Answers are grounded in this paper and cite their sources.
As multi-agent systems expand to massive scales, traditional Multi-Agent Reinforcement Learning (MARL) is severely constrained by the curse of dimensionality, environmental non-stationarity, and intensified credit assignment issues. Existing surveys focus on general MARL or small-scale cooperation, lacking a systematic synthesis of approaches specifically designed for large-scale settings. This survey systematically reviews a curated body of papers published between 2016 and 2026, identified through keyword-based search across six major databases (IEEE Xplore, ACM Digital Library, arXiv, Web of Science, Scopus, and Google Scholar) with defined inclusion and exclusion criteria. We propose a novel taxonomy that categorizes large-scale MARL algorithms based on their underlying problem-solving philosophies: abstraction (temporal, spatial, and population) to reduce model complexity, and decomposition combined with progressive evolution to reduce task difficulty. For each category, we provide critical comparative analysis, discuss failure modes and limitations, and map methods to the specific scalability bottlenecks they address. We further review representative engineering applications—ranging from computational job scheduling and intelligent transportation systems to physical swarm confrontation—and critically evaluate standard benchmark environments. Finally, we identify open challenges, formulate concrete research questions, and propose a three-stage roadmap for the field. This survey aims to serve as a structured guide for researchers entering LS-MARL and a critical reference for practitioners designing scalable multi-agent systems.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: