Multi-Agent Collaboration Mechanisms: A Survey of LLMs

📝 Paper Summary

Multi-agent collaboration Survey / Taxonomy

This survey proposes a comprehensive taxonomy for LLM-based Multi-Agent Systems (MAS), characterizing collaboration via actors, types, structures, and strategies to advance collective intelligence beyond isolated models.

Core Problem

Existing research lacks a unified framework to understand collaboration mechanisms in LLM-based Multi-Agent Systems, as current surveys focus primarily on single-agent capabilities or specific application domains without dissecting how agents interaction works.

Why it matters:

Transitioning from isolated LLMs to collaborative MAS is essential for achieving artificial collective intelligence and horizontal scaling
Current literature overlooks complex collaboration modes like competition and coopetition, limiting the design of robust multi-agent systems
A lack of standardized taxonomy hinders the comparison and development of new coordination protocols for complex, multi-step tasks

Concrete Example: While tools like Auto-GPT demonstrate agent capabilities, prior surveys (e.g., Xi et al., 2023) focus on the individual agent's 'brain' and 'perception' but fail to explain the mechanisms that enable multiple agents to negotiate or debate to solve a shared problem.

Key Novelty

Five-Dimensional Collaboration Framework for MAS

Characterizes MAS not just by application, but by five core dimensions: Actors (agents involved), Types (cooperation, competition, coopetition), Structures (topology), Strategies (role/model-based), and Coordination protocols
Integrates concepts from human collective intelligence (e.g., Society of Mind) into the design of LLM-based agent interactions

Architecture

Conceptual illustration of LLM-based Multi-Agent Systems having multiple collaboration channels with different characteristics

Breakthrough Assessment

7/10

A comprehensive survey that fills a significant gap by structuring the fragmented field of multi-agent collaboration, though it is a review/framework paper rather than a new algorithmic contribution.

⚙️ Technical Details

Problem Definition

Setting: Survey and taxonomy construction for Large Language Model-based Multi-Agent Systems (MAS)

Inputs: Existing literature on LLMs, MAS, and Collaborative AI

Outputs: A unified framework classifying collaboration by Actors, Types, Structures, Strategies, and Protocols

Comparison to Prior Work

vs. Xi et al. (2023): Focuses on multi-agent collaboration mechanisms rather than single-agent architecture
vs. Guo et al. (2024): Expands beyond communication structures to include collaboration types (competition, coopetition) and strategies
vs. Lu et al. (2024): Covers a broader range of collaboration mechanisms beyond just merging/ensemble, including negotiation and debate
+ 1 more
vs. Talebirad and Nadiri (2023): Provides detailed characterization and implementation analysis rather than just a high-level conceptual framework

Limitations

The framework is conceptual and relies on the state of current literature, which is rapidly evolving
Does not propose a new specific algorithm or model, but rather categorizes existing ones
The survey's scope is limited to LLM-based agents, potentially excluding non-LLM MAS research relevance

Reproducibility

This is a survey paper; no specific code or datasets were introduced for reproduction. It references existing works.

📚 Prerequisite Knowledge

Prerequisites

Multi-Agent Systems (MAS)
Large Language Models (LLMs)
Agentic AI concepts

Key Terms

MAS: Multi-Agent Systems—systems composed of multiple interacting intelligent agents that coordinate to solve complex tasks

Agentic AI: AI systems where LLMs act as the 'brain' to perceive, reason, plan, and execute actions using external tools

Coopetition: A collaboration type combining both cooperation and competition, where agents may compete individually but collaborate for a shared higher-level goal

Society of Mind: A theory by Marvin Minsky proposing that intelligence emerges from the interactions of many smaller, simple agents

Hallucination: A phenomenon where LLMs generate plausible but factually incorrect or nonsensical information

Horizontal scaling: Increasing system capability by adding more agents to work together, rather than just increasing the size of a single model

Scaling laws: Empirical observations describing how model performance improves with increases in model size, data, and compute