This survey provides a comprehensive overview of the research on Large Language Model (LLM)-based multi-agent systems. It discusses the key aspects of these systems, including:
Agents-Environment Interface: The ways in which agents interact with and perceive their operational environments, categorized into Sandbox, Physical, and None.
Agent Profiling: The methods used to define agent traits, actions, and skills, including Pre-defined, Model-Generated, and Data-Derived approaches.
Agent Communication: The communication paradigms (Cooperative, Debate, Competitive), structures (Layered, Decentralized, Centralized, Shared Message Pool), and content exchanged between agents.
Agent Capability Acquisition: The feedback sources (Environment, Agent Interactions, Human) and strategies (Memory, Self-Evolution, Dynamic Generation) employed by agents to enhance their abilities.
The survey also categorizes the current applications of LLM-MA systems into two main streams: Problem Solving (Software Development, Embodied Agents, Science Experiments, Science Debate) and World Simulation (Societal, Gaming, Psychology, Economy, Recommender Systems, Policy Making, Disease Propagation).
Additionally, the paper provides an overview of the commonly used implementation frameworks, datasets, and benchmarks in this research field. Finally, it discusses the key challenges and opportunities for future research, including advancing into multi-modal environments, improving agent orchestration, enhancing transparency and interpretability, and addressing safety and ethical concerns.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Taicheng Guo... at arxiv.org 04-22-2024
https://arxiv.org/pdf/2402.01680.pdfDeeper Inquiries