<p>Cloud task scheduling faces significant challenges due to resource heterogeneity, conflicting optimization objectives, and dynamic workload fluctuations. Traditional heuristic algorithms often necessitate comprehensive knowledge of environmental parameters, significantly constraining their efficacy in dynamic cloud computing environments. While Deep Reinforcement Learning (DRL) methods have shown promise in intelligent scheduling via continuous environment interaction, they suffer from limited generalization to diverse cloud scenarios and lack decision interpretability. To address these shortcomings, this paper proposes LarS, a scheduling framework that employs Large Language Models (LLMs) as high-level decision agents for cloud task scheduling. In LarS, DRL agents trained in carefully chosen representative cloud environments generate a high-quality dataset of scheduling decisions, which is used to fine-tune an LLM. By jointly optimizing average response time, task success rate, and average rental cost, LarS achieves strong generalization across heterogeneous cloud deployments. Experimental results demonstrate that LarS surpasses current approaches in average response time, success rate, and average cost, and maintains strong generalization performance under varied experimental settings.</p>