Categorizes multi-agent decision-making, prioritizing MARL and LLM-based over traditional methods.
A Comprehensive Survey on Multi-Agent Cooperative Decision-Making: Scenarios, Approaches, Challenges and Perspectives categorizes multi-agent decision-making approaches into five main types: rule-based (primarily fuzzy logic), game theory-based, evolutionary algorithms-based, deep multi-agent reinforcement learning (MARL)-based, and large language models (LLMs) reasoning-based. The survey emphasizes that MARL and LLMs-based methods offer significant advantages over traditional approaches such as rule-based, game-theoretic, and evolutionary methods, and thus focuses extensively on these two paradigms.
In particular, MARL techniques are analyzed through key training and execution paradigms: Centralized Training with Centralized Execution (CTCE), Decentralized Training with Decentralized Execution (DTDE), and Centralized Training with Decentralized Execution (CTDE). The survey highlights how CTDE balances comprehensive learning during training with autonomous execution, making it particularly effective in complex, dynamic environments. Methodological frameworks within MARL include value function decomposition, actor-critic architectures, and policy optimization algorithms such as Proximal Policy Optimization (PPO).
The integration of LLMs into multi-agent systems is identified as a transformative advancement, enabling agents to reason, communicate, and coordinate using natural language, thereby improving task decomposition, resource allocation, and collaborative strategies. Simulation environments critical for evaluating these systems include the Multi-Agent Particle Environment (MPE), StarCraft Multi-Agent Challenge (SMAC), and Google Research Football, each offering distinct task formats, reward structures, and levels of partial observability.
The survey also addresses persistent challenges in multi-agent cooperation, such as non-stationarity, credit assignment, scalable communication, and decentralization of tasks. Future research directions emphasize enhancing robustness, adaptability, and theoretical foundations for cooperative decision-making in real-world applications like autonomous driving, drone navigation, disaster rescue, and military simulations.
This survey provides a systematic taxonomy of Multi-Agent Cooperative Decision-Making (MACDM), focusing on the transition from traditional distributed planning to modern, data-driven paradigms. The authors categorize the field across diverse application scenarios—ranging from autonomous driving and cooperative robotics to complex gaming environments—and analyze the algorithmic architectures that enable agents to coordinate effectively. While acknowledging classical game-theoretic and planning-based methods, the text prioritizes the examination of Multi-Agent Reinforcement Learning (MARL) and the emerging domain of Large Language Model (LLM)-based agents, highlighting how these technologies are reshaping cooperative intelligence.
Key contributions of this work include a detailed comparative analysis of MARL techniques—which excel in learning optimal policies through environmental interaction—and LLM-based approaches, which leverage pre-trained knowledge for reasoning and natural language communication. The survey dissects the strengths and limitations of each paradigm, exploring how LLMs can address credit assignment and generalization issues inherent in traditional MARL. Furthermore, the authors outline critical challenges such as scalability, communication efficiency, and the alignment of multi-agent goals, offering a forward-looking perspective on open problems and future research directions.
This material is essential reading for researchers navigating the rapidly evolving landscape of distributed AI. By framing the discussion around the convergence of learning (MARL) and reasoning (LLMs), the survey highlights a pivotal shift in the field: moving agents from reactive, reward-optimizing entities toward proactive, communicative collaborators capable of complex social reasoning. It serves as a foundational resource for understanding how to integrate emergent generative capabilities into robust, cooperative multi-agent systems.
This survey provides a structured overview of multi-agent cooperative decision-making (MACDM), with a particular emphasis on Multi-Agent Reinforcement Learning (MARL) and Large Language Model (LLM)-based approaches, while framing traditional methods (e.g., game theory, classical optimization) as foundational but less dominant in modern research. The paper categorizes MACDM scenarios across domains such as robotics, autonomous driving, and social simulation, highlighting key challenges like non-stationarity, credit assignment, and scalability. It systematically reviews state-of-the-art MARL algorithms (e.g., MADDPG, COMA, MAPPO) and emerging LLM-augmented frameworks, which leverage natural language for interpretability and generalization in cooperative tasks. The authors also discuss hybrid approaches combining RL with symbolic reasoning or LLMs to address limitations in pure RL-based methods.
The survey’s key contributions lie in its critical synthesis of recent trends, particularly the shift toward LLM-augmented MARL, which promises to bridge the gap between declarative communication and learned strategies. It identifies open challenges, such as alignment of agent goals, robustness to adversarial behavior, and real-world deployment constraints, while outlining future directions like self-improving agent societies and human-AI collaboration. For researchers, this work serves as a roadmap for advancing MACDM beyond traditional RL, emphasizing the need for interdisciplinary methods (e.g., integrating planning, language models, and causal reasoning). Its relevance extends to practitioners in robotics, autonomous systems, and AI-driven coordination, offering a timely snapshot of where the field is heading.