Scholarlink

<p>The combination of Agentic Artificial Intelligence (AI) with Multi-Agent Reinforcement Learning (MARL) and Large Language Models (LLMs) now provides researchers with an effective method to create autonomous systems which can learn and reason and work together in decentralized systems. Agentic AI frameworks use adaptive policy learning together with language-based reasoning and communication to create a new system which differs from previous rule-based and single-agent designs. The paper delivers a comprehensive integrative review which examines current LLM–MARL research through various aspects of agent architectures and coordination mechanisms and memory augmentation and evaluation methodologies used in different fields of study. An analysis of the literature reveals three recurring patterns: the use of LLMs for inter-agent communication and high-level planning, the application of hierarchical control structures to MARL agents, and memory-driven decision-making supported by retrieval mechanisms. The study presents supply chain management as an application domain which demonstrates how hybrid systems enable organizations to make decentralized decisions during times of uncertainty. The review identifies three main categories of challenges which organizations face when they attempt to achieve scalability and robustness and accountability and evaluation standardization in their operations. The research results demonstrate that LLM–MARL-based agentic systems show considerable potential, with ongoing advancements steadily improving their readiness for safe real-world organizational deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integration of Multi-agent Reinforcement Learning (MARL) and Large Language Models (LLMs) in Agentic AI Systems: A Narrative Review

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Harsh Shingavi,
  • Yash Ranbhare,
  • Esha Shasri,
  • Jyoti Wadmare

摘要

The combination of Agentic Artificial Intelligence (AI) with Multi-Agent Reinforcement Learning (MARL) and Large Language Models (LLMs) now provides researchers with an effective method to create autonomous systems which can learn and reason and work together in decentralized systems. Agentic AI frameworks use adaptive policy learning together with language-based reasoning and communication to create a new system which differs from previous rule-based and single-agent designs. The paper delivers a comprehensive integrative review which examines current LLM–MARL research through various aspects of agent architectures and coordination mechanisms and memory augmentation and evaluation methodologies used in different fields of study. An analysis of the literature reveals three recurring patterns: the use of LLMs for inter-agent communication and high-level planning, the application of hierarchical control structures to MARL agents, and memory-driven decision-making supported by retrieval mechanisms. The study presents supply chain management as an application domain which demonstrates how hybrid systems enable organizations to make decentralized decisions during times of uncertainty. The review identifies three main categories of challenges which organizations face when they attempt to achieve scalability and robustness and accountability and evaluation standardization in their operations. The research results demonstrate that LLM–MARL-based agentic systems show considerable potential, with ongoing advancements steadily improving their readiness for safe real-world organizational deployment.