-MAS:

• Optimizing System Design: Designing scalable hybrid architectures, efficient communication paradigms, managing computational overhead, ensuring accurate information interpretation across diverse message types, and mitigating LLM hallucinations within the MAS context remain difficult.
• Advancing Research on Agent Competition: Developing frameworks that effectively balance competition and cooperation, creating scalable competitive strategies, and ensuring the safety and ethical alignment of competitive behaviors are key research gaps.
• Communicating Multimodal Content: Integrating, representing, coordinating, and communicating information across diverse modalities (text, image, audio, video) within an MAS framework presents substantial technical hurdles.
• Communication Security: Protecting inter-agent communication channels from threats like eavesdropping, tampering, and spoofing, especially in dynamic, open, or decentralized systems, requires novel security protocols tailored to LLM-MAS characteristics. Ensuring confidentiality, integrity, and authenticity is paramount.
• Benchmarks and Evaluation: The lack of comprehensive, multi-domain benchmarks and standardized evaluation metrics specifically designed to assess system-level communication effectiveness, collaboration quality, and emergent collective intelligence hinders progress and comparative analysis. Existing benchmarks often focus on individual agent performance or task success without deeply evaluating the communication dynamics.

Future Research Directions

The survey identifies several promising avenues for future research to address the aforementioned challenges:

• Development of sophisticated hybrid communication architectures and more efficient, scalable communication paradigms (e.g., adaptive strategies, optimized blackboard systems).
• Research into optimizing computational resource allocation for communication-intensive tasks and improving agents' abilities to filter, interpret, and synthesize large volumes of potentially noisy information.
• Investigation into the nuanced interplay between competition and cooperation, aiming for systems that can dynamically adapt their interaction styles, along with the development of safe and beneficial competitive frameworks.
• Advancements in multimodal fusion techniques and agent capabilities for processing, generating, and communicating complex multimodal information seamlessly within the MAS.
• Creation of robust, lightweight encryption, authentication, and access control mechanisms specifically designed for the dynamic and potentially decentralized nature of LLM-MAS.
• Establishment of comprehensive benchmarks and evaluation methodologies focusing explicitly on communication efficiency, robustness, collaborative problem-solving capabilities, and security aspects of LLM-MAS.

In conclusion, analyzing LLM-based multi-agent systems through the lens of communication provides a structured approach to understanding their capabilities and limitations. Focusing on architectural design, communication goals, internal mechanisms, and addressing the associated challenges related to scalability, multimodality, security, and evaluation is crucial for advancing the field towards more sophisticated and reliable intelligent systems.