Citation

Hossain, Elias and Nipu, Md. Mehedi Hasan Bhuiyan and Mahmood, Mohammad Sakib and Hossen, Md. Jakir and Mridha, M. F. (2026) Safe and Scalable Collaboration in Multiagent LLM Systems: A Comprehensive Review. IEEE Transactions on Systems, Man, and Cybernetics: Systems. pp. 1-17. ISSN 2168-2216

Text 2.pdf - Published Version Restricted to Repository staff only Download (8MB)

Abstract

Large language model (LLM)-based multiagent systems are a rapidly evolving frontier of artificial intelligence that enables complex coordination, communication, and autonomous reasoning across distributed agents. This article presents a comprehensive review of multiagent LLM ecosystems, organized around four foundational pillars: coordination strategies, communication frameworks, safety challenges, and trust and alignment mechanisms. We synthesize a range of architectural approaches to agent interaction, including shared latent representations, strategic autonomy, task decomposition, and neuro-symbolic communication protocols. We then systematically examine the critical vulnerabilities distinctive to multiagent environments, namely prompt injection, perceptual adversarial attacks, reward misalignment, and operational overreach. To address these issues, we propose the trust stack: a multilayered trust framework spanning epistemic reliability, behavioral alignment, role awareness, social accountability, and institutional governance. Beyond synthesizing the literature, this review contributes a novel integrated perspective. It models the interdependencies among these four pillars, formalizes the trust stack as an operational framework with runtime-computable metrics, and derives design principles from a cross-system analysis of real multiagent deployment failures—contributions absent from prior surveys. We further investigate the prospects for superalignment through constitutional AI principles, multiagent value aggregation, and dynamic preference shaping. Collectively, this work establishes a theoretical and practical foundation for the secure, reliable, and dependable deployment of multiagent LLM systems and offers a roadmap for future research at the intersection of coordination, communication, and alignment.

Item Type:	Article
Uncontrolled Keywords:	Alignment and superalignment, cognitive security
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA165 Engineering instruments, meters, etc. Industrial instrumentation
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	04 Aug 2026 02:52
Last Modified:	04 Aug 2026 02:52
URII:	http://shdl.mmu.edu.my/id/eprint/16466

Downloads

Edit (login required)