Adaptive Cloud Resource Management Using Large Language Models and Reinforcement Learning : A Critical Review and Conceptual Framework

Abstract

Adaptive cloud resource management increasingly requires systems that can respond not only to changing workloads and resource states but also to the broader operational context in which management decisions are made. This study critically reviews recent research on cloud resource management, with particular emphasis on the complementary roles of Large Language Models (LLMs) and Reinforcement Learning (RL), and identifies the main gaps that constrain their effective integration. The reviewed literature covers adaptive scheduling, resource allocation, auto-scaling, workload management, and emerging uses of LLMs for interpreting operational logs, policies, intents, and other unstructured information. The analysis shows that RL provides a strong foundation for sequential and adaptive decision-making through continuous interaction with the environment; however, existing approaches largely rely on predefined numerical states and reward functions, limiting their ability to capture richer operational context. LLMs, in contrast, offer valuable semantic reasoning capabilities but remain constrained by hallucination, reliability, inference latency, and computational cost, and cannot independently provide robust long-term adaptive control. To bridge these limitations, this study proposes a conceptual framework in which LLMs support contextual understanding and decision formulation, while RL drives adaptive resource decisions based on environmental feedback. Safety validation, execution, and feedback are incorporated into a closed-loop architecture. The study argues that such functional integration provides a conceptual foundation for moving from state-aware toward context-aware adaptive cloud management, while highlighting the need for empirical validation under realistic, large-scale cloud environments.