多智能体系统：从经典范式到大基础模型驱动的未来

Summary of "Multi-Agent Systems: From Classical Paradigms to Large Foundation Model-Enabled Futures"

Core Content

This survey explores the evolution of Multi-Agent Systems (MASs) from Classical MASs (CMASs) to Large Foundation Model-based MASs (LMASs). It emphasizes the transition from environment-specific, task-driven coordination mechanisms to more general-purpose, cognitively empowered systems that leverage Large Foundation Models (LFMs) for semantic-level reasoning, planning, and zero/few-shot generalization. The paper presents a comprehensive review of both paradigms, comparing their architectures, operating mechanisms, adaptability, and applications, while also identifying key challenges and future research opportunities in the development of MASs.

Main Contributions

1. Comprehensive Overview: A detailed review of core theories and recent advances in both CMASs and LMASs.
2. Comparative Analysis: A systematic comparison of CMASs and LMASs from theoretical and practical perspectives, highlighting their similarities, differences, and the paradigm shift they represent.
3. Future Directions: Identification of open challenges and potential research opportunities for the development of future MAS systems.

Key Dimensions of Classical MASs (CMASs)

CMASs are structured around a closed-loop coordination framework, with four fundamental dimensions:

• Perception: Involves processing agent state, sensor data, and signals from other agents. It includes early, intermediate, and late fusion methods, with intermediate fusion being the dominant approach due to its balance of communication efficiency and perception performance.
• Communication: Refers to the exchange of information among agents. It is analyzed from three perspectives:
  • Topological: Communication networks are modeled as graph structures with nodes representing agents and edges representing communication links.
  • Frequency: Communication can be event-triggered or time-triggered, with the former reducing redundant data and improving bandwidth efficiency.
  • Content: Includes explicit and implicit communication, where the latter relies on environment-mediated interactions.
• Decision-Making: Governed by cooperative, competitive, and hybrid interaction modes. It is categorized into model-based and learning-based approaches:
  • Model-based: Uses rule-based, game theory, and evolutionary optimization to derive decision strategies.
  • Learning-based: Embraces multi-agent reinforcement learning (MARL) for end-to-end optimization in complex, partially observable environments.
• Control: Focuses on distributed mechanisms for coordination, stability, and global objectives. Two main paradigms are:
  • Consensus Control: Ensures asymptotic convergence of agent states to a common value, with recent studies incorporating learning-based methods to achieve closed-loop stability.
  • Formation Control: Enables agents to form and maintain spatial structures, with approaches ranging from leader-follower to learning-based methods.

These dimensions are interconnected, forming a closed-loop system that enables system-level intelligence.

Key Dimensions of Large Foundation Model-based MASs (LMASs)

LMASs integrate LFMs to enhance reasoning, planning, and generalization capabilities. The paper reviews LMASs across four core aspects:

• Core Modules:

  • Role Definition: Assigns distinct responsibilities to agents, with manually defined and task-adaptive roles.
  • Perception: Structured into semantic, situational, and cognitive levels, enabling task-aware interpretation and anticipatory understanding.
  • Planning: Transforms high-level instructions into executable steps, with strategies including structured planning, feedback-driven optimization, and reliability enhancement.
  • Memory: Stores and retrieves task-relevant knowledge, with short-term and long-term components, enhancing reasoning efficiency and task execution.
  • Execution: Converts reasoning and planning into actions, leveraging language generation and external tools for task-specific performance.

• Interaction Mechanisms:

  • Collaborative Architecture: Includes cooperative, competitive, and hybrid paradigms, enabling dynamic interaction.
  • Task Orchestration: Ranges from structured workflows to autonomous coordination, with frameworks like ReSo and MAS² supporting self-organization and self-evolution.
  • Communication: Emphasizes structured protocols (e.g., MCP, A2A, ANP) and cost reduction strategies, such as pruning-based methods.
  • Human-Agent Interaction: Supports goal alignment, ethical oversight, and iterative feedback, with frameworks like AutoGen and InteRecAgent facilitating collaborative decision-making.

• Hierarchical Optimization:

  • Model Layer: Optimizes reasoning and decision-making through supervised fine-tuning (SFT) and reinforcement learning (RL).
  • Knowledge Layer: Enables cross-task experience accumulation and transfer, enhancing generalization and adaptability.
  • System Layer: Supports cross-environment task execution across device-edge-cloud infrastructures.

Comparative Analysis

Future Perspectives

The paper identifies several key challenges and research opportunities in the development of MASs:

• Scalability: Enhancing system scalability in complex and dynamic environments.
• Robustness: Improving robustness and reliability of LFM-based systems.
• Interpretability: Increasing model interpretability to ensure trustworthy and transparent decision-making.
• Human-in-the-Loop: Developing effective human supervision mechanisms to guide goal alignment and ethical compliance.
• Generalization: Enabling zero-shot and few-shot learning to improve task adaptability.
• Integration with Embodied Intelligence: Combining LFMs with sensory input to enable physical interaction and autonomous execution.

Conclusion

LMASs represent a paradigm shift in multi-agent systems, moving from task-specific and model-dependent frameworks to general-purpose, cognitively empowered, and semantically rich systems. While CMASs remain essential for reliable low-level control, LMASs offer greater flexibility, adaptability, and intelligence through semantic reasoning, planning, and generalization. This survey aims to bridge the gap between these paradigms, promoting synergistic co-evolution and joint advancement in future research.